PaperlayerView exact revision on arXiv

Source-linked AI summary

MOSS-Image Technical Report

MOSI. AI

arXiv:2610.23374v1cs.CV

TL;DR

Open-weight text-to-image models still struggle with dense text and intricate layouts, and concise requests leave content and spatial organization underspecified. MOSS-Image pairs a pretrained text-and-layout diffusion transformer with an intent-aware prompt enhancer and introduces benchmarks for rendering and end-to-end generation. It achieves leading open-weight LLTB performance and ranks second on RLA end to end, while its backbone ranks highest under shared enhanced prompts.

  • Problem

    Open-weight text-to-image models struggle with dense text and intricate layouts, while concise requests often leave content, information hierarchy, and spatial organization unspecified.

    Source: §1 · p. 1

  • Method

    MOSS-Image pairs a 5B diffusion transformer pretrained from scratch for text and layout rendering with an agentic enhancer that turns concise requests into detailed generation specifications.

    Source: §1 · p. 2

  • Results

    96.3 Mean-6 on LLTB gives MOSS-Image state-of-the-art open-weight performance; it ranks second on RLA end to end, while its backbone ranks highest under shared enhanced prompts.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The results support combining a capable renderer with explicit generation specifications for complex graphics.

    Source: §9 · p. 10

    The backbone does not directly support reference-image conditioning or image editing, and prompt-enhancer quality can be limited by external language models’ aesthetic judgment, reference interpretation, and factual errors.

    Source: §8 · p. 10

Abstract

from arXiv · show

Recent closed-source image generation systems can produce diverse visual content from simple compositions to information-dense graphics, yet open-weight text-to-image models still struggle with dense text and intricate layouts. To narrow this gap, we introduce MOSS-Image, an open generation system that combines high-fidelity visual synthesis with intent-aware composition. First, we pretrain a 5B single-stream diffusion transformer (DiT) from scratch with a carefully designed recipe, enabling bilingual long-form text rendering and precise layout control. Second, we develop an intention-aware agentic prompt enhancement (PE) module that translates diverse user requests into detailed content and layout specifications. Together, they enable MOSS-Image to generate text-rich and layout-coherent images. We further introduce LongLongTextBench (LLTB) for evaluating long-form English and Chinese text rendering, and RealLayoutArena (RLA) for end-to-end layout-rich generation. MOSS-Image achieves state-of-the-art performance among open-weight models on LLTB, with an average text coverage of 96.3%. On RLA, MOSS-Image ranks second in the end-to-end system comparison, while its backbone ranks first among generators evaluated with shared enhanced prompts. We will release the model, training recipe, and evaluation framework to support further research and promote the text-to-image community.

1 Introduction

MOSS-Image addresses open-weight models’ difficulty with dense text and intricate layouts by pairing a text-and-layout-capable generator with an agent that expands concise requests. It introduces benchmarks for long-form text rendering and end-to-end graphic generation.

1 Introduction

  • Open-weight models still struggle with dense text and intricate layouts, while concise requests often leave content, information hierarchy, and spatial organization unspecified.A plausible overall appearance does not guarantee legible text, complete content, or coherent spatial organization.Source: §1 · p. 1
  • MOSS-Image pairs a 5B single-stream diffusion transformer pretrained from scratch with an agent that enriches sparse requests and designs deliberate layouts.The training recipe combines broad data retention, structure-aware supervision, and progressive-resolution training.Source: §1 · pp. 2–3
  • LLTB evaluates extended English and Chinese text rendering, while RLA uses blind pairwise human comparisons to assess end-to-end graphics from short design requests.RLA jointly assesses instruction following, content completeness, text readability, and visual organization; shared-prompt comparisons examine specification and generator roles.Source: §1 · p. 2
  • 96.3 Mean-6 on LLTB gives MOSS-Image state-of-the-art performance among open-weight models; it ranks second on RLA and its backbone ranks highest under shared enhanced prompts.Prompt comparisons also report higher human preference for detailed specifications than concise requests.Source: §1 · p. 3
  • The intention-aware agent routes requests to task-specific guidance, retrieves knowledge and visual references, and produces structured layout specifications for the generator.It can turn requests of fewer than ten words into detailed specifications while adapting information density to the request.Source: §1 · pp. 2–3

2 System Overview

MOSS-Image combines a pretrained text-and-layout generator with an intent-aware prompt enhancer. The enhancer specifies content and layout from concise requests, and the generator renders those specifications through a shared text-conditioning interface.

