Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge
打开互动全文版（逐段中英对照 + 图/公式 + 针对本篇提问）→视频扩散模型正快速扩展并展现出增强的生成能力。在近期进展中，MiniMax-H3 作为一个高度 capable、生产级开源模型脱颖而出。然而，其 330 亿参数和多步迭代去噪过程带来了巨大的计算开销。因此，其实际生产受到云部署（如 NVIDIA-GB200）中生成延迟的阻碍，同时边缘设备（如 DGX-Spark）的严格内存限制也带来了进一步挑战。为了解决从云到边缘设备的多样化硬件瓶颈，我们提出了一种全栈推理流水线，将高效算法设计与优化算子实现相结合。在算法层面，我们引入了一种跨分辨率两阶段生成调度器，利用扩散的逐步特性：早期低分辨率步骤快速建立全局布局，而后期高分辨率步骤则专注于局部和感知细节的细化。这些阶段通过一个学习到的潜在到潜在映射模块连接，完全消除了跨不同分辨率传输时计算昂贵的 VAE 解码-重编码循环。在算子实现方面，我们部署了一个递归自我改进（RSI）循环，搜索内核融合和内存布局，同时评估延迟和数值一致性。这些优化共同实现了高达 30 倍的端到端加速和 20%的内存降低：在 8xGB200 节点上，生成带音频的 5 秒 1344x768 视频比实时快 3.5 倍，而在单个 DGX Spark 上，完全内存驻留生成时间不到一分钟。
Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.
（*）同等贡献。NVIDIA Research，高效 AI 团队与新加坡实验室。
(*) Equal contribution. NVIDIA Research, Efficient AI Team & Singapore Lab.
视频扩散模型正快速 Scaling（规模扩张），并展现出更强的生成能力。在近期进展中，MiniMax-H3 作为一款能力出众、达到生产级水平的开源模型尤为突出。然而，其 330 亿参数以及多步迭代去噪过程带来了巨大的计算开销。因此，其在 NVIDIA-GB200 等云端部署中受生成延迟制约，同时在 DGX-Spark 等边缘设备上还面临严格的内存限制，这些因素阻碍了其实际生产落地。为应对从云端到边缘设备的各种硬件瓶颈，我们提出了一套全栈推理流水线，将高效的算法设计与优化的算子实现相结合。在算法层面，我们引入了一种跨分辨率两阶段生成调度器，利用扩散过程的逐步特性：早期低分辨率步骤快速建立全局布局，而后期高分辨率步骤则专注于局部与感知细节的精修。这两个阶段通过一个可学习的 latent-to-latent 映射模块连接，彻底消除了跨不同分辨率转换时计算昂贵的 VAE 解码-重编码循环。在算子实现方面，我们部署了一个递归自我改进（RSI）循环，用于搜索核融合与内存布局，并同时评估延迟与数值一致性。这些优化确保我们精简的算法设计在不同硬件架构上充分发挥其性能潜力与实际延迟收益。通过结合这些创新，我们的流水线实现了高达 30 \(\times\) 的加速和 20% 的内存节省，在 8 \(\times\) GB200 云节点上实现最高 3.5 \(\times\) 的实时生成加速，并在 DGX-Spark 上实现不到 1 分钟的全内存驻留执行。
Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. These optimizations ensure our streamlined algorithmic design fully realizes its performance potential and practical latency benefits across varying hardware architectures. By combining these innovations, our pipeline achieves exceptional efficiency up to 30 \(\times\) speedup and 20% memory saving, delivering up to 3.5 \(\times\) faster-than-real-time generation on 8 \(\times\) GB200 cloud nodes and less than 1-min fully memory-resident execution on DGX-Spark.
链接：Github 代码 | 项目页面
Links: Github Code | Project Page
在模型与数据 Scaling（规模扩张）的推动下，视频扩散模型已展现出卓越的生成能力，近期开源发布如 MiniMax-H3 更是达到了生产级质量。尽管取得了这些进展，该模型庞大的 330 亿参数以及多步迭代去噪过程带来了沉重的算力负担。此外，推理过程中遇到的具体瓶颈在不同硬件平台和部署场景下差异巨大。在云环境中，系统吞吐量和生成成本是主要障碍，最终决定了服务的经济可行性。相反，边缘部署则受到内存限制的严格约束；超出可用 HBM 会迫使系统进行 CPU 卸载，导致推理延迟灾难性崩溃。鉴于这些硬件约束，单一优化策略无法同时解决这两个问题，使得整体优化过程困难得多。
Driven by model and data scaling, video diffusion models have achieved remarkable generative prowess, with recent open-source releases like MiniMax-H3 delivering production-grade quality. Despite these advances, the model's massive 33-billion parameters and the multi-step iterative denoising process impose a severe computational burden. Furthermore, the specific bottlenecks encountered during inference vary drastically across different hardware platforms and deployment scenarios. In cloud environments, system throughput and generation cost act as the primary hurdles, ultimately determining the economic viability of the service. Conversely, edge deployments are tightly constrained by memory limits; exceeding the available HBM forces the system to CPU-offloading, precipitating a catastrophic collapse in inference latency. Given these hardware constraints, a single optimization strategy cannot solve both issues, making the overall optimization process much more difficult.
扩散模型的加速依赖于两种不同的范式。第一种是算法加速，其本质上通过减少采样步数、缓存去噪计算、实现稀疏注意力机制或修改生成分辨率等方法来改变生成过程，这些方法通常是有损的。第二种是系统级加速，它在保持数学输出不变的同时应用底层系统优化，例如算子融合、内存布局重构和通信调度。
The acceleration of diffusion models has relied on two distinct paradigms. The first is algorithmic acceleration, which inherently alters the generation process through methods such as reducing sampling steps, caching denoising computation, implementing sparse attention mechanisms, or modifying generation resolutions, which are often in a lossy manner. The second is system-level acceleration, which preserves the mathematical output while applying low-level system optimizations, such as kernel fusion, memory layout restructuring, and communication scheduling.
在本报告中，我们从算法和实现两个角度优化推理流水线。对于算法加速，我们利用扩散模型的一个基本演化特性：早期去噪步骤决定视频的整体结构布局，而后期步骤则合成高频细节。基于此，我们设计了一种跨分辨率两阶段生成流水线，旨在最大化每个阶段的效率。第一阶段在低分辨率下运行，快速建立全局布局和结构骨架。第二阶段切换到高分辨率，完全专注于细化视觉细节。为了无缝连接这些阶段，我们训练了一个特征映射模块，执行直接的 latent-to-latent 转换，完全消除了传统像素级解码-重编码循环的延迟和内存开销。对于系统级实现，我们引入递归自我改进（RSI）循环来进行迭代且可验证的优化。在严格的正确性验证下运行，该 RSI 循环自动探索一个庞大的搜索空间，涵盖算子融合、内存布局和通信边界。自动搜索持续发现并优化底层计算和通信瓶颈，在广泛的硬件平台上实现巨大的算子级加速。这些优化在不同硬件架构上实现了我们高效推理流水线的实际延迟收益，同时减少了人工投入。
In this report, we optimize the inference pipeline from two perspectives: algorithm and implementation. For algorithmic acceleration, we leverage a fundamental evolutionary property of diffusion models: early denoising steps determine the overall structural layout of the video, while later steps synthesize high-frequency fine details. Building on this, we design a cross-resolution two-stage generation pipeline tailored to maximize efficiency at each phase. The first stage operates at a low resolution to rapidly establish the global layout and structural skeleton. The second stage shifts to a high resolution to focus entirely on refining visual details. To seamlessly connect these stages, we train a feature mapping module that performs direct latent-to-latent translation, completely eliminating the latency and memory overhead of the traditional pixel-level decode-and-reencode cycle. For system-level implementation, we introduce a Recursive Self-Improvement (RSI) loop to conduct iterative and verifiable optimization. Operating under strict correctness verification, this RSI loop automatically explores a massive search space encompassing kernel fusions, memory layouts, and communication boundaries. The automated search consistently uncovers and optimizes low-level compute and communication bottlenecks, delivering massive kernel-level speedups across a wide range of hardware platforms. These optimizations realize the practical latency benefits of our efficient inference pipeline across varying hardware architectures, while requiring less human effort.
整合这些算法和系统优化，我们构建了一个跨多种硬件平台的高效全栈推理流水线。在 8 \(\times\) GB200 节点上，Sol-H3 在 1.434 秒内生成带有原生音频的五秒视频，比实时播放快约 3.5 \(\times\)（表 1）。在单个 GB200 上，它实现了端到端 6.23 秒的生成，相比服务基线加速 22.2 \(\times\)。此外，在单个 DGX Spark 上，整个流水线完全驻留在 HBM 中，在 56.2 秒内完成视频生成，未触发 CPU 卸载，实现了近 20% 的内存减少和超过 \({30\times}\) 的加速。
Integrating these algorithmic and systematic optimizations, we build a highly efficient full-stack inference pipeline across various hardware platforms. On an 8 \(\times\) GB200 node, Sol-H3 generates a five-second video with native audio in 1.434 seconds, approximately 3.5 \(\times\) faster than real-time playback (Tab. 1). On a single GB200, it achieves an end-to-end 6.23s generation, a 22.2 \(\times\) speedup over the serving baseline. Furthermore, on a single DGX Spark, the entire pipeline remains fully HBM resident, completing the video generation in 56.2 seconds without triggering CPU offloading, achieving a near 20% memory reduction and more than \({30\times}\) speedup.
我们在下面总结本报告的主要贡献。
We summarize the main contributions of this report below.
一种跨分辨率的两阶段生成流水线。利用扩散模型的逐步特性，我们将低分辨率的早期步骤分配给全局布局生成，将高分辨率的后期步骤分配给精细细节细化，从而减少均匀分辨率去噪过程中的算法冗余。
A cross-resolution two-stage generation pipeline. Exploiting the step-wise nature of diffusion models, we allocate low-resolution earlier steps to global layout generation and high-resolution later steps to fine detail refinement, reducing the algorithmic redundancy of the uniform-resolution denoising process.
面向 MiniMax-H3 的全栈推理实现。通过算法与系统级优化的协同设计，我们的流水线在云端和边缘环境中实现了约 30\(\times\) 的加速和 20% 的内存降低。这为 MiniMax-H3 的实际部署解锁了可能。
A full-stack inference implementation for MiniMax-H3. Through the co-design of algorithmic and system-level optimizations, our pipeline achieves an approximate 30 \(\times\) acceleration and a 20% memory reduction across cloud and edge environments. This unlocks practical deployment for MiniMax-H3.
通过递归自我改进（Recursive Self-Improvement，RSI）实现的自动化系统级优化。我们引入了一个 RSI 循环，用于搜索核融合、内存布局、集合通信布局和传输精度，分析其延迟，并检查数值一致性。我们区分了位精确的布局变换与算术融合及近似执行模式，从而为所报告的加速给出了明确的数值范围。
Automated system-level optimization via Recursive Self-Improvement (RSI). We introduce an RSI loop that searches kernel fusions, memory layouts, collective layouts, and transport precision, profiles their latency, and checks numerical agreement. We distinguish bit-exact layout transformations from arithmetic fusions and approximate execution modes, giving the reported accelerations an explicit numerical scope.
当前生成视频领域的格局由开放的大规模扩散 Transformer 所定义。突出的基础模型包括 CogVideo 系列、HunyuanVideo 及其后继者 HunyuanVideo 1.5，以及 Wan、LongCat-Video、SANA-Video、Cosmos 3 和 JoyAI-Echo。本报告专门服务于 MiniMax-H3，一个 330 亿参数的联合音视频生成器，以及 LTX-2.5 及其前身 LTX-2.3，它们提供第二阶段细化。我们的潜在交接结合了空间上采样与在 H3 和 LTX VAE 表示之间的单独学习转换。与仅视觉模型相比，MiniMax-H3 的联合音视频特性引入了独特的服务挑战：音频轨道必须在应用于管道的每一个优化中存活下来，而不仅仅是视觉帧。
The current landscape of generative video is defined by open, large-scale diffusion transformers. Prominent base models include the CogVideo family, HunyuanVideo and its successor HunyuanVideo 1.5, as well as Wan, LongCat-Video, SANA-Video, Cosmos 3, and JoyAI-Echo. This report specifically serves MiniMax-H3, a 33-billion parameter joint audio-video generator, alongside LTX-2.5 and its predecessor LTX-2.3, which provide second-stage refinement. Our latent handoff combines spatial upsampling with a separate learned translation between the H3 and LTX VAE representations. The joint audio-video nature of MiniMax-H3 introduces a distinct serving challenge compared to vision-only models: the audio track must survive every optimization applied to the pipeline, not merely the visual frames.
