acceptodds
Under review as a conference paper at ICLR 2027

MORTIS: Advancing Complex Layout and High-Fidelity Text Rendering in Text-to-Image Generation

Abstract

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 **MORTIS**, 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 **MORTIS** 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. **MORTIS** achieves state-of-the-art performance among open-weight models on **LLTB**, with an average text coverage of 96.3%. On **RLA**, **MORTIS** 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.

Primary area:
generative models
Keywords:
Text-to-image generation, Diffusion Models, Layout-guided Generation, Visual Text Rendering, Multimodal Foundation Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_R2RbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MORTIS combines a 5B diffusion transformer pretrained from scratch with an agentic prompt enhancer for text-rich, layout-controlled image generation. The training recipe includes multi-granularity captions, structured layout annotations, and progressive-resolution training. The prompt enhancer converts short requests into detailed content and layout specifications. The paper also introduces LLTB for bilingual long-text rendering and RLA for human evaluation of complex graphic designs, reporting competitive performance on both.

Critical strengths
  1. This paper targets to address a useful problem: generating text and layouts for practical images such as poster and infographic images.

  2. Training a capable model from scratch represents substantial engineering effort. The reported long-text rendering and layout-control results are promising.

3.The paper provides considerable detail about the training recipe and evaluates both the generator and the complete system.

Critical weaknesses
  1. The distinction from existing work could be clearer. I appreciate the substantial effort involved in building and training this system. However, the paper currently reads more like a technical report, and several design choices have close precedents: Z-Image uses multi-granularity captions and OCR supervision, Qwen-Image adopts progressive training, and Ideogram 4 uses structured JSON conditioning with bounding boxes and LLM-based prompt expansion. These similarities do not diminish the engineering value, but I would appreciate a clearer explanation of the core contribution beyond these existing practices, supported by targeted comparisons or ablations.

  2. The release scope of the data-related resources remains unclear. The paper promises to release the model, training recipe, and evaluation framework, but does not specify whether this includes the data-processing pipeline, filtering configurations, generated captions, structured layout annotations, or final training-data manifest. Access to the original public datasets alone would not be sufficient to reconstruct this supervision. Given the importance of data preparation to this work, the authors should clarify which resources will be released and how others can reproduce the key processing steps.

Questions for authors

Could the authors clarify the core research contribution beyond integrating practices already described in reports such as Z-Image, Qwen-Image, and Ideogram 4? In particular, which methodological distinction or generalizable empirical finding is specific to MORTIS, and what evidence supports it?

This is my main concern. I recognize the substantial engineering effort and strong reported results, and a convincing clarification of the contribution beyond the overall system implementation could lead me to raise my score.

3 comments
AuthorsOpenReview ↗

Response to Reviewer R2Rb

Thank you for the careful review and for recognizing the practical relevance of dense-text and layout generation. Your question about the research contribution is helpful: we should make the empirical findings and the design choices they support more explicit.

Weakness 1 & Question 1: The core contribution and its relationship to prior work.

MORTIS demonstrates that a 5B DiT, paired with a strong language encoder and detailed visual supervision, can achieve strong dense-text rendering and layout generation. The central training recipe is minimal filtering + maximal captioning: retain varied visual data and describe it thoroughly, including its text, line structure, layout, and imperfections. We see three main findings from this study.

1. Preserving data diversity matters for text rendering. We retain images with properties such as watermarks, blur, or compression artifacts and describe those properties in the captions. This preserves information-rich examples that additional quality filtering can remove. Appendix B.2 and Table 10 test this choice directly: at 1.6M training steps, adding heuristic filtering reduces the trained model's OneIG text score from 0.5243 to 0.1180. The result supports a concrete data-design choice for dense-text generation: retain the visual content and provide supervision that describes it accurately.

2. Detailed text and layout conditions make a measurable difference. Appendix C describes how we build multi-granularity captions, verbatim text with line breaks, and structured layout annotations. For structured conditions, Appendix C.3 uses multiple schemas, field dropout, element-order randomization, and several coordinate conventions. This provides varied supervision for a single model that accepts both natural language and structured specifications.

Fig. 4(b) compares structured specifications with natural-language descriptions of the same plans and shows the benefit of making the layout explicit. We have also added a fixed-model test of line-break conditioning: retaining line breaks increases average LLTB coverage from 88.3% to 96.3%, with improvements of 11.8 points for long English text and 10.1 points for long Chinese text. Appendix K, Fig. 10 further illustrates layout transfer to an unseen resolution, where center-relative pixel coordinates better preserve the intended spatial scale than normalized coordinates.