2 System Overview

  • MOSS-Image pairs a pretrained text-and-layout generator with an intent-aware enhancer that turns concise requests into content and layout specifications for image rendering.The two modules communicate through the generator’s text-conditioning interface.Source: §2 · p. 3
  • The pretrained 5B single-stream DiT supports text rendering and spatial constraints through one text-conditioning pathway for natural-language descriptions and structured layouts.Broad data retention, multi-granularity and structure-aware supervision, and progressive training build this capability.Source: §2 · p. 3
  • The prompt enhancer resolves choices left unspecified by short requests by selecting task-specific guidance, retrieving knowledge and visual references as needed, and constructing detailed specifications.Its layout output uses the same textual interface learned by the generator.Source: §2 · p. 3

3 Pretraining Strategy

MOSS-Image’s pretraining strategy develops a text-and-layout renderer through corpus construction, multi-granularity and structure-aware supervision, and staged training. It retains broad visual coverage while adding textual and spatial supervision.

3 Pretraining Strategy

  • The 5B single-stream DiT jointly processes text tokens and image latents, using joint self-attention, shared modulation, and SiTU-GLU feed-forward layers.The recipe develops the renderer; the inference-time enhancer supplies content and layout specifications through its conditioning interface.Source: §3 · pp. 3–4

3.1 Pretraining Corpus Construction

  • After deduplication and AI-image removal, the corpus receives no additional filtering; undesirable attributes are described rather than used as exclusion criteria.This choice aims to retain broad visual coverage, while allowing residual watermarks, blurry faces, and compression artifacts into textual supervision.Source: §3.1 · p. 5

3.2 Multi-granularity and Structure-aware Supervision

  • Training uses short, medium, and detailed captions, plus structured annotations with localized bounding boxes and, where applicable, transcribed text and typography attributes.Preserving annotated line breaks provides explicit supervision for how multi-line text is organized in images.Source: §3.2 · pp. 5–6

3.3 Pretraining Schedule and Optimization

  • Training increases image resolution and text-context capacity across three stages, while caption and layout-annotation sampling depends on available descriptions and domain-specific policies.The final stage uses a 4096-token text limit; standard caption sampling uses original captions and recaptions at a 20:80 ratio when both are available.Source: §3.3 · p. 6
  • A 1B ablation reports lower denoising loss with hybrid Muon–AdamW than AdamW in early and late training, using 33.2% and 28.6% fewer steps, respectively.The RoPE design retains a reference-index axis for future reference-conditioned extensions; without reference inputs, that axis remains unrotated.Source: §3.3 · p. 6

4 Agentic Prompt Enhancement

The prompt enhancer turns underspecified requests into content and layout conditions the generator can execute. It routes by task, retrieves relevant knowledge or visual references when needed, and validates structured layout plans.

4 Agentic Prompt Enhancement

  • The enhancer converts concise requests into content and layout conditions before rendering.The request may leave literal content and visual organization unspecified.Source: §4 · p. 6
  • GPT-6 selects task-specific modules while a shared core maintains the JSON interface and validation.Modules can be extended or customized without changing the interface or pretrained DiT.Source: §4 · p. 7
  • The agent retrieves relevant knowledge and inspects visual references as needed, adapting the specification to the target canvas.References guide planning but do not directly condition the DiT.Source: §4 · p. 7
  • The enhancer specifies text, elements, bounding boxes, typography, and spatial relationships in JSON, then checks constraints, completeness, consistency, and text fit.The validated specification passes through the generator’s shared text-conditioning interface.Source: §4 · p. 7

5 Experiments

Experiments evaluate the pretrained backbone on text rendering and layout control, and compare complete systems and prompt configurations on RLA. MOSS-Image leads open-weight models on LLTB and shared-prompt generator ratings, while ranking second among end-to-end systems.