为了管理高分辨率生成的计算复杂性，级联和多分辨率生成调度是文献中成熟的方案。我们采用类似的范式，利用来自 MiniMax-H3 的低分辨率草稿，随后用 LTX-2.5 进行高分辨率细化。我们强调，低分辨率草稿和高分辨率细化策略是现有技术；我们的贡献，详见第 3.1 节，在于服务实现和端到端执行栈，而非概念调度本身。
To manage the computational complexity of high-resolution generation, cascaded and multi-resolution generation schedules are well-established recipes in the literature. We employ a similar paradigm, utilizing a low-resolution draft from MiniMax-H3 followed by a high-resolution refinement pass with LTX-2.5. We emphasize that the low-resolution draft and high-resolution refine strategy is prior art; our contribution, detailed in Sec. 3.1, lies in the serving realization and the end-to-end execution stack rather than the conceptual schedule itself.
扩散 Transformer 的推理加速通常通过三个主要杠杆实现。第一个杠杆是稀疏注意力机制，它缓解了视频生成中序列长度的二次方缩放问题。近期方法包括 Sparse VideoGen 和 Sparse VideoGen2、可训练稀疏注意力（VSA）、Sliding Tile Attention、SpargeAttn、XAttention、Radial Attention、DraftAttention 和 PISA。此外，SageAttention 系列工作提供了高度优化的注意力后端。我们在稀疏注意力实现中利用了类似原理，详见第 3.4 节。第二个杠杆涉及缓存机制和减少有效步数的技术，从而限制每步评估的网络部分。TeaCache、EasyCache、Pyramid Attention Broadcast、TaylorSeer 和 Cache-DiT 等方法利用了时间冗余。这些步数减少技术与我们的两阶段调度正交；缓存减少了固定分辨率轨迹内的计算，而我们的调度改变了分辨率轨迹本身。
Inference acceleration for diffusion transformers generally operates across three primary levers. The first lever is sparse attention, which mitigates the quadratic scaling of sequence length in video generation. Recent approaches include Sparse VideoGen and Sparse VideoGen2, trainable sparse attention (VSA), Sliding Tile Attention, SpargeAttn, XAttention, Radial Attention, DraftAttention, and PISA. Additionally, the SageAttention line of work provides highly optimized attention backends. We leverage similar principles for our sparse attention implementation, detailed in Sec. 3.4. The second lever involves caching mechanisms and techniques to reduce the number of effective steps, thereby limiting how much of the network is evaluated per step. Methods such as TeaCache, EasyCache, Pyramid Attention Broadcast, TaylorSeer, and Cache-DiT exploit temporal redundancy. These step-reduction techniques are orthogonal to our two-stage schedule; while caching reduces computation within a fixed resolution trajectory, our schedule alters the resolution trajectory itself.
第三个杠杆包括量化、token 剪枝和内核级优化。扩散模型的量化技术包括 ViDiT-Q、PTQ4DiT、Q-DiT、SVDQuant、AWQ 以及微缩放格式的采用。token 级减少策略，如 ToMeSD、Astraea、TAPE 和 CoReDiT，进一步压缩了计算图。这些优化通常部署在现代服务系统和序列并行框架的 broader 上下文中，包括 vLLM、SGLang、DeepSpeed Ulysses 和 LongLive-2.0。
The third lever encompasses quantization, token pruning, and kernel-level optimizations. Quantization techniques for diffusion models include ViDiT-Q, PTQ4DiT, Q-DiT, SVDQuant, AWQ, and the adoption of microscaling formats. Token-level reduction strategies, such as ToMeSD, Astraea, TAPE, and CoReDiT, further compress the computational graph. These optimizations are typically deployed within the broader context of modern serving systems and sequence parallelism frameworks, including vLLM, SGLang, DeepSpeed Ulysses, and LongLive-2.0.
在本报告中，我们重点依赖推理优化中的一个基本区分：保持生成契约的方法与改变调度或步数的方法。算子融合（在第 3.5 节讨论）、精确注意力后端和无损集合布局等技术提供了同等的运行时加速而不改变输出。相反，调度修改产生根本不同的产物。虽然这两类优化在实践中有效组合，但在报告加速时不能混为一谈，因为评估不同的生成契约不等同于加速固定的生成契约。
Throughout this report, we lean heavily on a fundamental distinction in inference optimization: methods that preserve the generation contract versus methods that alter the schedule or the step count. Techniques such as operator fusion (discussed in Sec. 3.5), exact attention backends, and lossless collective layouts provide like-for-like runtime speedups without changing the output. Conversely, schedule modifications produce a fundamentally different artifact. While these two categories of optimization compose effectively in practice, they must not be conflated when reporting speedups, as evaluating a different generation contract is not equivalent to accelerating a fixed one.
我们递归自我改进循环的总体背景是智能体式编码与机器学习实验系统的迅速涌现。SWE-agent、AutoCodeRover、Agentless 和 OpenHands 等框架自动化了软件工程任务，而 AgentBench、MLAgentBench、The AI Scientist 和 AI Harness Engineering 等基准与系统则探索自主研究与实验。更具体地说，本报告属于旨在编写和优化 GPU 内核的智能体系列。KernelBench 基准表明，前沿模型在没有迭代反馈的情况下执行此任务表现不佳。后续系统已弥补这一差距：CUDA-LLM 和 CudaForge 专注于内核生成，AccelOpt 引入了一个自我改进的智能体系统，它整理慢-快内核对的优化记忆，而 AlphaEvolve 采用进化编码智能体进行算法发现。
The general context for our recursive self-improvement loop is the rapid emergence of agentic coding and machine-learning experimentation systems. Frameworks such as SWE-agent, AutoCodeRover, Agentless, and OpenHands automate software engineering tasks, while benchmarks and systems like AgentBench, MLAgentBench, The AI Scientist, and AI Harness Engineering explore autonomous research and experimentation. More specifically, this report belongs to the lineage of agents designed to write and optimize GPU kernels. The KernelBench benchmark demonstrated that frontier models perform this task poorly without iterative feedback. Subsequent systems have addressed this gap: CUDA-LLM and CudaForge focus on kernel generation, AccelOpt introduces a self-improving agentic system that curates an optimization memory of slow-fast kernel pairs, and AlphaEvolve employs an evolutionary coding agent for algorithmic discovery.
我们在此领域的直接前身是 Sol 视频推理引擎，一个智能体原生的全栈加速框架。Sol 通过专门技能智能体的层次结构和中央集成智能体，组合了缓存、稀疏注意力机制、token 剪枝、量化和内核融合。本报告代表了 Sol 框架对单一模型家族的深度应用，通过引入严格的故障关闭契约作为优化过程的显式约束，扩展了先前的工作。
Our direct predecessor in this domain is the Sol video inference engine, an agent-native, full-stack acceleration framework. Sol composes caching, sparse attention, token pruning, quantization, and kernel fusion through a hierarchy of specialized skill agents and a central integrator agent. This report represents the deep application of the Sol framework to a single model family, extending the prior work by introducing a strict fail-closed contract as an explicit constraint on the optimization process.
这种故障关闭约束构成了我们方法论的关键定位论据。无约束的自我改进智能体可能会无意中重新定义它们正在优化的任务——例如，通过静默减少步数或改变数值轨迹——这使得它们报告的性能提升难以归因。我们的递归自我改进循环被刻意限制：它搜索固定生成契约的实现（如数值格式、内核边界、注意力机制后端、集合布局和内存驻留），但从不改变契约本身。因此，被接受的更改保证是在相同硬件上针对相同基线的同类减少，而被拒绝的候选者则产生精确测量，而非未量化的判断。
This fail-closed constraint forms a critical positioning argument for our methodology. Unconstrained self-improving agents may inadvertently redefine the task they are optimizing—for instance, by silently reducing step counts or altering the numerical trajectory—which makes their reported performance gains difficult to attribute. Our recursive self-improvement loop is deliberately restricted: it searches realizations of a fixed generation contract (such as numerical formats, kernel boundaries, attention backends, collective layouts, and memory residency) but never alters the contract itself. Consequently, an accepted change is guaranteed to be a like-for-like reduction measured against the same baseline on the same hardware, and a rejected candidate yields a precise measurement rather than an unquantified judgment.
鉴于视频生成对算力的巨大需求，云环境成为主要的部署载体。为最大化服务效率并降低每次生成的算力成本，我们将原本单阶段、统一分辨率的去噪过程重构为非对称的两阶段生成流水线。我们进一步应用递归自我改进（RSI）循环来迭代优化执行栈。该 RSI 过程系统性地消除各阶段的冗余，提升计算效率与生成吞吐量。
Given the massive computational demands of video generation, cloud environments serve as the primary deployment vehicle. To maximize serving efficiency and reduce the compute cost per generation, we restructure the original single-stage, uniform-resolution denoising process into an asymmetric two-stage generation pipeline. We further apply a Recursive Self-Improvement (RSI) loop to iteratively optimize the execution stack. This RSI process systematically eliminates redundancy across each stage, elevating computational efficiency and generation throughput.
本节详细阐述 H3 运行时及其跨分辨率扩展在算法、内核与系统层面的优化。适用配置在第 6 节中分别评估。特别地，第 6.2 节中在 8 \(\times\) GB200 上的单阶段 Sol-H3 可在 1.434 秒内生成一段五秒视频，约比实时播放快 3.5 \(\times\)。
In this section, we detail algorithm, kernel, and system optimizations for the H3 runtime and its cross-resolution extension. The applicable configurations are evaluated separately in Sec. 6. In particular, single-stage Sol-H3 on 8 \(\times\) GB200 in Sec. 6.2 generates a five-second video in 1.434 seconds, approximately 3.5 \(\times\) faster than real-time playback.
虽然已发布的流水线在全分辨率下运行整个去噪轨迹，但我们将其转换为两阶段高效推理流水线。由于早期扩散迭代建立全局布局和运动，而后续步骤合成高频细节，因此在原生分辨率下执行完整轨迹会消耗大部分算力来解决结构组成，之后精细特征才出现。为了解决这个问题，Sol-H3 流水线动态分配空间分辨率，将传统的全分辨率（\(1344 \times 768\)）下的 49 个去噪步骤重构为非对称多阶段调度：
While the released pipeline runs the entire denoising trajectory at full resolution, we convert it into a two-stage efficient inference pipeline. Because early diffusion iterations establish global layout and motion while later steps synthesize high-frequency detail, executing the full trajectory at native resolution expends most compute resolving structural composition before fine features emerge. To address this, the Sol-H3 pipeline allocates spatial resolution dynamically, restructuring the conventional 49 denoising steps at full resolution (\(1344 \times 768\) ) into an asymmetric multi-stage schedule:
阶段 1 使用 MiniMax-H3 并增强 FastH3 VSA DataFree LoRA 来生成核心视觉元素和主音轨。它在 \(672 \times 384 \times 124\) 帧上执行四次更新迭代。阶段 2 使用 LTX-2.5 合成高频细节。它在 \(1344 \times 768\) 潜在画布上执行三次联合音视频更新，使用常驻 BF16 LTX-2.5 开发骨干网络，并以强度 0.8 应用蒸馏 LoRA 450。
Stage 1 generates core visual elements and the primary audio track using MiniMax-H3 augmented with the FastH3 VSA DataFree LoRA. It executes four update iterations at \(672 \times 384 \times 124\) frames. Stage 2 synthesizes high-frequency detail using LTX-2.5. It executes three joint audio–video updates on the \(1344 \times 768\) latent canvas using a resident BF16 LTX-2.5 dev backbone with distilled LoRA 450 applied at strength 0.8.
