PaperlayerView exact revision on arXiv

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

arXiv:2504.11346v3cs.CV

TL;DR

Seedream 2.0 remained limited in fine-grained typography and visual aesthetics. Seedream 3.0 upgrades the generation pipeline and reports broader capability gains, especially in text rendering, alongside native 2K output.

  • Problem

    Seedream 2.0 remained limited in fine-grained typography, including small text and multi-line compositions, and in capturing nuanced visual aesthetics.

    Source: §1 · p. 4

  • Method

    Seedream 3.0 applies system-level upgrades to data construction, pretraining, post-training, and model acceleration.

    Source: §4 · p. 19

  • Results

    Seedream 3.0 reports gains in text-image alignment, compositional structure, aesthetics, and text rendering, particularly for small Chinese and English text and high-aesthetic long-text layouts.

    Source: §1 · p. 4

  • Takeaways & Limitations

    Native 2K output avoids post-processing and supports diverse aspect ratios.

    Source: §1 · p. 4

    Without an API for large-scale GPT-4o image generation, systematic evaluation was not conducted, leaving only selected-case comparisons.

    Source: §3.5 · p. 16

Abstract

from arXiv · show

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 prompt-following, typography, aesthetic-fidelity, and resolution limitations through pipeline-wide data, training, post-training, and acceleration improvements. The report highlights stronger generation capabilities, native 2K output, and faster inference.

Motivation

  • Seedream 2.0’s identified limitations include complicated-prompt alignment, fine-grained typography, visual aesthetics and fidelity, and native resolution restricted to small outputs such as 512 × 512px.Source: §1 · p. 4

Method

  • The pipeline doubles dataset size through dynamic sampling, adds mixed-resolution and alignment-focused pre-training, uses aesthetic-caption SFT and VLM rewards, and accelerates sampling with consistent noise expectation.The dynamic sampler uses image-cluster distribution and textual semantic coherence; pre-training also uses cross-modality RoPE and resolution-aware timestep sampling.Source: §1 · p. 4

Results

  • Seedream 3.0 shows advances over Seedream 2.0 in user preference and text-image alignment, compositional structure, aesthetic quality, and text rendering.The report identifies these as comprehensive capability improvements.Source: §1 · p. 4
  • Seedream 3.0 improves small-text generation in Chinese and English and high-aesthetic long-text layouts, outperforming human-designed Canva templates in graphic design output.Source: §1 · p. 4
  • Seedream 3.0 substantially improves aesthetic quality, especially in cinematic scenarios and portrait realism.Source: §1 · p. 4
  • Seedream 3.0 supports native 2K output without post-processing and adapts to higher resolutions and diverse aspect ratios.Source: §1 · p. 4
  • Seedream 3.0 generates a 1K image in 3.0 seconds without PE, reducing inference cost and running faster than other commercial models.Source: §1 · p. 4

2 Technical Details · 2.1 Data · 2.2 Model Pre-training

Seedream 3.0 expands and better balances training data, then combines architecture and training changes to improve resolution generalization, visual-text alignment, and training convergence. Its pre-training adapts both token-position handling and timestep sampling to the image resolution.

2.1 Data

  • 21.7% more training data is enabled by retaining images with less than 20% detected defect area and masking defect-region gradients during diffusion loss.The defect detector was trained on 15,000 manually annotated samples selected by active learning; the area threshold is configurable.Source: §2.1 · p. 5
  • Dual-axis sampling balances visual morphology and text semantics, while cross-modal retrieval adds targeted concepts, similarity-weighted sampling, and neighboring image-text pairs.Visual patterns are balanced with hierarchical clustering; semantic balance uses term frequency and inverse document frequency (TF-IDF). The retrieval system creates a joint image-text embedding space and reports state-of-the-art performance across benchmark tests.Source: §2.1 · p. 5

2.2.1 Model Architectures

  • Seedream 3.0 retains Seedream 2.0’s MMDiT image-text architecture while increasing base-model parameters to improve scalability, generalizability, and visual-language alignment.Seedream 2.0 [4] uses MMDiT [3] to process image and text tokens and capture relationships between the modalities.Source: §2.2.1 · p. 5
  • Mixed-resolution training packs varied aspect ratios and resolutions, progressing from average resolution 256^2 to fine-tuning at 512^2–2048^2 with size embeddings.This increases data diversity and improves generalizability to unseen resolutions.Source: §2.2.1 · p. 5
  • Cross-modality RoPE treats text tokens as 2D tokens and extends their position IDs after image tokens to model intra- and cross-modal relationships.The method extends Scaling RoPE and is intended to improve visual-text alignment and text-rendering accuracy.Source: §2.2.1 · p. 5