5.1 Experimental Setup and Benchmarks

  • LLTB targets average 1,339.8 characters, approximately 11.1× LongTextBench’s 120.3; its shortest target is 444 characters.The evaluation uses six equally averaged language–length groups; LongTextBench scores cluster near saturation, limiting discrimination of sustained long-text rendering.Source: §5.1 · p. 8
  • RLA compares complete systems on concise design requests, with raters jointly judging instruction following, rendering quality, and spatial organization.The backbone evaluations omit the agentic prompt enhancer, and the pretrained results do not isolate the causal contribution of individual pretraining choices.Source: §5.1 · pp. 7–8, §5.2 · p. 8

5.2 Base DiT Evaluation

  • 96.3 Mean-6 on LLTB is the highest score among the open-weight models listed; the longest English and Chinese scores are 97.1 and 92.4.Text recall is high overall but lower for the longest Chinese texts.Source: §5.2 · p. 8
  • 0.88 mIoUdet and 0.70 Layout AUC lead evaluated models, compared with Ideogram 4’s 0.44 and 0.20.The evaluation measures mean IoU over detected target objects and Layout AUC.Source: §5.2 · p. 8

5.3 End-to-End Evaluation on RLA

  • 1698.70 Elo places MOSS-Image second on RLA, behind GPT Image 2 (1844.35) and ahead of Seedream 5.0 Pro (1661.16).The 95% confidence interval is 1655.96–1741.43; Figure 4(c) reports the end-to-end comparison.Source: §5.3 · p. 8, Figure 4 · p. 9

5.4 Ablation Studies

  • The prompt-strategy comparison tests short and detailed prompts, direct expansion, and natural-language or JSON layout plans.The experiments also compare MOSS-Image and other generators using shared enhanced natural-language prompts.Source: §5.4.1 · p. 9, §5.4 · p. 9

5.4.1 Prompt Strategies

  • JSON layout receives the highest prompt-strategy rating at 1622.87 Elo; detailed and direct-expansion prompts score 1509.21 and 1541.45, versus 1081.70 for short requests.Natural-language layout scores 1506.20 Elo.Source: §5.4.1 · p. 9

5.4.2 Backbones under Shared Enhanced Natural Prompts

  • Under shared enhanced natural-language prompts, MOSS-Image leads nine generators at 1807.34 Elo, ahead of ERNIE Image (1748.51).The comparison uses the same natural-language rewrite of each Layout PE specification without additional prompt enhancement.Source: §5.4.2 · p. 9

6 Case Study

The case studies illustrate layout-guided edits, line-break effects on text rendering, and coordinate choices for layout generalization.

6 Case Study

  • Layout-guided regeneration supports targeted content replacement and element rearrangement while retaining the overall design.The edits require neither the original image as conditioning nor additional training.Source: §6 · p. 9
  • Explicit line breaks help establish each line’s character content earlier during denoising, leading to better final text rendering.The comparison is qualitative.Source: §6 · p. 9
  • Zero-centered absolute bounding-box coordinates generalize better to unseen resolutions than coordinates normalized to [0, 1000].The comparison is qualitative.Source: §6 · p. 10

7 Related Works

The detailed related-work discussion is provided in Appendix L.

7 Related Works

  • The detailed related-work discussion is provided in Appendix L.The section points readers to the appendix for this discussion.Source: §7 · p. 10

8 Limitations

The backbone is currently limited to text-to-image generation: it does not directly accept reference images or support image editing. The prompt enhancer also depends on external language models whose limitations can reduce specification quality.

8 Limitations

  • The backbone does not directly condition on reference images or support image editing.Image-conditioned generation and iterative design refinement remain future work under the current computational budget.Source: §8 · p. 10
  • Image-conditioned generation and iterative design refinement would broaden the system's use in creative workflows.These capabilities remain future work under the current computational budget.Source: §8 · p. 10
  • External language models can weaken prompt-specification quality through limitations in aesthetic judgment, reference interpretation, and factual accuracy.This can occur despite faithful rendering by the system.Source: §8 · p. 10

9 Conclusion

MOSS-Image combines a from-scratch 5B diffusion transformer with an agentic prompt enhancer, and the conclusion reports strong open-weight long-text performance alongside a second-place end-to-end RLA ranking. The appendices describe the model, training, and corpus preparation behind the system.