这两个阶段使用独立训练的 VAE，因此仅匹配空间分辨率并不能使它们的潜变量可互换。H3 以 24 个通道和 \（16\times）的空间压缩表示视频，而 LTX-2.5 Conv Video VAE 使用 128 个通道和 \（32\times）的空间压缩。直接连接会将 H3 潜变量解码为像素，并用 LTX VAE 重新编码这些像素，从而在交接中增加两个大型网络及其中间激活。我们用学习到的潜变量到潜变量转换器替代这一往返过程。空间放大仍是一个独立操作：学习到的 H3 \（\times 2）上采样器首先提高草稿分辨率，然后转换器将该表示映射到相同像素分辨率的 LTX。
The two stages use independently trained VAEs, so matching spatial resolution alone does not make their latents interchangeable. H3 represents video with 24 channels and spatial compression of \(16\times\), whereas the LTX-2.5 Conv Video VAE uses 128 channels and spatial compression of \(32\times\). A direct connection would decode the H3 latent into pixels and re-encode those pixels with the LTX VAE, adding two large networks and their intermediate activations to the handoff. We replace this round trip with a learned latent-to-latent translator. Spatial enlargement remains a separate operation: the learned H3 \(\times 2\) upscaler first raises the draft resolution, and the translator then maps that representation to LTX at the same pixel resolution.
主要的对齐困难是时间上的。H3 以名义时间压缩 \（4\times）编码独立的 17 帧块，在块边界产生非均匀的 token 位置；LTX 使用因果 \（8\times）时间网格。因此，我们根据 token 的物理帧位置而非张量索引进行对齐。一个固定的前端将每个 LTX 时间位置上线性插值的 H3 特征与包含最近邻源 token 的三个槽位拼接。每个源 token 在打包槽位中恰好保留一次。然后空间像素反混洗将每个 \（2\times 2）邻域移入通道，在 LTX 网格上产生 \(24\times(1+3)\times 4=384) 个通道。这种无参数打包保留了源特征，而不是在学习跨 VAE 映射之前将其平均掉。
The main alignment difficulty is temporal. H3 encodes independent 17-frame chunks with nominal temporal compression of \(4\times\), producing nonuniform token positions across chunk boundaries; LTX uses a causal \(8\times\) temporal grid. We therefore align tokens by their physical frame positions rather than their tensor indices. A fixed front end concatenates a linearly interpolated H3 feature at each LTX time position with three slots containing the nearest-bin source tokens. Each source token is retained exactly once in the packed slots. A spatial pixel-unshuffle then moves each \(2\times 2\) neighborhood into channels, producing \(24\times(1+3)\times 4=384\) channels on the LTX grid. This parameter-free packing preserves the source features instead of averaging them away before learning the cross-VAE mapping.
在这个对齐网格上，适配器结合了一个冻结的 \（1\times 1\times 1）仿射跳跃连接（通过岭回归初始化）和一个学习到的残差网络。该网络包含 22 个宽度为 752 的块，每个块结合了空间卷积、深度时间卷积和门控通道 MLP。完整的转换器有 194.76M 参数，并预测归一化的 128 通道 LTX 潜变量。
On this aligned grid, the adapter combines a frozen \(1\times 1\times 1\) affine skip, initialized by ridge regression, with a learned residual network. The network contains 22 blocks of width 752, each combining spatial convolution, depthwise temporal convolution, and a gated channel MLP. The complete translator has 194.76M parameters and predicts normalized 128-channel LTX latents.
训练对是通过使用两个冻结 VAE 的后验模式对同一调整大小的视频进行编码而获得的。令 \（z_{H}）和 \（z_{L}）表示归一化的 H3 和 LTX 潜变量，\(G) 表示固定对齐，\（A_{\theta}）表示适配器。我们监督潜变量预测 \(\hat{z}_{L}=A_{\theta}(G(z_{H}))) 及其解码视频：
Training pairs are obtained by encoding the same resized video with both frozen VAEs using their posterior modes. Let \(z_{H}\) and \(z_{L}\) denote the normalized H3 and LTX latents, \(G\) the fixed alignment, and \(A_{\theta}\) the adapter. We supervise both the latent prediction \(\hat{z}_{L}=A_{\theta}(G(z_{H}))\) and its decoded video:
其中 \（D_{L}）包括逆潜变量归一化、冻结的 LTX Conv 解码器以及转换为限制在 \（[0,1]）的 RGB 值。梯度通过解码器传递到适配器，而 VAE 权重保持固定。这种解码监督考虑了不同潜变量方向上的误差对视觉的不同影响。最终训练阶段使用 65,536 个 \（1344\times 768）的配对视频；实现和评估细节见附录 A。
where \(D_{L}\) includes inverse latent normalization, the frozen LTX Conv decoder, and conversion to RGB values clamped to \([0,1]\). Gradients pass through the decoder to the adapter while the VAE weights remain fixed. This decoded supervision accounts for the unequal visual effects of errors in different latent directions. The final training phases use 65,536 paired videos at \(1344\times 768\); implementation and evaluation details appear in Appendix A.
在推理时，上采样器保持其规定的输入与输出归一化，适配器输出以归一化的 LTX 坐标直接进入精炼器。若再次施加 LTX 归一化，将会改变输入分布。对于五秒服务配置，时间转换产生 17 个潜在帧；精炼器消费前 16 个以生成 121 个输出帧。适配器仅作用于视频潜在表示，保持 Stage-1 音频不变。移除中间的 H3 解码器与 LTX 编码器降低了交接环节的算力与内存需求，从而支持第 4 节所述的共驻部署。
At inference, the upscaler retains its prescribed input and output normalization, and the adapter output enters the refiner directly in normalized LTX coordinates. Applying LTX normalization a second time would change the input distribution. For the five-second serving configuration, temporal conversion produces 17 latent frames; the refiner consumes the first 16 to produce 121 output frames. The adapter operates only on video latents, leaving the Stage-1 audio unchanged. Removing the intermediate H3 decoder and LTX encoder reduces the handoff's compute and memory requirements, supporting the co-resident deployment described in Sec. 4.
我们在多个 GPU 上并行化推理，但通信与张量布局变换限制了实际获得的加速比。注意力机制需要在按序列分片与按头分片的布局之间进行数据交换，这使得其周边的集合通信成为 RSI 优化循环的关注重点。
We parallelize inference across GPUs, but communication and tensor layout transformations limit the realized speedup. Attention requires exchanges between sequence-sharded and head-sharded layouts, making the surrounding collectives a focus of the RSI optimization loop.
在一个具有代表性的全分辨率 H3 性能剖析中，布局变换和集合通信在每个去噪步骤中累计耗时 94 ms，此外还有 341 ms 的 SDPA 注意力机制计算。这 94 ms 包括 42 ms 的布局复制和 52 ms 的通信。基线在每个注意力机制层执行三次 QKV 集合通信和一次输出集合通信。保留打包的 QKV 存储将三次输入交换合并为一次，从而将总数减少到两次集合通信，而不改变未量化的有效载荷量。在批大小为 1 时，扁平（total, heads, head_dim）布局还消除了冗余的批次维度。然而，一个朴素的 PyTorch 实现仅达到参考吞吐量的 \(0.953\) 倍，即降低了 4.7%：先堆叠，然后进行目标主序的 permute(...).contiguous()，在通信前复制了激活量两次。优化的步幅感知内核则通过现有的步幅读取 Q、K 和 V，并一次性直接写入交换缓冲区。逆头合并使用相同的方法在返回路径上。这些仅布局的内核与它们所替代的参考排列在比特级完全相同。
In a representative full-resolution H3 profile, layout transformations and collectives accumulated 94 ms per denoising step, alongside 341 ms of SDPA attention computation. The 94 ms comprises 42 ms of layout copies and 52 ms of communication. The baseline executes three QKV collectives and one output collective per attention layer. Retaining packed QKV storage combines the three input exchanges into one, reducing the total to two collectives without changing the unquantized payload volume. A flat (total, heads, head_dim) layout also removes the redundant batch dimension at batch size one. However, a naive PyTorch implementation achieved only \(0.953\) times the reference throughput, a 4.7% reduction: stack followed by a destination-major permute(...).contiguous() copied the activation volume twice before communication. The optimized stride-aware kernels instead read Q, K, and V through their existing strides and write directly into the exchange buffer in one pass. The inverse head merge uses the same approach on the return path. These layout-only kernels are bit-identical to the reference permutations they replace.
在序列并行下，打包的 QKV 交换与注意力机制输出回传在每个注意力机制层需要两次 all-to-all 集合通信。block-INT8 QKV 格式将一条 token/head 记录存储为 384 个量化值字节、十二个 group-32 有限正数 UE5M3 缩放码以及四个零填充字节，总计 400 字节。我们评估了两种不同的输出格式。block-INT8 格式存储 128 个值字节和四个缩放码，从 132 字节填充至 144 字节，以使每条记录对齐到 16 字节；尽管有效载荷更大，这种对齐格式在 GB200 上提升了实测的 NCCL 传输性能。FP8 替代方案则对每个 token/head 传输 128 个原始 E4M3 字节，不带单独的缩放元数据或填充。所有传输的填充字节均被显式初始化。两种量化线格式均为近似格式，并与位精确的布局变换分开验证。
Under sequence parallelism, the packed QKV exchange and attention-output return require two all-to-all collectives per attention layer. The block-INT8 QKV format stores one token/head record as 384 quantized value bytes, twelve group-32 finite-positive UE5M3 scale codes, and four zero padding bytes, for 400 bytes in total. We evaluate two distinct output formats. The block-INT8 format stores 128 value bytes and four scale codes, padded from 132 to 144 bytes to align each record to 16 bytes; this aligned format improves the measured NCCL transfer performance on GB200 despite its larger payload. The FP8 alternative instead transmits 128 raw E4M3 bytes per token/head, without separate scale metadata or padding. All transmitted padding bytes are explicitly initialized. Both quantized wire formats are approximate and are validated separately from the bit-exact layout transformations.
注意力机制和前馈网络（FFN）的线性投影处理大规模 token 矩阵，是低精度 GEMM 加速的重要目标。其端到端收益还取决于量化、scale 布局转换以及周边数据移动的开销。因此，我们将算术内核与其生产者及通信接口一并评估。
Attention and FFN linear projections process large token matrices and are important targets for low-precision GEMM acceleration. Their end-to-end benefit also depends on the cost of quantization, scale layout conversion, and surrounding data movement. We therefore evaluate the arithmetic kernels together with their producers and communication interfaces.
Sol-H3 在已验证的 SM100 系列路径上，对零索引 Transformer 块 2–46（含两端）中的注意力机制和 FFN 投影使用 MXFP8 GEMM。权重和激活使用 E4M3 值，沿归约维度每 32 个值共享一个 E8M0 缩放因子。缩放缓冲区采用与 cuBLASLt 兼容的 SWIZZLE_32_4_4 布局。融合的 RMSNorm/调制和 SwiGLU 生产者直接写入 E4M3 激活和 swizzled 缩放，消除了中间 BF16 张量和单独的量化过程。边界块、AdaLN 投影、细化器、VAE 以及文本/音频编码器保留 BF16 计算。量化是近似的。
Sol-H3 uses MXFP8 GEMMs for attention and FFN projections in zero-indexed transformer blocks 2–46, inclusive, on the validated SM100-family path. Weights and activations use E4M3 values with one E8M0 scale for each group of 32 values along the reduction dimension. The scale buffers use the cuBLASLt-compatible SWIZZLE_32_4_4 layout. Fused RMSNorm/modulation and SwiGLU producers write the E4M3 activations and swizzled scales directly, eliminating the intermediate BF16 tensor and a separate quantization pass. Boundary blocks, AdaLN projections, refiners, VAE, and text/audio encoders retain BF16 computation. The quantization is approximate.
Sol-H3 复用注意力机制返回的原始 E4M3 值，作为量化输出投影的输入，并提供单位 E8M0 缩放因子。接收方将按 rank 排列的多头直接合并到投影的按行排列输入中，避免了中间的 BF16 张量以及随后的动态量化。将这种复用与融合的激活生成器相结合，消除了注意力机制、通信与线性计算之间冗余的数据转换。
Sol-H3 reuses the raw E4M3 values returned by attention as inputs to the quantized output projections, supplying unit E8M0 scales. The receiver merges the rank-major heads directly into the projection's row-major input, avoiding an intermediate BF16 tensor and subsequent dynamic quantization. Integrating this reuse with the fused activation producers removes redundant data conversions between attention, communication, and linear computation.