2.2.2 Model Training Details

  • Training combines flow matching with REPA, a cosine-distance alignment between intermediate MMDiT features and pre-trained DINOv2-L features.The representation alignment loss weight is λ = 0.5, and the authors report that it accelerates convergence of large-scale text-to-image generation.Source: §2.2.2 · p. 6
  • Resolution-aware timestep sampling shifts a logit-normal distribution toward lower SNRs at higher training resolutions, using dataset-average resolution for training and target resolution and aspect ratio for inference.The timestep distribution is adaptive to the dataset and the intended inference resolution.Source: §2.2.2 · p. 6

2.3 Model Post-training · 2.4 Model Acceleration

Post-training combines staged refinement, specialized captioning, resolution-balanced training, and scaled VLM rewards; acceleration redesigns generation trajectories and timestep selection to support high-quality few-step synthesis.

2.3 Model Post-training

  • Post-training uses Continuing Training, Supervised Fine-Tuning, RLHF, and prompt engineering, omitting a refiner because the model generates directly from 512^2 to 2048^2 resolution.The stages are similar to Seedream 2.0’s post-training process.Source: §2.3 · p. 7

2.3.1 Aesthetic Caption

  • Specialized caption models describe aesthetics, style, and layout, helping the model respond to relevant prompts with greater controllability and improved prompt-engineering performance.Multiple caption-model versions were trained for data used in CT and SFT.Source: §2.3.1 · p. 7

2.3.2 Model Training Details

  • Resolution balancing ensures adequate training-data sampling across resolutions, enhancing prompt following in varied scenarios.Source: §2.3.2 · p. 7

2.3.3 Reward Model Scaling

  • VLM reward models replace CLIP with normalized “Yes”-token probabilities for instruction queries; scaling from 1B to >20B parameters correlates with improved reward-modeling performance.The approach draws on pretrained LLM knowledge and its scaling potential.Source: §2.3.3 · p. 7

2.4 Model Acceleration

  • Building on Hyper-SD [17] and RayFlow [20], the framework guides each sample toward an instance-specific target distribution, reducing path collisions and improving stability and diversity.This replaces the shared path toward a standard Gaussian prior used in conventional diffusion models.Source: §2.4 · pp. 7–8
  • A unified noise expectation vector estimated from a pretrained model aligns denoising across timesteps, enabling fewer sampling steps without degrading image quality.Theoretical analysis indicates this design maximizes the probability of the forward-backward path and supports stable sampling and reliable reconstructions.Source: §2.4 · p. 8
  • Stochastic Stein Discrepancy [6] and a neural network learn data-dependent timestep importance, prioritizing loss-reducing steps for faster convergence and more efficient training.This replaces uniform timestep sampling, which the passage says increases loss variance and wastes computation on uninformative steps.Source: §2.4 · p. 8
  • The framework matches or surpasses baselines requiring 50 function evaluations (NFE) in aesthetic quality, text-image alignment, and structural fidelity.It supports efficient few-step sampling without compromising generation quality.Source: §2.4 · p. 8

3 Model Performance

Seedream 3.0 ranks first in a public evaluation of leading text-to-image models and shows improvements over Seedream 2.0 across key performance indicators. Its standout capabilities include dense text rendering and photorealistic human portrait generation.

3 Model Performance

  • Seedream 3.0 ranks first among top-tier text-to-image models in a public evaluation, including GPT-4o, Imagen 3, Midjourney v6.1, FLUX1.1 Pro, and Ideogram 3.0.The overall results are presented in Figure 1.Source: §3 · p. 8
  • Expert evaluations, conducted manually and through automated means, show Seedream 3.0 improves across all key performance indicators over Seedream 2.0 and outperforms industry-leading counterparts.Source: §3 · p. 8
  • Seedream 3.0 exhibits exceptional capabilities in dense text rendering and photorealistic human portrait generation.The paper discusses these capabilities in Sections 3.3 and 3.4, respectively.Source: §3 · p. 8

3.1 Artificial Analysis Arena

Seedream 3.0 ranked first overall in the Artificial Analysis Arena, outperforming GPT-4o and leading other listed models. It performed best across most reported style and subject sub-dimensions.

