Source-linked AI summary
Seedream 3.0 Technical Report
Yu Gao, Lixue Gong, Qiushan Guo, Xiaoxia Hou, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, Wei Liu, Yichun Shi, Shiqi Sun, Yu Tian, Zhi Tian, Peng Wang, Rui Wang, Xuanda Wang, Xun Wang, Ye Wang, Guofeng Wu, Jie Wu, Xin Xia, Xuefeng Xiao, Zhonghua Zhai, Xinyu Zhang, Qi Zhang, Yuwei Zhang, Shijia Zhao, Jianchao Yang, Weilin Huang
TL;DR
Seedream 3.0 addresses Seedream 2.0’s limitations in complex prompt alignment, fine-grained typography, visual quality, and native resolution. It upgrades data construction, training, post-training, and acceleration, and reports broader capabilities, stronger text rendering, native high-resolution output, and faster inference. A systematic evaluation limitation remains for GPT-4o because its lack of a large-scale image-generation API prevented a systematic comparison.
Problem
Seedream 2.0 had limitations in complex prompt adherence, fine-grained typography, nuanced visual aesthetics and fidelity, and native image resolution.
Method
Seedream 3.0 upgrades the pipeline through expanded defect-aware data, dual-axis sampling, mixed-resolution and cross-modality training, representation alignment, aesthetic-caption SFT, VLM reward modeling, and accelerated sampling.
Results
Seedream 3.0 shows comprehensive gains over Seedream 2.0, especially in text rendering, and provides native 2K output while generating 1K images in 3.0 seconds without PE.
Takeaways & Limitations
The model is presented as a high-quality, fast bilingual image-generation foundation model with stronger typography and native high-resolution generation.
Takeaways & Limitations
A systematic GPT-4o evaluation was not conducted because no API was available for large-scale image generation.
Abstract
from arXiv · showhide
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.
1 Introduction
Seedream 3.0 addresses Seedream 2.0’s limitations in prompt alignment, typography, visual quality, and native resolution through system-level improvements. It reports broader capabilities, stronger text rendering, high-resolution output, and faster inference.
- Challenges: Seedream 2.0 remained limited in numerical precision, multi-object spatial relationships, small text, multiline compositions, aesthetic fidelity, and native resolution.Its native output was limited to small resolutions such as 512 × 512px, requiring post-processing super-resolution.
- Technical improvements: The methodology approximately doubled the dataset and combines dynamic dual-axis sampling with mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling.Post-training adds diverse aesthetic captions and a VLM-based reward model to improve overall performance and visual-language alignment.
- Reported improvements: Seedream 3.0 delivers comprehensive gains in text-image alignment, compositional structure, aesthetic quality, and text rendering compared with Seedream 2.0.The reported improvements include stronger small-size Chinese and English text and high-aesthetic long-text layouts.
- Reported improvements: Seedream 3.0 provides native 2K output, supports higher resolutions and diverse aspect ratios, and removes the need for post-processing.The model is also reported to produce high-quality 1K images in 3.0 seconds without PE.
- Deployment: Seedream 3.0 was integrated into Doubao and Jimeng in early April 2025 as a practical productivity tool.The passage expresses the authors’ intended application scope across work and daily life.
2 Technical Details
Seedream 3.0 improves data construction, architecture, training, post-training, and acceleration across its image-generation pipeline. The reported design targets scalable multimodal alignment, controllable captions, and high-quality few-step synthesis.
- Data construction: Defect-aware training retains previously excluded samples while masking defect-region gradients, expanding the effective training dataset by 21.7% while maintaining stability.Seedream 2.0 had excluded samples with artifacts such as watermarks, overlaid text, subtitles, and mosaics.
- Data construction: Dual-axis collaborative sampling balances visual morphology and textual semantics, while cross-modal retrieval adds expert concepts and distribution calibration.The framework uses hierarchical clustering for visual patterns and TF-IDF for semantic balance.
- Model architecture: The architecture inherits MMDiT and increases base-model parameters to improve scalability, generalizability, and visual-language alignment.Cross-modality RoPE further models intra- and cross-modality relationships by applying 2D positional treatment to text tokens.
- Model training: Mixed-resolution training packs varied aspect ratios and resolutions, pretraining near 256^2 and finetuning from 512^2 to 2048^2 with size conditioning.Resolution balancing ensures adequate sampling across resolutions and supports prompt following in varied scenarios.
- Model training: Representation alignment uses cosine distance to DINOv2-L features with λ = 0.5, and the authors report faster convergence for large-scale text-to-image generation.The training objective combines flow matching with representation alignment loss.
- Post-training: Post-training uses Continuing Training, SFT, RLHF, and Prompt Engineering, with specialized aesthetic, style, and layout captions improving controllability.The model omits a Refiner stage because it directly generates images from 512^2 to 2048^2.
- Post-training: VLM reward modeling derives rewards from normalized “Yes”-token probabilities and benefits from scaling, with increased capacity correlating with improved reward-modeling performance.This replaces Seedream 2.0’s CLIP reward model.