在精炼阶段，token 数量随分辨率显著增长，急剧增加了注意力机制的计算成本。因此，我们采用 Sol-Attn，一种无需训练的块稀疏注意力方法，它使用在线阈值化来选择一小部分 KV 块进行计算。具体而言，对于每个查询块，阈值设为 \(\mu + \tau \sigma\)，其中 \(\mu\) 和 \(\sigma\) 分别表示其在 KV 块上的块级注意力分数的均值和标准差，而 \(\tau\) 是控制稀疏度的可调超参数。只有分数超过该阈值的 KV 块才会被选中进行注意力计算；较大的 \(\tau\) 会提高阈值，从而保留更少的块。我们观察到，在早期高噪声精炼步骤中，稀疏化对全局图像连贯性的影响比后期低噪声步骤更大。因此，我们采用逐步依赖的调度策略，逐步增加稀疏度，将三个精炼步骤的 \(\tau\) 分别设为 \(1.0\)、\(1.25\) 和 \(1.5\)。为了适应不同的 GPU 架构，我们使用 cuDNN 块稀疏注意力（BSA）和自定义 CuTe DSL 内核作为替代后端，提高了内核效率和多 GPU 服务性能。
In the refinement stage, the token count grows substantially with resolution, sharply increasing the computational cost of attention. We therefore employ Sol-Attn, a training-free block-sparse attention method that uses online thresholding to select a small subset of KV blocks for computation. Specifically, for each query block, the threshold is set to \(\mu + \tau \sigma\), where \(\mu\) and \(\sigma\) denote the mean and standard deviation of its block-level attention scores across KV blocks, and \(\tau\) is a tunable hyperparameter that controls sparsity. Only KV blocks whose scores exceed this threshold are selected for attention computation; a larger \(\tau\) raises the threshold and thus retains fewer blocks. We observe that sparsification has a greater impact on global image coherence during early, high-noise refinement steps than during later, low-noise steps. We therefore adopt a step-dependent schedule that progressively increases sparsity, setting \(\tau\) to \(1.0\), \(1.25\), and \(1.5\) for the three refinement steps, respectively. To accommodate different GPU architectures, we use cuDNN block-sparse attention (BSA) and a custom CuTe DSL kernel as alternative backends, improving kernel efficiency and multi-GPU serving performance.
在实践中，我们在整个去噪过程中保持前两个 Transformer 层为稠密计算，因为在全局结构仍在形成时引入的误差可能会持续影响后续层。在剩余的注意力机制层中，我们为前缀查询和目标视频查询使用不同的注意力路径。MiniMax-H3 将每个序列打包为 \([\,\texttt{text}\mid\texttt{conditioning video}\mid\texttt{audio}\mid\texttt{target video}\,]\)，因此目标视频之前的所有 token 构成一个连续前缀。来自该前缀的查询通过稠密交叉注意力机制关注所有 KV token。其余目标视频查询使用块稀疏注意力机制，其中前缀 KV 块作为始终被关注的汇点，无论路由阈值如何都会被精确计算。重要的是，这种保护覆盖整个前缀，而不仅仅是文本 token：音频 token 本身也是生成的，模型在这些位置预测音频速度。在验证中，我们观察到即使是在所比较的变体中视觉质量得分最高的案例，也出现了严重的对话退化，这凸显了在保持视觉质量的同时保持音频保真度的必要性。这种设计保持了多模态前缀与完整序列之间的稠密连接，同时在目标视频 token 之间保留了块稀疏注意力机制。
In practice, we keep the first two Transformer layers dense throughout denoising, as errors introduced while global structure is still forming may persist through subsequent layers. Within the remaining attention layers, we use distinct attention paths for prefix and target-video queries. MiniMax-H3 packs each sequence as \([\,\texttt{text}\mid\texttt{conditioning video}\mid\texttt{audio}\mid\texttt{target video}\,]\), so all tokens preceding the target video form a contiguous prefix. Queries from this prefix attend to all KV tokens through dense cross-attention. The remaining target-video queries use block-sparse attention, with the prefix KV blocks serving as always-attended sinks that are evaluated exactly regardless of the routing threshold. Importantly, this protection covers the entire prefix rather than text tokens alone: the audio tokens are themselves generated, with the model predicting audio velocities at these positions. In validation, we observed severe dialogue degradation even in a case that achieved the highest visual-quality score among the compared variants, highlighting the need to preserve audio fidelity alongside visual quality. This design preserves dense connectivity between the multimodal prefix and the full sequence while retaining block-sparse attention among target-video tokens.
DiT 反复将宽中间张量写入 HBM，并在后续逐元素操作中将其读回。融合通过在寄存器中消费中间结果来减少这种流量。H3 结合了 RMSNorm、部分旋转嵌入、SwiGLU 和逐行索引调制，这要求对为其他架构设计的融合核进行适配。RSI 循环搜索实现和启动配置，分析其延迟，并对照相应的参考检查其数值行为。
The DiT repeatedly writes wide intermediate tensors to HBM and reads them back for subsequent elementwise operations. Fusion reduces this traffic by consuming intermediates in registers. H3 combines RMSNorm, partial rotary embeddings, SwiGLU, and per-row indexed modulation, requiring adaptations to fusion kernels designed for other architectures. The RSI loop searches implementation and launch configurations, profiles their latency, and checks their numerical behavior against the corresponding reference.
RSI 循环在 50 个 H3 Transformer 块中识别出三个算子组：(1) 带索引门控的残差加法、RMSNorm 以及带索引的缩放/平移调制；(2) 带部分 RoPE 的 QK 归一化；(3) SwiGLU 拆分、SiLU 与乘法。在具有代表性的四秩上下文并行配置（本地行数 9562、隐藏宽度 5376）下，这些融合减少了宽中间张量的重复读写。纯布局变换与其参考置换逐位相同。算术融合保留了算子结构，但中间精度和浮点舍入的变化可能导致无法与原始 eager 实现逐位等价；我们将这些与下文单独验证的 LoRA 消费者融合等价性区分开来。
The RSI loop identifies three operator groups across the 50 H3 transformer blocks: (1) residual addition with indexed gating, RMSNorm, and indexed scale/shift modulation; (2) QK normalization with partial RoPE; and (3) SwiGLU splitting, SiLU, and multiplication. At a representative four-rank context-parallel shape of 9562 local rows and hidden width 5376, these fusions reduce repeated reads and writes of wide intermediate tensors. Pure layout transformations are bit-identical to their reference permutations. Arithmetic fusions preserve the operator structure, but changes in intermediate precision and floating-point rounding can prevent bitwise equivalence to the original eager implementation; we distinguish these from the separately validated LoRA consumer-fusion equivalence below.
权重合并移除了 LoRA 的额外 GEMM，但可能改变少步适配器的数值行为。当 \(W+BA\) 以 BF16 存储时，即使合并本身以 FP32 计算，微小的更新也可能舍入回原始权重。因此，实数算术恒等式 \((W+BA)x=Wx+B(Ax)\) 并不意味着在 BF16 下等价。在对所评估的八步适配器进行的受控诊断中，六个检查的投影中 86.1–94.4% 的非零权重更新舍入回其原始值。这些是逐投影的权重统计量，而非模型范围的分数或感知质量损失；附录 B 给出了诊断控制。
Weight merging removes LoRA's extra GEMMs, but can change a few-step adapter's numerical behavior. Small updates can round back to the original weights when \(W+BA\) is stored in BF16, even if the merge itself is computed in FP32. The real-arithmetic identity \((W+BA)x=Wx+B(Ax)\) therefore does not imply equivalence in BF16. In a controlled diagnostic of the evaluated eight-step adapter, 86.1–94.4% of nonzero weight updates in six inspected projections rounded back to their original values. These are per-projection weight statistics, not a model-wide fraction or a perceptual-quality loss; Appendix B gives the diagnostic controls.
我们改为保留基分支和 LoRA 分支，并将其加法融合到下一个消费者内核中。每个投影仍以不变的权重和归约顺序执行相同的三个 GEMM。融合的消费者在 QK 归一化、RoPE 和打包、残差调制或 SwiGLU 之前，显式地将 \(\mathrm{base}+\delta\) 舍入为 BF16；最终的 FFN 消费者还保留中间的 BF16 门乘积。保留这些舍入点消除了宽求和的 HBM 写/读对，同时保留了分离分支的算术。该实现专门针对单位缩放下的单个活动适配器，使用 BF16 线性计算，与适配器名称无关。其数值一致性和运行时在保持其他加速设置不变的情况下，与原生分支进行评估，见第 6.6 节。
We instead retain the base and LoRA branches and fuse their addition into the next consumer kernel. Each projection still executes the same three GEMMs with unchanged weights and reduction order. The fused consumer explicitly rounds \(\mathrm{base}+\delta\) to BF16 before QK normalization, RoPE and packing, residual modulation, or SwiGLU; the final FFN consumer also retains the intermediate BF16 gate product. Preserving these rounding points removes the wide sum's HBM write/read pair while retaining the separate-branch arithmetic. The implementation specializes to one active adapter at unit scale with BF16 linear compute, independent of the adapter's name. Its numerical agreement and runtime are evaluated against native branches with the other acceleration settings held constant in Sec. 6.6.
潜在扩散需要解码到像素。本小节中的优化针对单阶段 H3 输出路径中所使用的 H3 视频解码器。它们与第 3.1 节中的潜在交接相互独立，后者绕过 H3 解码，并将视频潜在表示传递给 LTX 精化与解码路径。
Latent diffusion requires decoding to pixels. The optimizations in this subsection concern the H3 video decoder used in single-stage H3-output paths. They are separate from the latent handoff in Sec. 3.1, which bypasses H3 decoding and passes video latents to the LTX refinement and decoding path.
上下文并行将 Transformer 堆栈分片，但其本身并不对视频解码进行划分。在所评估的 H3 配置中，否则每个 rank 都会解码完整视频。在 \(1344 \times 768\) 分辨率和 124 个输出帧下，H3 解码器处理七个时间片段，每个片段包含 28 个等大小的空间瓦片，共 196 次瓦片解码。将这些相互独立的瓦片分配到八个 rank 上，可将记录的解码延迟从 7.55 s 降至约 1.16 s，同时保留原始的拼接顺序。
Context parallelism shards the Transformer stack but does not by itself partition video decoding. In the evaluated H3 configuration, each rank would otherwise decode the full video. At \(1344 \times 768\) and 124 output frames, the H3 decoder processes seven temporal clips with 28 equal-sized spatial tiles per clip, for 196 tile decodes. Distributing these independent tiles over eight ranks reduces the recorded decode latency from 7.55 s to approximately 1.16 s while retaining the original stitch order.
分片使得每个 rank 在每个时间片段中拥有四个瓦片槽位。将这四个瓦片作为一个本地批次进行解码，而不是为每个瓦片单独启动解码器，可将延迟从 1.1597 秒降低到 0.6021 秒，在记录的配置中实现了 \（1.93\times）的加速。编译批处理解码器后达到 0.3593 秒，相对于 1.1597 秒的分片基线实现了 \（3.23\times）的加速，并将测得的解码器峰值内存从 18.8 GB 降低到 15.6 GB。这些是解码级别的测量结果，而非整个流水线的加速。编译与即时解码并非逐位相同；记录的最大绝对偏差约为 0.021。
Sharding leaves four tile slots per rank in each temporal clip. Decoding these four tiles as one local batch, rather than launching the decoder separately for each tile, reduces latency from 1.1597 s to 0.6021 s, a \(1.93\times\) speedup in the recorded configuration. Compiling the batched decoder reaches 0.3593 s, or \(3.23\times\) relative to the 1.1597 s sharded baseline, and reduces the measured decoder peak memory from 18.8 GB to 15.6 GB. These are decode-level measurements rather than whole-pipeline speedups. Compilation is not bit-identical to eager decoding; the recorded maximum absolute deviation is approximately 0.021.
一项独立的优化将全部七个时间片段的瓦片合并在一起。这样，196 个真实瓦片占据 200 个执行槽位，即每个 rank 25 个，仅需四个填充副本，而按片段调度则需要 224 个槽位和 28 个副本。全局路径只需调用一次批处理解码器和 gather，而非七次，并在之后恢复原始的瓦片与混合顺序。改变批形状可能改变所选的 GEMM 实现：与按片段路径的对比记录显示，在 \([-1,1]\) 范围内最大绝对偏差为 0.013672，解码器输出 PSNR 为 75.75 dB。这些数值仅表征该次测试，而非一般误差界或感知质量保证。此处讨论的测量与实现使用八个 rank；它们并不能证明在单块 GPU 上具有相同的批处理收益。
A separate optimization combines tiles across all seven temporal clips. The 196 real tiles then occupy 200 execution slots, or 25 per rank, with only four padded duplicates, instead of 224 slots and 28 duplicates for per-clip scheduling. The global path invokes the batched decoder and gather once rather than seven times and restores the original tile and blend order afterward. Changing the batch shape can change the selected GEMM implementation: the recorded comparison against the per-clip path has maximum absolute deviation 0.013672 and decoder-output PSNR 75.75 dB over the \([-1,1]\) range. These values characterize that test, rather than a general error bound or perceptual-quality guarantee. The measurements and implementation discussed here use eight ranks; they do not establish the same batching benefit on a single GPU.