3.1 Artificial Analysis Arena

  • Seedream 3.0 secured the top overall position, outperforming GPT-4o and leading Recraft V3, HiDream, Reve Image, Imagen 3 (v002), FLUX1.1 Pro, and Midjourney v6.1.Source: §3.1 · p. 8, Figure 5 · p. 9
  • Seedream 3.0 performed best across most reported style categories and subject categories, including portraits, groups and activities, fantasy, futuristic scenes, and physical spaces.Source: §3.1 · p. 9

3.2 Comprehensive Evaluation

Seedream 3.0 is evaluated on human and automatic benchmarks covering text-image alignment, structural fidelity, and image quality. It ranks first across automatic benchmarks and performs strongly in human evaluations across varied usage scenarios.

3.2.1 Human Evaluation

  • Bench-377 contains 377 prompts spanning cinematic, arts, entertainment, aesthetic design, and practical design scenarios, evaluated for text-image alignment, structural correctness, and aesthetics.Practical design includes routine tasks such as arranging icons in slides and designing illustrations for handwriting newspapers.Source: §3.2.1 · p. 9
  • In human evaluation, Seedream 3.0 outperforms Seedream 2.0 and competing models in text-image alignment and structural fidelity, and surpasses Midjourney overall in aesthetics.It is notably stronger than Midjourney in design, though slightly behind in art; Imagen 3 is competent in alignment and structure but underperforms in aesthetics.Source: §3.2.1 · p. 9, Figure 6 · p. 10

3.2.2 Automatic Evaluation

  • Seedream 3.0 consistently ranks first across all benchmarks and image-quality metrics—HPSv2, MPS, Internal-Align, and Internal-Aes—and outperforms Midjourney in aesthetic evaluation.Image-quality evaluation uses two external metrics, HPSv2 and MPS, and two internal models, Internal-Align and Internal-Aes. Seedream 3.0’s HPSv2 index exceeds 0.3 for the first time.Source: §3.2.2 · p. 10
  • On EvalMuse [7], Seedream 3.0 outperforms other models and improves over Seedream 2.0 in most fine-grained dimensions, especially objects, activities, locations, food, and space.Ideogram 2.0 is included for consistency with previously reported results.Source: Table 1 · p. 10, §3.2.2 · p. 10

3.2.1 Human Evaluation

  • Examples show that improved alignment, structural stability, and aesthetics support precise user-intent representation, complex descriptions, specialized terms, and more available outputs.The reported capabilities include conveying movie atmosphere through micro-expressions, interpreting “three-view” requests, preserving limbs in motion, and improving small-object detail, color, lighting, texture, and composition.Source: §3.2.1 · p. 10

3.3 Text Rendering

Seedream 3.0 improves text rendering across Chinese and English, including dense text, evaluated with availability and objective text metrics on a bilingual prompt benchmark.

3.3 Text Rendering

  • Seedream 3.0 reaches 94% text availability for both Chinese and English, with Chinese availability improving 16% over Seedream 2.0.The nearly equivalent availability and hit rates indicate minimal layout- or medium-related rendering errors; the results support the effectiveness of native text rendering over post-processing or external plugins.Source: Figure 11 · p. 13, §3.3 · p. 14
  • The evaluation benchmark contains 180 Chinese and 180 English prompts spanning logos, posters, electronic displays, printed text, and handwriting..Source: §3.3 · p. 13
  • Evaluation uses perceived text availability plus objective text accuracy and hit rates to assess rendering capability.Availability measures acceptable text rendering in context of integration and overall aesthetics; the objective measures use target-character count, minimum edit distance, and correctly rendered characters.Source: §3.3 · pp. 13–14
  • Seedream 3.0 excels at dense text, with precise small-character generation and natural text layouts.Dense text includes long passages with many small characters; Figures 12 and 13 illustrate the capability, while GPT-4o is slated for evaluation in later sections.Source: Figure 12 · p. 13, §3.3 · p. 14, Figure 13 · p. 14

3.4 Photorealistic Portrait

Seedream 3.0 addresses the artificial appearance of generated portraits with more realistic skin texture and native high-resolution output. In a public evaluation, it tied with Midjourney v6.1 for first place, significantly outperforming other models.

