We present Seedream 3.0, a high-performance Chinese-English bilingual image generation foundation model. We develop several technical improvements to address existing challenges in Seedream 2.0, including alignment with complicated prompts, fine-grained typography generation, suboptimal visual aesthetics and fidelity, and limited image resolutions. Specifically, the advancements of Seedream 3.0 stem from improvements across the entire pipeline, from data construction to model deployment. At the data stratum, we double the dataset using a defect-aware training paradigm and a dual-axis collaborative data-sampling framework. Furthermore, we adopt several effective techniques such as mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling in the pre-training phase. During the post-training stage, we utilize diversified aesthetic captions in SFT, and a VLM-based reward model with scaling, thereby achieving outputs that well align with human preferences. Furthermore, Seedream 3.0 pioneers a novel acceleration paradigm. By employing consistent noise expectation and importance-aware timestep sampling, we achieve a 4 to 8 times speedup while maintaining image quality. Seedream 3.0 demonstrates significant improvements over Seedream 2.0: it enhances overall capabilities, in particular for text-rendering in complicated Chinese characters which is important to professional typography generation. In addition, it provides native high-resolution output (up to 2K), allowing it to generate images with high visual quality.
核心贡献 · Key contributions
提出缺陷感知训练范式和双轴数据采样，将数据集规模翻倍并提升质量。 Proposes a defect-aware training paradigm and dual-axis data sampling to double dataset size and improve quality.
引入混合分辨率训练、跨模态 RoPE、表征对齐损失和分辨率感知时间步采样，提升可扩展性与对齐能力。 Introduces mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling for better scalability and alignment.
在 SFT 中使用多样化美学描述，并采用基于 VLM 的奖励模型与 Scaling，使输出对齐人类偏好。 Uses diversified aesthetic captions in SFT and a VLM-based reward model with scaling to align outputs with human preferences.
首创一致性噪声期望与重要性感知时间步采样的加速范式，实现 4-8 倍加速。 Pioneers an acceleration paradigm with consistent noise expectation and importance-aware timestep sampling, achieving 4-8x speedup.
实现原生高达 2K 分辨率输出，消除后处理超分辨率流程。 Achieves native high-resolution output up to 2K, eliminating post-processing super-resolution pipelines.
在文本渲染（尤其是密集中文字符）和逼真肖像生成方面达到最先进水平。 Demonstrates state-of-the-art performance in text rendering, especially for dense Chinese characters, and photorealistic portrait generation.
局限 · Limitations
评估聚焦单轮生成；多轮编辑和交错生成能力仍有限。 Evaluation focuses on single-round generation; multi-round editing and interleaved generation remain limited.
复杂多图像参考任务的性能尚未解决。 Performance on complex multi-image reference tasks is not yet addressed.
与 GPT-4o 的比较限于精选案例；因 API 不可用，系统性评估尚待进行。 Comparison with GPT-4o is limited to selected cases; systematic evaluation is pending due to API unavailability.
模型处理超出基准的极长或复杂提示的能力尚未充分探索。 The model's capabilities in handling extremely long or complex prompts beyond the benchmark are not fully explored.
训练数据中的潜在偏差及其对生成内容的影响未讨论。 Potential biases in training data and their impact on generated content are not discussed.