尽管由于巨大的计算需求，云环境是主要的部署载体，但在本地或边缘设备上执行这些生成模型仍有大量需求。为了实现高效的本地部署，我们开展了一系列全面的优化，旨在严格降低延迟和设备内存占用。该场景下的关键约束是设备内存驻留。它要求整个工作集，包括文本编码器、草稿模型、精炼器和 VAE，必须同时驻留在 HBM 中。如果工作集无法容纳，运行时必须将权重卸载到主机内存并按需流式加载回来，这会产生显著的主机-设备通信开销，并主导请求延迟。
While cloud environments serve as the primary deployment vehicle due to massive computational demands, there is substantial demand for executing these generative models locally or on edge devices. To enable efficient local deployment, we undertake a comprehensive set of optimizations targeting strict reductions in both latency and device memory footprint. The binding constraint in this regime is device memory residency. It requires that the entire working set, including the text encoder, drafting model, refiner, and VAE, must remain simultaneously resident in HBM. If the working set does not fit, the runtime must offload weights to host memory and stream them back on demand, incurring a substantial host-device communication cost that dominates request latency.
在单台 DGX Spark 上，CPU 和 GPU 共享 119.68 GiB 的统一内存池，我们的优化代表了可行部署与内存不足故障之间的关键差异。具体而言，朴素的两阶段部署需要 142.4 GiB，超出硬件容量 22.7 GiB。我们的优化配置仅消耗 116.9 GiB，在采样期间保留了 2.8 GiB 的安全余量。
On a single DGX Spark, where the CPU and GPU share a 119.68 GiB unified memory pool, our optimizations represent the critical difference between a viable deployment and an out-of-memory failure. Specifically, a naive two-stage deployment demands 142.4 GiB, exceeding the hardware capacity by 22.7 GiB. Our optimized configuration consumes only 116.9 GiB, preserving a 2.8 GiB safety margin during sampling.
在本节中，我们详细介绍了为实现这些延迟和内存目标而应用的具体优化技术，重点关注确定性调制缓存、提示缓存和权重量化等策略。最终，在单加速器边缘环境（如 GB10）上部署这一优化流水线，成功将完整的生成栈适配到设备上，同时提供高响应性的本地推理。
In this section, we detail the specific optimization techniques applied to achieve these latency and memory objectives, focusing on strategies such as deterministic modulation caching, prompt caching, and weight quantization. Ultimately, deploying this optimized pipeline on a single-accelerator edge environment, such as a GB10, successfully fits the complete generative stack onto the device while delivering highly responsive local inference.
对于 MiniMax-H3，模型 33B 参数中约有 13B 位于 adaln_proj 模块。在推理时，该模块会在每个去噪步骤重新计算调制表，每个生成的视频累计达 2500 次重算。由于该投影的唯一输入是时间步嵌入，而该嵌入仅依赖于去噪循环前已固定的采样调度，因此这些重算产生的每一个值都是事先已知的。所以，这种重复是纯粹冗余的，该模块在两个维度上造成浪费：设备内存被占用，因为宽投影权重必须常驻；算力和带宽被消耗，因为每个步骤都在重新生成同一张表。我们的运行时一次性预计算整张表，并维护一个以（block, step）为索引的缓存。这消除了两个严重瓶颈。第一，节省约 24 GB 设备内存，因为用约 1.5 GB 的预计算表替换 26 GB 的 adaln_proj 权重，使去噪器常驻内存从 61.7 GB 降至约 37 GB。第二，消除每步约 26 GB 的 HBM 读取；参考流水线每个 block 跨 HBM 流式传输 520 MB 权重数据，仅产生九行输出，带来带宽税且算术强度极低。
For MiniMax-H3, approximately 13B of the model's 33B parameters reside in the adaln_proj module. At inference time, this module recomputes the modulation table at every denoising step, amounting to 2500 recomputations per generated video. Because the projection's only input is the timestep embedding, and that embedding depends solely on the sampling schedule fixed before the denoising loop, every one of those recomputations produces a value that was already knowable in advance. Consequently, this repetition is pure redundancy, with the module imposing waste along two dimensions: device memory is consumed because the wide projection weights must stay resident, and compute and bandwidth are expended regenerating the same table at each step. Our runtime precomputes the entire table once and maintains a cache indexed by (block, step). This eliminates two severe bottlenecks. First, it saves about 24 GB of device memory, as replacing the 26 GB of adaln_proj weights with a precomputed table of \(\approx 1.5\) GB lowers denoiser residency from 61.7 GB to \(\approx 37\) GB. Second, it eliminates about 26 GB of HBM reads per step; the reference pipeline streams 520 MB of weight data across HBM per block to produce just nine output rows, incurring a bandwidth tax with minimal arithmetic intensity.
如第 3.1 节所述，Stage 1 已经确定了生成视频的全局布局、语义内容和运动。Stage 2 的作用主要是局部高频细化。因此，Stage 2 不需要针对每个样本的文本条件：它所需的样本特定信息已经由接收到的 Stage-1 潜在表示携带，一个通用的、与提示无关的条件信号就足够了。
As established in Sec. 3.1, Stage 1 has already fixed the global layout, semantic content, and motion of the generated video. The role of Stage 2 is mainly local high-frequency refinement. Therefore, Stage 2 does not require per-sample text conditioning: the sample-specific information it needs is already carried by the Stage-1 latent it receives, and a generic, prompt-independent conditioning signal suffices.
我们不是为 Stage 2 动态编码用户提示，而是离线一次性评估一个单一的规范通用细化提示：4K、精细、高质量、电影细节、干净纹理、自然运动。该提示通过 INT8 Gemma 文本编码器和 INT8 开发连接器处理，我们仅保留紧凑的连接器后视频和音频上下文张量，作为缓存馈送到 Stage 2。因此，在线推理期间永远不会加载 Stage-2 文本编码器或其连接器，从而释放 16.2 GiB 供其他模块使用。
Instead of encoding the user's prompt dynamically for Stage 2, we evaluate a single canonical generic refinement prompt once and offline, 4K, refined, high quality, cinematic detail, clean textures, natural motion. This prompt is processed through the INT8 Gemma text encoder and the INT8 dev connector, and we retain only the compact post-connector video and audio context tensors to feed to Stage 2 as a cache. Consequently, neither the Stage-2 text encoder nor its connector is ever loaded during online inference, freeing the 16.2 GiB for other modules.
为了在单台 DGX Spark（GB10）上部署，量化对于将两阶段流水线常驻在设备的 119.68 GiB CPU–GPU 统一内存池内至关重要。边缘配置将一个 FP8 H3 草稿 DiT 与一个 NVFP4 AWQ Qwen 提示编码器配对，报告的常驻内存占用分别为 23.4 GiB 和 14.6 GiB。运行时从检查点读取每一层的量化配置，允许不同的投影使用不同的精度设置。Qwen 在在线推理期间对每个传入的提示进行编码。INT8 Gemma 编码器和连接器仅离线用于构建第 4.2 节所述的缓存 Stage-2 条件，并且在请求处理期间两者均不保持常驻。
For deployment on a single DGX Spark (GB10), quantization is essential to keeping the two-stage pipeline resident within the device's 119.68 GiB CPU–GPU unified memory pool. The edge configuration pairs an FP8 H3 draft DiT with an NVFP4 AWQ Qwen prompt encoder, with reported resident footprints of 23.4 GiB and 14.6 GiB, respectively. The runtime reads each layer's quantization configuration from the checkpoint, allowing different projections to use different precision settings. Qwen encodes each incoming prompt during online inference. The INT8 Gemma encoder and connector are used only offline to construct the cached Stage-2 conditioning described in Sec. 4.2, and neither remains resident during request processing.
在许多实际部署场景中，生成过程以用户提供的参考材料为条件。在这种参考到视频与音频（ref2av）设定下，模型基于一组条件参考图像或视频合成输出视频和音频。随着参考材料数据量的增长，参考令牌与生成令牌之间的交互成为生成延迟的瓶颈。因此，我们提出一种参考令牌缓存策略，以进一步降低该设定下的延迟。
In many practical deployment scenarios, generation is conditioned on user-provided reference materials. In this reference-to-video and audio (ref2av) setting, the model synthesizes output video and audio based on a set of conditioning reference images or videos. As the volume of reference materials data grows, the interaction between reference tokens and generation tokens becomes a bottleneck of generation latency. Thus, we propose a reference token caching strategy to further reduce latency in this setting.
参考计算冗余。在标准扩散推理过程中，模型执行 \(N\) 个去噪步骤。在朴素实现中，这 \(N\) 个步骤中的每次自注意力评估都会完整重新计算参考令牌，不跳过任何计算。然而，与迭代去噪的目标视频令牌不同，输入到 Diffusion Transformer（DiT）的参考潜变量是干净的潜变量。由于参考输入未添加噪声，其表示随时间步嵌入的变化很微弱。因此，在每个去噪步骤中为参考令牌重新计算完整的自注意力会引入潜在的计算冗余。
Reference computation redundancy. During the standard diffusion inference process, the model executes \(N\) denoising steps. In a naive implementation, every self-attention evaluation across these \(N\) steps recomputes the reference tokens in full, with no computations skipped. However, unlike the target video tokens which are iteratively denoised, the reference latent fed into the Diffusion Transformer (DiT) is a clean latent. Because no noise is added to the reference inputs, their representations vary only weakly with the timestep embedding. Consequently, recomputing the full self-attention for the reference tokens at every denoising step introduces potential computational redundancy.
重条件生成。当处理大量参考输入时，这种冗余会成为严重瓶颈。单个生成请求可能提供多达 9 张参考图像或 3 个参考视频。在这种重条件负载下，参考令牌的数量很容易超过生成令牌本身的数量。在这种规模下，参考令牌的冗余重计算不再是边际开销，而是主导了整体推理成本，导致计算资源的显著浪费。
Generation with heavy conditioning. This redundancy becomes a severe bottleneck when processing extensive reference inputs. A single generation request may supply up to 9 reference images or 3 reference videos. Under these heavy conditioning workloads, the number of reference tokens can easily exceed the number of generation tokens themselves. At this scale, the redundant recomputation of reference tokens ceases to be a marginal overhead and instead dominates the overall inference cost, leading to a significant waste of computational resources.
为消除这种低效，我们引入了一种参考 KV 缓存机制。其核心原理是仅在第一个 DiT 步骤中完整评估参考分支。在此初始步骤中，计算并缓存参考 token 的键（K）和值（V）张量。对于所有后续去噪步骤，不再重新计算参考 token；相反，其缓存的 KV 张量被保留并原样重用（缓存的 KV 从不迭代或刷新）。后续步骤仅评估生成的 token，这些 token 关注缓存的参考 KV。由于第一步保持完整计算，条件信号被精确建立，仅跳过了后续步骤中的冗余重复。
To eliminate this inefficiency, we introduce a reference KV caching mechanism. The core principle is to evaluate the reference branch in full only during the first DiT step. During this initial step, the key (K) and value (V) tensors for the reference tokens are computed and cached. For all subsequent denoising steps, the reference tokens are no longer recomputed; instead, their cached KV tensors are retained and reused unchanged (the cached KV is never iterated or refreshed). The subsequent steps only evaluate the generated tokens, which attend to the cached reference KV. Because the first step remains a full computation, the conditioning signal is established exactly, and only the redundant repetitions in later steps are bypassed.
我们注意到，这种缓存策略是一种近似，而非逐位相同的变换。尽管输入的参考潜变量是干净的，但由于 DiT 块中时间步嵌入的变化，其中间表示会自然地在步骤间轻微漂移。然而，从定性比较中，我们发现忠实重计算流程与缓存流程显示出微小偏差，表明缓存策略是一种实用的近似方法，用于参考到视频生成，尤其是在参考材料繁重时。
We note that this caching strategy is an approximation rather than a bitwise-identical transformation. Although the input reference latent is clean, its intermediate representations naturally drift slightly across steps due to the changing timestep embeddings in the DiT blocks. However, from the qualitative comparison, we find that the faithful recomputation pipeline and caching pipeline show minor deviation, demonstrating that the caching strategy is a practical approximation method for reference to video generation especially when reference material is heavy.