- Model acceleration: Few-step sampling matches or surpasses 50-NFE baselines across aesthetic quality, text-image alignment, and structural fidelity.The acceleration framework uses adaptive per-sample trajectories and noise-consistency mechanisms to reduce computational cost.
3 Model Performance
Seedream 3.0 ranks first in public and automatic evaluations, with broad gains over Seedream 2.0 and competing models in alignment, structure, aesthetics, and text rendering. Human evaluation further shows strong performance across usage scenarios, while comparisons highlight superior dense-text rendering and photorealistic portrait generation.
- Artificial Analysis Arena: Seedream 3.0 ranks first among top-tier text-to-image models in a publicly conducted global evaluation.The comparison includes GPT-4o, Imagen 3, Midjourney v6.1, FLUX1.1 Pro, and Ideogram 3.0.
- Artificial Analysis Arena: Seedream 3.0 secures the top Artificial Analysis position, outperforming GPT-4o and leading models including Midjourney v6.1 and Imagen 3.It also shows the best performance across most listed style and subject sub-dimensions.
- Human Evaluation: Human experts find Seedream 3.0 significantly better than Seedream 2.0 and competing models in text-image alignment and structural fidelity.It scores higher than Midjourney overall in aesthetic performance and is notably superior in the design category, while lagging slightly in art.
- Automatic Evaluation: Seedream 3.0 consistently ranks first across automatic text-image alignment and image-quality benchmarks.It ranks first in EvalMuse, HPSv2, MPS, Internal-Align, and Internal-Aes; HPSv2 exceeds 0.3 for the first time.
3.3 Text Rendering
Seedream 3.0 substantially improves bilingual text rendering, especially for dense, small Chinese and English text, using a benchmark that combines perceptual and objective metrics.
- The benchmark contains 180 Chinese and 180 English prompts spanning logos, posters, electronic displays, printed text, and handwriting.
- Text rendering is evaluated with availability rate, text accuracy rate, and hit rate.Availability incorporates text correctness, integration with other content, and overall aesthetic quality.
- Seedream 3.0 achieves a 94% text availability rate for both Chinese and English characters.Chinese text availability improves by 16% over Seedream 2.0.
- Nearly equivalent availability and hit rates indicate minimal layout or medium-related rendering errors.
- Seedream 3.0 handles dense text with high precision for small characters and aesthetically demanding long-text compositions.The paper illustrates these capabilities in Figures 12 and 13.
3.4 Photorealistic Portrait
Seedream 3.0 targets the synthetic appearance of generated portraits and evaluates whether its outputs improve realism, texture, and visual quality at higher resolution.
- AI-generated portraits often appear artificial because of overly smooth skin and oily texture.
- The portrait evaluation set contains 100 prompts covering expressions, postures, angles, hair, skin texture, clothing, and accessories.Participants select preferred portraits in an Elo-style comparison and justify their choices.
- Seedream 3.0 directly generates images at 2048×2048 resolution.
- Higher resolution enhances portrait texture and advances generated portrait quality toward professional photography standards.Figure 16 provides examples of high-resolution portraits.
3.5 Comparison with GPT-4o
Against GPT-4o, Seedream 3.0 shows complementary strengths: GPT-4o is strong on small English text and some LaTeX symbols, while Seedream 3.0 excels at dense Chinese typography and image quality.
- A systematic GPT-4o evaluation was not conducted because no API for large-scale image generation was available.The comparison therefore relies on selected cases.
- GPT-4o renders small English characters and certain LaTeX symbols accurately but struggles with Chinese fonts.
- Seedream 3.0 outperforms GPT-4o in dense Chinese text generation, typesetting, and aesthetic composition.
- Image Editing: In image editing, GPT-4o handles many requirements but struggles to preserve original-image IP and ID consistency, while Gemini-2.0 can preserve pixels but may reduce color naturalness and quality.SeedEdit 1.6 offers a more balanced profile for typical editing needs.
- Text Editing: SeedEdit inherits Seedream 3.0’s text capabilities and produces satisfying results on text writing, removal, and modification tasks.
- Generation Quality: GPT-4o-generated images tend to show a dark yellowish hue and significant noise that affect usability.Generation quality is assessed through color, texture, clarity, and aesthetic appeal.
4 Conclusion
Seedream 3.0 introduces system-level upgrades across data, training, post-training, and acceleration to address resolution, prompt adherence, typography, aesthetics, and fidelity challenges.
- Seedream 3.0 achieves comprehensive improvements, native high-resolution output, stronger text rendering, enhanced visual appeal, and extreme generation speed.The paper describes potential integration into platforms including Doubao and Jimeng.
A Contributions and Acknowledgments
All Seedream contributors are listed alphabetically by last name.
- Contributors are listed in alphabetical order by their last names.
A.1 Core Contributors
The core contributors are listed by name.
- The core contributor list includes Yu Gao, Lixue Gong, Qiushan Guo, and other contributors.
A.2 Contributors
Additional contributors are listed by name.
- The contributor list includes Haoshen Chen, Kaixi Chen, Xiaojing Dong, and other contributors.