9 Conclusion

  • The system pairs a 5B diffusion transformer pretrained from scratch with an agentic enhancer that turns diverse requests into detailed generation specifications.The backbone is described as supporting bilingual long-form text rendering and layout control.Source: §9 · p. 10
  • 96.3 Mean-6 on LLTB gives MOSS-Image state-of-the-art performance among open-weight models; it ranks second on end-to-end RLA while its backbone rates highest with shared enhanced prompts.Prompt comparisons show detailed content and design instructions improve human preference over concise requests.Source: §9 · p. 10

A.1 Model Configuration

  • The 5B designation refers to the diffusion generator, excluding the frozen text encoder and VAE from that designation.Source: §A.1 · p. 14

A.2 Training Configuration

  • Pretraining uses three stages with increasing resolution and text length, ending at step 310,000 with EMA weights.The final-stage text condition limit is 4096 tokens, while training uses a hybrid Muon–AdamW optimizer.Source: §A.2 · pp. 14–15

B Pretraining Dataset Details

  • The corpus combines multiple image-text sources, uses SSCD-based deduplication and an in-house AI-image detector, and preserves source proportions.Captions are predominantly English, without dedicated Chinese-caption annotation; the additional heuristic filter studied separately is not part of this corpus-preparation stage.Source: B Pretraining Dataset Details · p. 16, §B.1 · p. 16

B.2 Preliminary Filtering Experiments

The preliminary filtering comparison found worse generation quality after additional filtering, informing the decision to retain broad data coverage. The appendices also describe multi-granularity captions, annotation checks, and structured layout conditions used for training.

B.2 Preliminary Filtering Experiments

  • 0.5243 to 0.1180: the final OneIG text score fell after additional filtering, which also worsened GenEval and FDR6-LAION while leaving DPG-Bench nearly unchanged.The filtered model had a higher, worse FDR6-LAION score at all 32 evaluated checkpoints; the authors retain broad data coverage and omit the additional heuristic filter.Source: Table 10 · p. 17, §B.2 · p. 17, Figure 5 · p. 18

C.1 Caption Generation and Sampling

  • Caption generation supplies short, medium, and detailed descriptions that vary in semantic detail, and preserves visible text's wording and layout attributes.When all three are available, training samples uniformly among the nonempty description lengths.Source: §C.1 · pp. 18–19

C.2 Annotation Validation

  • Caption validation checks JSON syntax and three nonempty descriptions; layout validation checks required fields, bounding boxes, coordinate ranges, identifiers, references, and color formats.The annotation generators allow up to three retries for invalid outputs.Source: §C.2 · p. 19

C.3 Conditioning Schemas and Coordinate Conventions

  • Structured conditions combine global descriptions with localized content and boxes, using scene/element or summary/items schemas, plus a supplementary item-level format.The scene/element branch retains global brief and elements while dropping optional fields and randomizing element order; the summary/items branch discards element identifiers and parent references.Source: §C.3 · p. 19

D RoPE Ablation and Reference-Axis Design

The RoPE ablations show similar overall image-generation performance across coordinate designs, with trade-offs among metrics. The chosen sequential coordinates preserve a dedicated reference-index axis for possible extension to reference-conditioned generation.

D RoPE Ablation and Reference-Axis Design

  • The sequential design keeps r=0 for all tokens and places image sequence positions after text positions, preserving r for reference indexing.Without reference inputs, this makes the associated rotations identity transformations and leaves that feature subspace unrotated.Source: D RoPE Ablation and Reference-Axis… · p. 21
  • The comparison evaluates DPG-Bench, GenEval, and OneIG, with OneIG averaging alignment, text, reasoning, style, and diversity scores.The four designs are compared at 200K steps, with longer comparisons spanning ten checkpoints from 800K to 1.25M steps.Source: D RoPE Ablation and Reference-Axis… · p. 21
  • Sequential text–image coordinates retain nearly the baseline DPG-Bench score at 200K steps while improving OneIG and lowering GenEval; longer comparisons show r-axis separation improves OneIG but reduces DPG-Bench and GenEval.The authors retain a dedicated reference-index axis for extensibility.Source: Table 13 · p. 21, D RoPE Ablation and Reference-Axis… · p. 21

E Optimizer Ablation: Training-Loss Convergence

This section presents supplementary analyses of MOSS-Image, including optimizer convergence comparisons across training stages and aggregation widths, rubric comparisons, qualitative tests of layout and text control, and the relationship of its training and evaluation contributions to prior work.