我们在单块 NVIDIA GB200 上评估了两阶段跨分辨率流水线的结果。对于 1344 \(\times\) 768 分辨率的 5 秒工作负载，该流水线实现了 6.852 秒的端到端延迟，其中 Stage 1 耗时 4.173 秒，Stage 2 耗时 2.679 秒。相较于 SGLang 基线 152.3 秒，实现了 22.2 \(\times\) 的加速。对于相同 1344 \(\times\) 768 分辨率的 10 秒工作负载，端到端延迟为 14.931 秒（Stage 1 为 9.820 秒，Stage 2 为 5.111 秒），相较于 SGLang 基线 414.1 秒，实现了 27.7 \(\times\) 的加速。
We evaluate the two-stage cross-resolution pipeline results measured on a single NVIDIA GB200. For a 5 s workload at 1344 \(\times\) 768 resolution, the pipeline achieves an end-to-end latency of 6.852 s, comprising 4.173 s for Stage 1 and 2.679 s for Stage 2. This represents a 22.2 \(\times\) speedup over the SGLang baseline of 152.3 s. For a 10 s workload at the same 1344 \(\times\) 768 resolution, the end-to-end latency is 14.931 s (9.820 s for Stage 1 and 5.111 s for Stage 2), yielding a 27.7 \(\times\) speedup over the SGLang baseline of 414.1 s.
对于边缘设备，单台 DGX Spark 部署在相同的 5 秒 1344 \(\times\) 768 工作负载下端到端达到 56.17 秒，相较于约 1740.81 秒的基线，实现了 31 \(\times\) 的加速。该部署的延迟占比分解如下（图 5）：LTX 3 步细化占 44.0%，H3 4 步生成占 35.2%，LTX VAE 解码占 12.5%，H3 上采样器和 VAE 适配器占 3.8%，Qwen 提示编码占 2.4%。生成帧的定性比较见图 6。
For edge device, the single DGX Spark deployment reaches 56.17 s end to end for the same 5 s 1344 \(\times\) 768 workload, yielding a 31 \(\times\) speedup over the baseline of roughly 1740.81 s. The latency share breakdown for this deployment is as follows (Fig. 5): LTX 3-step refinement accounts for 44.0%, H3 4-step generation for 35.2%, LTX VAE decode for 12.5%, H3 upscaler and VAE adapter for 3.8%, Qwen prompt encoding for 2.4%. A qualitative comparison of the generated frames is provided in Fig. 6.
相同的两阶段流水线也可在单块消费级 GPU 上运行。在一块 RTX 5090 上，5 秒 \(1344\times 768\) 请求在 39.80 秒内完成，而同一显卡上测得的全分辨率 49 步基线为 1045.40 秒，实现了 26.3 \(\times\) 的加速。其延迟构成与数据中心和 DGX Spark 部署均不同：Qwen 提示编码是最大的单一模块，占 34.1%，高于 H3 4 步生成的 28.2% 和 LTX 3 步细化的 19.7%，LTX VAE 解码占 5.1%，上采样器和潜在适配器占 2.8%。文本编码器占主导并非因为编码本身昂贵，而是因为编码器在每个请求时都会被重建：在该模块的 13.568 秒中，NVFP4 AWQ Qwen 前向传播仅占 0.518 秒，其余为加载和卸载。与一次性启动成本不同，这一开销在每个请求上都会产生，因为 5090 无法将编码器与两个生成阶段同时保存在 GPU HBM 中，因此运行在 CPU 卸载模式下。因此，在消费级硬件上，去噪本身已不再是限制项；提示编码器的驻留才是，这与第 4.2 节中 Stage-2 条件缓存为 DGX Spark 缓解的压力相同。
The same two-stage pipeline also runs on a single consumer GPU. On one RTX 5090 the 5 s \(1344\times 768\) request completes in 39.80 s, against a 1045.40 s full-resolution 49-step baseline measured on the same card, a 26.3 \(\times\) speedup. The latency composition differs from both datacenter and DGX Spark deployments: Qwen prompt encoding is the single largest block at 34.1%, ahead of H3 4-step generation at 28.2% and LTX 3-step refinement at 19.7%, with LTX VAE decode at 5.1% and the upscaler and latent adapter at 2.8%. The text encoder dominates not because encoding is expensive but because the encoder is rebuilt for every request: of the 13.568 s attributed to this block, the NVFP4 AWQ Qwen forward pass accounts for 0.518 s and the rest is loading and teardown. Unlike a one-time startup cost, this is charged on each request, because the 5090 cannot hold the encoder in GPU HBM together with both generation stages and therefore runs under CPU offload. Denoising itself is therefore no longer the limiting term on consumer hardware; prompt-encoder residency is, which is the same pressure the Stage-2 conditioning cache in Sec. 4.2 relieves on DGX Spark.
我们在 1344 \(\times\) 768 分辨率和 24 FPS 下，在 1、4 和 8 块 NVIDIA GB200 GPU 上评估 Sol-H3，生成无参考视频及立体声音频。所有九种 GPU 数量与输出时长的组合均使用相同的提示词、种子 20260903 和合并的四步 LoRA。所评估的系统结合了 SOL/BSA 注意力机制、量化 QKV 与注意力输出通信、带有融合激活生成器和输出复用的 MXFP8 GEMM，以及第 3 节所述的 VAE 优化。
We evaluate Sol-H3 on one, four, and eight NVIDIA GB200 GPUs at 1344 \(\times\) 768 resolution and 24 FPS, generating reference-free video with stereo audio. All nine combinations of GPU count and output duration use the same prompt, seed 20260903, and merged four-step LoRA. The evaluated system combines SOL/BSA attention, quantized QKV and attention-output communication, MXFP8 GEMMs with fused activation producers and output reuse, and the VAE optimizations described in Sec. 3.
NVIDIA GB200，1344 \(\times\) 768，24 FPS。延迟单位为秒；加速比 = Base H3 / Sol-H3。Base H3：49 次 DiT 前向；Sol-H3：4 次 DiT 前向。
NVIDIA GB200, 1344 \(\times\) 768, 24 FPS. Latency in seconds; speedup = Base H3 / Sol-H3. Base H3: 49 DiT forwards; Sol-H3: 4 DiT forwards.
每个 Sol-H3 延迟是两次预热后五次调用的中位数，每次调用验证四次去噪评估。计时涵盖文本编码、去噪、视频/音频 VAE 解码和同步，不包括模型加载、编译、预热和 MP4 编码。单 GPU 执行使用相同的 SOL/BSA 策略和本地 QKV/输出量化，无分布式集合通信。它使用编译的逐瓦片 VAE 解码器；多 GPU 执行使用编译的、全局批处理的分布式解码器。Base H3 参考使用 49 次去噪评估。
Each Sol-H3 latency is the median of five calls after two warmups, with four denoising evaluations verified per call. Timing covers text encoding, denoising, video/audio VAE decoding, and synchronization, excluding model loading, compilation, warmup, and MP4 encoding. Single-GPU execution uses the same SOL/BSA policy and local QKV/output quantization without distributed collectives. It uses the compiled per-tile VAE decoder; multi-GPU execution uses the compiled, globally batched distributed decoder. The Base H3 reference uses 49 denoising evaluations.
如表 1 所示，Sol-H3 在评估的工作负载上相比 Base H3 实现了 12.73 \(\times\) –16.99 \(\times\) 的加速。在八块 GB200 GPU 上，五秒、十秒和十五秒视频分别耗时 1.434 秒、3.350 秒和 6.062 秒。五秒工作负载的生成速度约为实时播放的 3.5 \(\times\) 倍。
As shown in Tab. 1, Sol-H3 achieves 12.73 \(\times\) –16.99 \(\times\) speedup over Base H3 across the evaluated workloads. On eight GB200 GPUs, five-, ten-, and fifteen-second videos take 1.434 s, 3.350 s, and 6.062 s, respectively. The five-second workload is generated approximately 3.5 \(\times\) faster than real-time playback.
参考条件显著增加了计算开销。如表 2（左）所示，与无参考生成过程相比，一张图像增加了 14,344 个有效参考 token，并将 DiT 时间提升至 1.71 \(\times\)；一个视频增加了 31,792 个 token，达到 2.83 \(\times\)；而一个视频与一张图像共同作为参考时，增加了 46,136 个 token，达到 3.94 \(\times\)。随着参考材料增多，开销由参考 token 数量决定，这与注意力机制随序列长度呈二次增长的趋势一致。参考 KV 缓存减少了重复的参考分支计算，实现了 1.14 \(\times\)–1.92 \(\times\) 的加速，如表 2（右）所示。这种收益随着每种模态内参考数量的增加而增长。
Reference conditioning substantially increases computation overhead. As shown in Table 2 (left), one image adds 14,344 effective reference tokens and raises the DiT time to 1.71 \(\times\) compared with the reference-free generation process; one video adds 31,792 tokens for 2.83 \(\times\), and a video together with an image adds 46,136 tokens for 3.94 \(\times\). Cost is governed by the reference-token count as the reference materials grow, consistent with the quadratic growth of attention in sequence length. Reference KV caching reduces repeated reference-branch computation, yielding speedups of 1.14 \(\times\)–1.92 \(\times\), demonstrated in Table 2 (right). The benefit grows with the number of references within each modality.
尽管缓存策略是一种近似方法，而非无损的系统优化，但定性比较表明，这种近似仅引入了轻微的感知差异。如图 7 所示，我们可以看到，与全计算基线相比，使用缓存策略生成的样本保持了场景结构、主体身份和运动连续性，并严格遵循视频到视频编辑的指令。
Although the caching strategy is an approximation method instead of a lossless system optimization, the qualitative comparison demonstrates that this approximation only introduces minor perceptual differences. As shown in Figure 7, we can see that the generated samples with caching strategy preserve scene structure, subject identity and motion continuity against the full-compute baseline and strictly follow the instruction of video-to-video editing.
我们评估了 120 个文本到视频和音频（T2VA）任务以及 120 个参考到视频和音频（Ref2VA）任务，涵盖多样化的主题、视觉风格、动作和镜头运动。Ref2VA 包括单参考和多参考输入、直接视频编辑，以及身份、运动或镜头行为的迁移。候选模型按任务类型和函数评估次数（NFE）分组，其中 T2VA 在 4 和 8 NFE 下分别有四个和六个候选，Ref2VA 有两个和三个候选。对每个任务进行组内穷举配对，共产生 3,000 对比较：这四个组中分别为 720、1,800、120 和 360 对。除了下文讨论的适配器外，T2VA 池还包括 4 NFE 下的 FastH3 Preview v1 LoRA 和 8 NFE 下的 FastH3 V2 完整蒸馏检查点基线。对于每个任务，比较的运行使用冻结的重写提示、参考资产以及相同的初始视频/音频噪声。
We evaluate 120 text-to-video-and-audio (T2VA) tasks and 120 reference-to-video-and-audio (Ref2VA) tasks, spanning diverse subjects, visual styles, actions, and camera movements. Ref2VA includes single- and multiple-reference inputs, direct video editing, and transfer of identity, motion, or camera behavior. Candidates are grouped by task type and number of function evaluations (NFE), with four and six candidates for T2VA at 4 and 8 NFE, respectively, and two and three candidates for Ref2VA. Exhaustive within-group pairing for every task yields 3,000 pairwise comparisons: 720, 1,800, 120, and 360 in these four groups. In addition to the adapters discussed below, the T2VA pool includes FastH3 Preview v1 LoRA at 4 NFE and the FastH3 V2 full distilled-checkpoint baseline at 8 NFE. For each task, the compared runs use the frozen rewritten prompt, reference assets, and the same initial video/audio noise.
我们将对部分配对的人类评估与对全部 3,000 对池的 VLM 评估相结合，同时保持两种判断来源相互独立。两种协议均隐藏模型身份，随机化 A/B 呈现顺序，并允许平局。人类评分者可以播放带音频的生成视频，并提供整体偏好。对于仅视觉的 VLM 评估，采用高推理强度的 Astra，接收生成指令、每个候选视频中均匀采样的 16 帧以及可用的视觉参考材料；参考视频同样被采样为 16 帧。不提供音频也不对其进行评分，音频相关的提示要求被排除在其判断之外。每对在一个全新的上下文中获得一次 VLM 判断。
We combine human assessment on a subset of pairs with VLM assessment of the full 3,000-pair pool, keeping the two sources of judgments separate. Both protocols hide model identities, randomize the A/B presentation, and allow ties. Human raters can play the generated videos with audio and provide an overall preference. For the visual-only VLM assessment, Astra with high reasoning effort receives the generation instruction, 16 uniformly sampled frames from each candidate video, and the available visual reference material; reference videos are likewise sampled into 16 frames. No audio is supplied or scored, and audio-related prompt requirements are excluded from its judgment. Each pair receives one VLM judgment in a fresh context.