3.4 Photorealistic Portrait

  • The portrait evaluation used 100 prompts and Elo battles, with realism and emotion as criteria, comparing Seedream 3.0 against four other models.The competitors were Seedream 2.0, Midjourney v6.1, FLUX-Pro 1.1, and Ideogram 3.0.Source: §3.4 · p. 14
  • After over 50,000 public battle rounds, Seedream 3.0 and Midjourney v6.1 tied for first, significantly outperforming other models.Some model variants are not displayed in Figure 14.Source: §3.4 · p. 15, Figure 14 · p. 15
  • Seedream 3.0 produces realistic skin details such as wrinkles, fine facial hair, and scars, while still generating flawless skin when prompted.Midjourney v6.1’s portrait texture is described as slightly inferior, but its emotional expressions contribute to its high ranking.Source: §3.4 · p. 15, Figure 15 · p. 15
  • Seedream 3.0 can directly generate images at higher resolutions such as 2048×2048; its generated portraits show enhanced texture and promising progress toward professional photography standards.Figure 16 shows higher-resolution human-portrait examples with enhanced texture and clarity.Source: §3.4 · p. 16, Figure 16 · p. 16

3.5 Comparison with GPT-4o

Because large-scale GPT-4o image generation could not be systematically evaluated, selected cases show distinct strengths and weaknesses between GPT-4o and Seedream 3.0. Comparisons also examine SeedEdit against GPT-4o and Gemini-2.0 on image-editing tasks.

3.5.1 Dense Text Rendering

  • Across comparable cases, Seedream 3.0 handles dense Chinese text with ease and outperforms GPT-4o in typesetting and aesthetic composition, while GPT-4o renders small English characters and certain LaTeX symbols more accurately.Source: §3.5.1 · p. 16

3.5.2 Image Editing

  • SeedEdit 1.6 balances editing capability with preservation and prompt following, while GPT-4o struggles to preserve image identity and Gemini-2.0 can have color-naturalness and image-quality issues.GPT-4o handles a wide range of editing requirements; Gemini-2.0 maintains the original image at pixel level. SeedEdit 1.6 is limited on complex multi-image-reference and multi-round editing.Source: §3.5.2 · pp. 16–18, Figure 18 · p. 17
  • SeedEdit detects text in images for precise deletion or modification and integrates added text into the layout, whereas GPT-4o fails to preserve the original image during text editing.The examples cover writing, removing, and modifying text.Source: §3.5.2 · p. 18, Figure 19 · p. 18

3.5.3 Generation Quality

  • GPT-4o-generated images tend to have a dark yellowish hue and significant noise, which impacts usability across scenarios.The comparison assesses color, texture, clarity, and aesthetic appeal.Source: §3.5.3 · p. 18, Figure 20 · p. 19

4 Conclusion · Appendix

Seedream 3.0 uses system-level upgrades across data construction, pretraining, post-training, and acceleration to address limitations in resolution, attribute adherence, typography, aesthetics, and fidelity. The paper reports comprehensive improvements over Seedream 2.0, including native high-resolution output, better text rendering and visual appeal, and faster generation.

4 Conclusion

  • Seedream 3.0 targets limited image resolution, complex attribute adherence, fine-grained typography, and suboptimal visual aesthetics and fidelity.Source: §4 · p. 19
  • Its system-level upgrades span data construction, model pretraining, post-training, and model acceleration.Source: §4 · p. 19
  • The paper reports comprehensive improvements across multiple aspects compared with Seedream 2.0.Source: §4 · p. 19
  • Seedream 3.0 provides native high-resolution image output.Source: §4 · p. 19
  • The paper highlights superior text-rendering quality and enhanced visual appeal.Source: §4 · p. 19
  • Seedream 3.0 is described as offering extreme generation speed.Source: §4 · p. 19
  • Integration into Doubao and Jimeng is presented as showing potential for productivity use across work and daily-life scenarios.Source: §4 · p. 19

Related papers

Loading 2504.11346v3…