E Optimizer Ablation: Training-Loss Convergence

  • Convergence is measured by interpolating the first crossing of a target loss on non-overlapping 10K-step bins of training denoising mean-squared error.Loss values recorded every 10 steps are averaged in the bins, which require at least 90% coverage.Source: E Optimizer Ablation: Training-Loss… · p. 22
  • Muon–AdamW’s loss gap remains positive from 500K to 1.6M steps, ending at 1.655 × 10−3; final mean losses are 0.40082 versus 0.40247 for AdamW.The figure presents the matched-run loss gap and convergence curves across early and late training.Source: Figure 7 · p. 22, E Optimizer Ablation: Training-Loss… · pp. 22–23
  • Muon–AdamW uses 33.2% fewer updates than AdamW to reach loss 0.41 and 28.6% fewer to reach loss 0.403.The corresponding AdamW/Muon–AdamW step ratios are 1.50× at loss 0.41 and 1.40× at loss 0.403.Source: E Optimizer Ablation: Training-Loss… · p. 22
  • First-crossing step ratios (AdamW/Muon–AdamW) remain above 1 across 5K, 10K, and 20K bins: 1.37, 1.50, and 1.47 at loss 0.41, and 1.35, 1.40, and 1.35 at loss 0.403.This sensitivity analysis finds the convergence advantage at both early and late loss targets across all three aggregation widths.Source: E Optimizer Ablation: Training-Loss… · p. 23

H Supplementary Rubric-Based Evaluation

  • 94.93 for MOSS-Image versus 97.72 for GPT Image 2 in the system comparison; under shared enhanced prompts, MOSS-Image scored 87.62, followed by SenseNova U1 at 79.51.The shared-prompt comparison evaluates generators, while the system comparison evaluates end-to-end systems.Source: H Supplementary Rubric-Based Evaluation · p. 30, Table 21 · p. 31
  • A 0.52-point spread separates the rubric scores for enhancement without agentic planning (95.59), detailed prompts (95.46), and JSON layout (95.07), so human preferences provide the primary ranking evidence.The rubric is less sensitive to fine differences in complex content organization and visual design; human evaluation also provides an assessment independent of GPT-6, which serves as planner and automated judge.Source: H Supplementary Rubric-Based Evaluation · p. 30

I Layout-Guided Control: Case Studies

  • The editing cases show coordinated layout changes: infographic steps 2 and 3 switch order with their content, while a poster title changes from KRASNOYARSK to CALIFORNIA and retains its typographic composition.The infographic keeps its seven-panel structure and visual elements recognizable; the poster preserves its photograph-filled lettering style and overall design, though smaller city labels remain unchanged.Source: I Layout-Guided Control: Case Studies · p. 32

J Line-Break Alignment: Denoising Case Study

  • With explicit line breaks, many lines become distinguishable by denoising step 8 and clearer thereafter; without them, rows remain ambiguous and malformed characters persist in the final displayed step.Both runs use 30 denoising updates with timestep shift s=3.Source: J Line-Break Alignment: Denoising Case… · p. 33

K Bounding-Box Coordinates: Generalization to Unseen Resolutions

  • At an unseen higher resolution, center-relative pixel coordinates distribute content across the canvas and better preserve the lower-resolution design’s spatial scale, while normalized coordinates concentrate it centrally with wider margins.The comparison uses normalized coordinates in [0, 1000] and zero-centered absolute coordinates following a center-relative pixel convention.Source: K Bounding-Box Coordinates:… · p. 34

L Related Work

  • MOSS-Image distinguishes its work from Qwen-Image-Agent [12] by pairing agent-produced layout specifications with a newly pretrained text-rich renderer, and from Ideogram 4 by documenting a 5B pretraining recipe.The report describes corpus construction, supervision design, and the training schedule for bilingual long passages and dense graphic layouts.Source: L Related Work · p. 35
  • LLTB evaluates bilingual rendering of longer passages, while RLA jointly assesses content, layout, and text quality in complete systems responding to realistic graphic-design requests.RLA is presented as an evaluation of end-to-end system performance.Source: L Related Work · p. 35

Related papers

Loading 2610.23374v1…