我们的比较结果倾向于在 4 NFE 的 T2VA 中使用 Larry 的 Turbo LoRA（v4，第 600 步）和 LightX2V FL2VA Turbo（4 步 v1.2），在 8 NFE 中使用阿里巴巴 PAI PDD Acc8 和官方 VDN-H3 stage-dmd-step-250 适配器。对于 Ref2VA，阿里巴巴 PAI 的 4-NFE PDD 方案和 LightX2V Ref2VA Turbo（v0.1）在 4 NFE 下表现良好，而公开的 HyperFlow v1.0 版本是我们在 8 NFE 下的首选。
Our comparisons favor Larry’s Turbo LoRA (v4, step 600) and LightX2V FL2VA Turbo (4-step v1.2) for T2VA at 4 NFE, and Alibaba PAI PDD Acc8 and the official VDN-H3 stage-dmd-step-250 adapter at 8 NFE. For Ref2VA, both Alibaba PAI’s 4-NFE PDD recipe and LightX2V Ref2VA Turbo (v0.1) perform well at 4 NFE, while the public HyperFlow v1.0 release is our preferred choice at 8 NFE.
我们将适配器与两阶段生成器分开评估，以单独衡量转换成本和重建保真度。质量评估使用 256 个留出测试视频，与训练、开发及早期评估队列均不重叠。预测潜在表示和配对的 LTX 教师潜在表示均使用同一个冻结的 Conv Video VAE 进行解码。我们报告解码后视频片段的平均逐视频 PSNR，以及在 16 个均匀采样帧上的 SSIM。这些指标衡量的是与源视频的 LTX 重建结果的一致性，且在任何 Stage-2 去噪之前。
We evaluate the adapter separately from the two-stage generator to isolate conversion cost and reconstruction fidelity. The quality evaluation uses 256 held-out test videos, disjoint from training, development, and earlier evaluation cohorts. Both the predicted latent and the paired LTX teacher latent are decoded with the same frozen Conv Video VAE. We report mean per-video PSNR over the decoded clips and SSIM over 16 uniformly sampled frames. These metrics measure agreement with the LTX reconstruction of the source video, before any Stage-2 denoising.
如表 3 所示，最终适配器达到 30.618 dB PSNR 和 0.8962 SSIM。相对于相同架构的早期检查点，配对 PSNR 增益为 1.205 dB，95% 自助法区间为 \([1.166, 1.245]\) dB；全部 256 个片段在 PSNR 和 SSIM 上均有提升。最高运动四分位增益为 1.366 dB，而最低运动四分位为 1.059 dB。这一改进来自持续的解码器感知训练，使用了更大的学习率、批量大小和曝光预算，且未增加推理时的模型规模。它并未分离出每项训练改动的贡献。
As shown in Tab. 3, the final adapter reaches 30.618 dB PSNR and 0.8962 SSIM. Relative to the earlier checkpoint with the same architecture, the paired PSNR gain is 1.205 dB, with a 95% bootstrap interval of \([1.166, 1.245]\) dB; all 256 clips improve in both PSNR and SSIM. The highest-motion quartile gains 1.366 dB, compared with 1.059 dB in the lowest-motion quartile. This improvement comes from continued decoder-aware training with a larger learning rate, batch size, and exposure budget, without increasing inference-time model size. It does not isolate the contribution of each training change.
对于转换成本，表 4 在单张 H100 80GB 上以 BF16 对批量大小为 1、192 帧、分辨率为 \(1344 \times 768\) 的片段进行了性能分析。适配器（包括其固定的几何变换）耗时 63.1 ms，而 H3 解码后接 LTX 编码耗时 12,748.6 ms，中位转换延迟降低了 \(202.0\times\)。峰值分配内存从 13.79 GiB 降至 0.785 GiB。在相同输入几何下，Tiny AutoEncoder 往返耗时 201.9 ms。此性能分析使用的是同一 194.76M 参数架构的早期检查点，而表 3 评估的是最终训练权重。测量排除了空间上采样、去噪、最终视频解码和音频处理；因此它们量化的是交接模块的成本，而非端到端加速或对完整 VAE 往返的质量等效替代。
For conversion cost, Tab. 4 profiles a batch-one, 192-frame clip at \(1344 \times 768\) on one H100 80GB in BF16. The adapter, including its fixed geometry transform, takes 63.1 ms versus 12,748.6 ms for H3 decoding followed by LTX encoding, a \(202.0\times\) reduction in median conversion latency. Peak allocated memory falls from 13.79 to 0.785 GiB. A Tiny AutoEncoder round trip takes 201.9 ms under the same input geometry. This profile uses an earlier checkpoint of the same 194.76M-parameter architecture, whereas Tab. 3 evaluates the final trained weights. The measurements exclude spatial upsampling, denoising, final video decoding, and audio processing; they therefore quantify the cost of the handoff module, not an end-to-end speedup or a quality-equivalent replacement for the full VAE round trip.
我们首先验证 consumer fusion 能够保持所选 adapter 的计算结果。在使用所评估的八步 adapter 与 BF16 线性计算时，consumer fusion 在三个 15 秒样本上，与原生分离分支逐字节一致，覆盖全部八次去噪评估以及四个 GPU rank，包括最终输出的原始视频与音频张量。两种模式均保留相同的其他加速设置，包括稀疏注意力机制与低精度注意力传输。这确立了在测试输入与软件环境下的结果一致性，而非与完全未加速模型的等价性。
We first verify that consumer fusion preserves the selected adapter's computation. With the evaluated eight-step adapter and BF16 linear compute, consumer fusion matches native separate branches byte for byte on three 15-second samples, across all eight denoising evaluations and four GPU ranks, including the final raw video and audio tensors. Both modes retain the same other acceleration settings, including sparse attention and low-precision attention transport. This establishes agreement for the tested inputs and software environment, rather than equivalence to an entirely unaccelerated model.
我们使用一个固定的八步适配器和一个 15 秒的动画场景来隔离执行模式，场景中一个女孩遇到一只打喷嚏喷出蝴蝶的恐龙。图 8 中的三个分支共享提示词、随机种子、实际初始噪声、条件、采样计划以及其他运行时优化。权重合并改变了恐龙的外观、女孩的发型、它们的相对位置以及蝴蝶爆发的时间。放大视图还显示，在合并权重输出中蝴蝶周围出现网格状颜色图案，这些图案在视频编码前的解码帧中可见。原生分支和消费者融合保留了相同的生成场景：重跑验证了在所有四个 rank 上的每次去噪评估中、在最终解码的视频和音频中、以及在所有 362 个导出帧中的字节一致性。这个例子说明了数值变化如何改变生成轨迹；它不是学习到的适配器之间的比较，也不是定量的感知质量评估。
We isolate the execution mode using a fixed eight-step adapter and a 15-second animated scene in which a girl meets a dinosaur that sneezes butterflies. The three arms in Fig. 8 share the prompt, seed, actual initial noise, conditioning, sampling schedule, and other runtime optimizations. Weight merging changes the dinosaur's appearance, the girl's hairstyle, their relative positions, and the timing of the butterfly burst. The magnified view also shows grid-like color patterns around the butterflies in the merged-weight output, visible in the decoded frames before video encoding. Native branches and consumer fusion preserve the same generated scene: the rerun verifies byte agreement at every denoising evaluation on all four ranks, in the final decoded video and audio, and in all 362 exported frames. This example illustrates how a numerical change can alter a generation trajectory; it is not a comparison between learned adapters or a quantitative perceptual-quality evaluation.
在本报告中，我们提出了 Sol-H3，一个面向 MiniMax-H3 视频-音频生成器的全栈高效推理流水线。通过将跨分辨率两阶段生成流水线与通过递归自我改进（RSI）发现的核级优化栈相结合，Sol-H3 在云端和边缘环境中均实现了前所未有的生成速度。
In this report, we present Sol-H3, a full-stack efficient inference pipeline for the MiniMax-H3 video-audio generator. By combining a cross-resolution two-stage generation pipeline with a kernel-level optimization stack discovered via Recursive Self-Improvement (RSI), Sol-H3 achieves unprecedented generation speeds across both cloud and edge environments.
Sol-H3 开发过程中的一个关键启示是，人类架构洞察与自动化优化之间的互补关系。在我们的工作中，宏观层面的推理框架从根本上由人类设计驱动，而微观层面的实现优化则由 RSI 循环系统性地执行。
A key takeaway from the development of Sol-H3 is the complementary relationship between human architectural insight and automated optimization. In our work, the macro-level inference framework is fundamentally driven by human design, whereas the micro-level implementation optimizations are systematically executed by the RSI loop.
具体而言，RSI 被部署用于处理那些优化空间巨大但结果严格可验证（例如，保持数学正确性）的执行瓶颈。其在该流水线中的主要贡献包括：
Specifically, RSI is deployed to handle execution bottlenecks where the optimization space is vast but the outcomes are strictly verifiable (e.g., maintaining mathematical correctness). Its primary contributions in this pipeline include:
通信集合与量化：联合搜索数据压缩和网络传输的配置空间，以消除分布式推理瓶颈。
Communication collectives and quantization: Jointly searching the configuration space of data compression and network transfer to eliminate distributed inference bottlenecks.
融合核实现：自动发现并生成最优的算子分组，以最小化冗余的 HBM 流量。
Fusion kernel implementation: Automatically discovering and generating optimal operator groupings to minimize redundant HBM traffic.
稀疏注意力机制实现：设计高效且硬件感知的注意力机制，以处理超长序列长度。
Sparse attention implementation: Engineering highly efficient, hardware-aware attention mechanisms to manage massive sequence lengths.
精度保持的 LoRA 集成：确保低秩适应操作在加速的同时，不损害数值保真度或生成质量。
Precision-preserving LoRA integration: Ensuring that low-rank adaptation operations are accelerated without compromising numerical fidelity or generation quality.
相比之下，流水线中开放式的、固有有损的以及架构性的组件，需要一种范式转变的直觉，而 RSI 目前无法提供这种直觉。这些由人类设计的组件包括：
In contrast, the open-ended, inherently lossy, and architectural components of the pipeline require a level of paradigm-shifting intuition that RSI currently cannot provide. These human-designed components include:
跨分辨率两阶段框架：构思总体算法策略，以利用扩散模型的逐步特性，平衡早期全局布局生成与后期高分辨率细化。
The cross-resolution two-stage framework: Conceptualizing the overarching algorithmic strategy to exploit the step-wise nature of diffusion models, balancing early global layout generation with later high-resolution refinement.
潜在适配器：认识到标准 VAE 解码器/编码器的系统级权衡，并设计自定义适配器，以严格场景身份不变性为代价，高效桥接从粗到细的生成阶段。
The latent adapter: Recognizing the system-level trade-offs of standard VAE decoders/encoders, and designing a custom adapter to efficiently bridge the coarse-to-fine generation stages at the cost of strict scene-identity invariance.
基于这一二分法，我们观察到 RSI 特别适合系统地解决确定性的、边界明确的工程问题，这类问题的成功可以对照基线进行数学验证。然而，对于需要引入有损近似或整体性权衡的开放式架构流水线设计，人类工程仍然不可或缺。最终，人类算法设计与 RSI 驱动的内核优化之间的这种协同，实现了快于实时的高分辨率视频合成，大幅降低了在生产环境中部署大规模视频生成模型的门槛。
Based on this dichotomy, we observe that RSI is exceptionally well-suited for systematically resolving deterministic, well-bounded engineering problems where success can be mathematically verified against a baseline. However, for open-ended, architectural pipeline designs that necessitate lossy approximations or holistic trade-offs, human engineering remains indispensable. Ultimately, this synergy of human algorithmic design and RSI-driven kernel optimization enables faster-than-real-time high-resolution video synthesis, drastically lowering the barrier to deploying massive video generation models in production.
源语料库包含 80,000 个视频，存储在八个归档中。每个源视频在编码前以相同方式调整大小，然后使用冻结的 H3 和 LTX-2.5 Conv VAE 进行编码。我们使用确定性后验模式和发布的逐通道归一化统计量。对于 \(1344 \times 768\) 的 192 帧片段，H3 张量在通道–时间–高度–宽度顺序下的形状为 \(24 \times 57 \times 48 \times 84\)。LTX 将输入填充至 193 帧，并产生 \(128 \times 25 \times 24 \times 42\)。H3 标记位置遵循 17 帧分块规则，锚点为 \(0,4,8,12,16,17,21,\ldots\)，并配置移除三个尾部标记。第 3.1 节中的固定对齐产生 \(384 \times 25 \times 24 \times 42\) 特征，在空间像素反混洗之前，将每个 H3 标记保留在三个时间槽之一中。跳跃投影将这些特征映射到 128 个通道，并在岭回归初始化后保持固定；残差网络进行端到端训练。
The source corpus contains 80,000 videos stored in eight archives. Each source video is resized identically before encoding with the frozen H3 and LTX-2.5 Conv VAEs. We use deterministic posterior modes and the released per-channel normalization statistics. For a 192-frame clip at \(1344 \times 768\), the H3 tensor has shape \(24 \times 57 \times 48 \times 84\) in channel–time–height–width order. LTX pads the input to 193 frames and produces \(128 \times 25 \times 24 \times 42\). H3 token positions follow the 17-frame chunking rule, with anchors \(0,4,8,12,16,17,21,\ldots\) and the configured removal of three trailing tokens. The fixed alignment in Sec. 3.1 yields \(384 \times 25 \times 24 \times 42\) features, retaining each H3 token in one of three temporal slots before spatial pixel-unshuffle. The skip projection maps these features to 128 channels and remains fixed after ridge-regression initialization; the residual network is trained end to end.
最终的解码器感知优化从先前训练好的适配器开始，使用 65,536 个原生分辨率的缓存视频对。两个阶段分别进行 2,500 和 5,000 次更新，全局批量大小为 32，额外呈现 240,000 个样本。每个阶段使用 AdamW，峰值学习率为 \(8 \times 10^{-5}\)，\(\beta=(0.9,0.95)\)，零权重衰减，梯度裁剪为 1.0。相应的预热长度分别为 100 和 150 次更新，随后余弦衰减至峰值学习率的 5%；第二阶段从第一阶段的权重重新启动优化器。训练使用 BF16 自动混合精度，并通过残差堆栈和冻结解码器进行激活检查点。公式 1 中的目标函数在完整空间网格上评估，解码视频监督权重为 \(\alpha=160\)。仅更新适配器的残差路径；仿射跳跃连接和两个 VAE 保持固定。报告的最终检查点直接使用训练后的权重，未使用 EMA 平均。
The final decoder-aware optimization starts from an earlier trained adapter and uses 65,536 cached video pairs at native resolution. Two phases of 2,500 and 5,000 updates use global batch size 32, for 240,000 additional sample presentations. Each phase uses AdamW with peak learning rate \(8 \times 10^{-5}\), \(\beta=(0.9,0.95)\), zero weight decay, and gradient clipping at 1.0. The respective warmup lengths are 100 and 150 updates, followed by cosine decay to 5% of the peak learning rate; the second phase restarts the optimizer from the first phase’s weights. Training uses BF16 autocast and activation checkpointing through the residual stack and the frozen decoder. The objective in Eq. 1 is evaluated on the full spatial grid, with decoded-video supervision weight \(\alpha=160\). Only the adapter’s residual path is updated; the affine skip and both VAEs remain fixed. The reported final checkpoint uses the trained weights directly, without EMA averaging.
样本标识符通过稳定哈希以 98/1/1 的比例分配给训练、验证和测试分区。最终的开发集和测试集各包含 256 个视频，并排除了早期评估队列中使用的标识符。检查点选择使用开发数据。在最终评估器中，预测和教师重建被限制在 \([0,1]\) 范围内，并在它们共同的解码时间范围上进行比较，包括 VAE 填充，而不是显式裁剪回 192 帧。PSNR 按视频计算，然后在视频间取平均；SSIM 使用 16 个均匀采样的解码帧。运动四分位数由教师潜在变量的时间差异定义。配对置信区间使用对相同片段标识符的 10,000 次自助重采样。这些重建测量不评估提示遵循度、音频质量或后续 LTX 细化的输出。
Sample identifiers are assigned to train, validation, and test partitions by a stable hash with a 98/1/1 split. The final development and test cohorts contain 256 videos each and exclude identifiers used in earlier evaluation cohorts. Checkpoint selection uses development data. In the final evaluator, predictions and teacher reconstructions are clamped to \([0,1]\) and compared over their common decoded temporal extent, including the VAE padding rather than an explicit crop back to 192 frames. PSNR is computed per video and then averaged across videos; SSIM uses 16 uniformly sampled decoded frames. Motion quartiles are defined by temporal differences in the teacher latents. Paired confidence intervals use 10,000 bootstrap resamples over the same clip identifiers. These reconstruction measurements do not evaluate prompt adherence, audio quality, or the output of subsequent LTX refinement.
表 4 中的三个分支均使用相同的缓存 H3 潜变量，并以相同的像素分辨率输出归一化的 LTX 潜变量。完整的往返过程使用已发布的 H3 解码器和分块 LTX 卷积编码器。Tiny AutoEncoder 分支使用 TAEH3 解码和 TAELTX 编码，并采用其并行执行模式。两个往返分支均使用一次预热和三次测量迭代；适配器使用三次预热和 20 次测量迭代。CUDA 同步用于界定墙钟时间测量，模型加载和磁盘 I/O 被排除在外。所分析的适配器是一个较早的检查点，其几何结构、宽度、深度和参数数量与最终质量检查点相同。峰值分配因每个实现而异，且不包括共驻的生成流水线；它并非对 Spark 总内存使用量的估计。
All three arms in Tab. 4 consume the same cached H3 latent and output normalized LTX latents at the same pixel resolution. The full round trip uses the released H3 decoder and tiled LTX Conv encoder. The Tiny AutoEncoder arm uses TAEH3 decoding and TAELTX encoding with their parallel execution mode. Both round-trip arms use one warmup and three measured iterations; the adapter uses three warmups and 20 measured iterations. CUDA synchronization brackets wall-time measurements, and model loading and disk I/O are excluded. The profiled adapter is an earlier checkpoint with the same geometry, width, depth, and parameter count as the final quality checkpoint. Peak allocation is specific to each implementation and excludes the co-resident generative pipeline; it is not an estimate of total Spark memory usage.
合并诊断固定检查点、提示、初始噪声和采样计划，并在禁用其他运行时优化的情况下，比较在四个 GB200 GPU 上的首次去噪评估。在将线性计算保持为 FP32 的情况下，合并到 BF16 权重在视频预测中相对于原始分离分支基线产生 35.26% 的相对 RMSE；将合并后的权重保留为 FP32 可将此值降至 6.24%。因此，仅使用 FP32 GEMM 无法恢复在权重存储期间已丢失的更新。将分离分支本身改为 FP32 相对于同一基线给出 6.08% 的相对 RMSE，这表明为什么更高精度也不足以精确复现原始混合精度计算。这些是在一次去噪评估中对一个输入的预测差异，而非感知质量分数。第 3.5 节中的六投影计数使用 \(W+BA\) 的 FP32 重建，并统计 \(BA\) 中非零元素的数量，这些元素在将和转换为 BF16 后返回原始 \(W\) 条目。
The merge diagnostic fixes the checkpoint, prompt, initial noise and sampling schedule, and compares the first denoising evaluation on four GB200 GPUs with the other runtime optimizations disabled. With linear computation held in FP32, merging into BF16 weights gives 35.26% relative RMSE in the video prediction against the original separate-branch baseline; retaining the merged weights in FP32 reduces this to 6.24%. FP32 GEMMs alone therefore do not recover updates already lost during weight storage. Changing the separate branches themselves to FP32 gives 6.08% relative RMSE against the same baseline, showing why higher precision is also insufficient to reproduce the original mixed-precision computation exactly. These are prediction discrepancies for one input at one denoising evaluation, not perceptual-quality scores. The six-projection counts in Sec. 3.5 use FP32 reconstructions of \(W+BA\) and count nonzero entries of \(BA\) for which casting the sum to BF16 returns the original \(W\) entry.
发布的实现[^1]支持公开的 PEFT LoRA 和 FastVideo v2 混合格式。混合 `.diff`/`.diff_b` 修正以 FP32 累加方式应用于基础参数，然后再转换为参数数据类型；低秩因子保持独立。消费者融合需要推理模式、一个激活的 scale-1 适配器、普通的 BF16 线性层、恒等 LoRA dropout，且无投影钩子。不支持的安装配置在安装消费者补丁之前被拒绝；不支持的运行时状态回退到原生 PEFT。QKV 消费者融合使用多 GPU Ulysses 路径，而单 GPU 密集 Q/K/V 投影以及主 Transformer 块之外的投影保留原生执行。 [^1]: NVlabs/Sana PR #503, commit 51b14fd6f262。
The released implementation[^1] accepts the supported public PEFT LoRA and FastVideo v2 hybrid formats. Hybrid `.diff`/`.diff_b` corrections are applied to the base parameters with FP32 accumulation before conversion to the parameter dtype; low-rank factors remain separate. Consumer fusion requires inference mode, one active scale-1 adapter, ordinary BF16 linear layers, identity LoRA dropout, and no projection hooks. Unsupported installation configurations are rejected before consumer patches are installed; unsupported runtime states fall back to native PEFT. QKV consumer fusion uses the multi-GPU Ulysses path, while single-GPU dense Q/K/V projections and projections outside the main transformer blocks retain native execution. [^1]: NVlabs/Sana PR #503, commit 51b14fd6f262.
公开运行时默认采用合并模式。分离模式和融合模式保留 LoRA 分支，并要求 compute-quant=none；后者启用消费者内核。这会禁用量化线性计算，但不会禁用独立的注意力机制或通信优化。表 5 中的受控比较明确选择了每种模式；启用消费者融合并非在每个流水线测量中都隐含。发布的回归测试套件覆盖 BF16 舍入边界、非连续输入、索引门、提升数据类型回退、BF16/INT8 QKV 打包，以及超过 \(2^{31}\) 个元素的行寻址。全流水线字节一致性仍限于所测试的输入及其记录的 GPU 和软件配置。
The public runtime defaults to merged. The separate and fused modes retain LoRA branches and require compute-quant=none; the latter enables the consumer kernels. This disables quantized linear compute, not the independent attention or communication optimizations. The controlled comparison in Tab. 5 explicitly selects each mode; enabling consumer fusion is not implicit in every pipeline measurement. The released regression suite covers BF16 rounding boundaries, non-contiguous inputs, indexed gates, promoted-dtype fallback, BF16/INT8 QKV packing, and row addressing beyond \(2^{31}\) elements. Full-pipeline byte agreement remains scoped to the tested inputs and their recorded GPU and software configurations.
图 8 中的额外重跑使用种子 42、八次去噪评估、四块 GB200 GPU，以及 362 帧，分辨率为 \(1344\times 768\)，帧率为 24 FPS。每个分支执行其自然的随机抽取，并在重放前根据相同的存储参考验证生成的张量和生成器状态；初始条件和视频/音频采样网格也会被检查。原生分支和融合分支在每条记录的预测以及完整的解码视频和音频中均匹配。我们在 MP4 编码前将所有帧直接导出为无损 8 位 RGB PNG，并比较其像素哈希。全帧行使用常见的从零开始的索引 180 和 300。而 \(480\times 320\) 的细节裁剪则遵循明确标记时间处的蝴蝶爆发，因为合并会改变其时间以及外观。这些审计运行包括张量哈希和帧导出，与表 5 中的热延迟测量分开。
The additional rerun in Fig. 8 uses seed 42, eight denoising evaluations, four GB200 GPUs, and 362 frames at \(1344\times 768\) and 24 FPS. Each arm performs its natural random draws and verifies the resulting tensors and generator states against the same stored reference before replay; initial conditioning and the video/audio sampling grids are also checked. The native and fused arms match in every recorded prediction and in the complete decoded video and audio. We export all frames directly to lossless 8-bit RGB PNGs before MP4 encoding and compare their pixel hashes. The full-frame rows use common zero-based indices 180 and 300. The \(480\times 320\) detail crops instead follow the butterfly burst at the explicitly labeled times, because merging changes its timing as well as its appearance. These audited runs include tensor hashing and frame export and are separate from the warm latency measurements in Tab. 5.