3. A compact diffusion backbone can make effective use of this supervision. The 5B DiT is paired with a separately specified 9B text encoder, which supports language understanding and generalization across input formats. In our additional LLTB comparisons, MORTIS reaches a 96.3% six-subset average, compared with 92.93% for Qwen Image 2.1, 95.43% for MiniMax H3 Image (4096), and 93.69% for MiniMax H3 Image (2048). The two comparison models use 8B and 32B text encoders, respectively. MORTIS's results show that strong text rendering can be achieved with a compact DiT through the combination of language understanding and detailed visual supervision; a larger text encoder alone does not determine the outcome.

The evaluation also makes this recipe easier to study. RLA measures complete systems, Fig. 4(a) compares renderers with shared enhanced prompts, and Fig. 4(b) changes the specification format while fixing the renderer. Together, these experiments examine how planning, conditioning, and rendering contribute to the final image.

Multi-granularity captions, OCR supervision, progressive training, and JSON conditions have precedents in the works you mention. Our contribution is the documented training recipe, the resulting 5B rendering model, and the empirical findings above. We will revise the discussion of Z-Image, Qwen-Image, and Ideogram 4 to make these connections and findings easier to identify.

Weakness 2 & Question 2: Release scope and reproduction of data preparation.

Thank you for asking us to make the release scope concrete. We plan to release the model, training recipe, and evaluation framework. Appendices B and C describe the data processing and annotation procedure, and we will provide reference captioning prompts and the annotation models' inference parameters to help others reproduce that procedure.

The annotated training corpus, including the generated captions and structured layout annotations, is outside the planned release. We will state this explicitly in the paper, together with the resources and settings that will be available.

Thank you again for the constructive feedback. It helped us sharpen the presentation of both the contribution and the resources we intend to share.

AuthorsOpenReview ↗

To Reviewer R2Rb

Dear Reviewer R2Rb,

We hope this message finds you well. As the discussion period draws to a close, we would like to kindly follow up to ensure that we have fully addressed your concerns. If you have any further questions or feedback you would like us to consider, we would be grateful if you could let us know. We particularly appreciate your thoughtful suggestions on clarifying MORTIS’s core contribution beyond existing practices, and sincerely welcome any further guidance on how we could make this distinction clearer.

Thank you very much for the time and effort you have dedicated to reviewing our paper.

Best regards,

The Authors

I recognize the practical contribution of the model and its documented training recipe. However, the core techniques largely follow existing practices, and the experiments do not sufficiently establish a distinct research contribution beyond them. In its current form, I view the work as a useful technical report, but I am not convinced that the research contribution is strong enough to support conference acceptance.

Reviewer_dQihOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes MORTIS, an open text-to-image generation system for complex layouts and high-fidelity text rendering. The system combines a 5B single-stream diffusion transformer pretrained from scratch, which renders both natural-language descriptions and explicit layout specifications through a unified text-conditioning pathway, with an intention-aware agentic prompt enhancement module that routes requests, retrieves knowledge and visual references, and converts brief requests into structured content and layout plans. The paper further introduces benchmarks for long-form bilingual text rendering and end-to-end layout-rich generation. Experiments show that MORTIS achieves state-of-the-art performance among open-weight models on long-text rendering, ranks competitively in end-to-end complex graphic generation, and its backbone performs strongly under shared enhanced natural-language prompts.

Critical strengths
  1. The paper presents a well-documented 5B single-stream DiT pretrained from scratch, where natural-language descriptions and explicit layouts share one text-conditioning pathway, enabling bilingual long-form text rendering and layout control.

  2. It pairs this backbone with an intention-aware agentic prompt enhancer that turns brief requests into executable layout specifications, and supports the system-level claims with LLTB, RLA, and prompt/backbone ablations.

Critical weaknesses
  1. The end-to-end RLA evaluation relies on a proprietary GPT-6 planner and human pairwise judgments, which makes exact reproduction and comparison across systems somewhat difficult. This is consequential because the reported system-level ranking may partly reflect the external planner and rater protocol, although the shared-prompt backbone comparison helps separate the renderer’s contribution.

  2. The benchmark coverage is relatively focused: LLTB uses literary passages, and RLA contains 61 poster, infographic, and UI tasks. This is consequential because the conclusions may not fully generalize to other languages, document genres, or more diverse real-world design requests, though the paper's main claims are still reasonably supported.

Questions for authors
  1. Could the authors clarify how the prompt enhancer checks text fit and line breaks when the requested canvas aspect ratio differs substantially from the training resolutions?

  2. LLTB contains very long text passages, and OCR accuracy may degrade as text length increases, which could affect the reliability of the reported coverage scores. Have you evaluate the accuracy of the OCR-based evaluator itself, for example by reporting a human-verified OCR error rate or agreement metric?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dQih ( 1/2 )

Thank you for the thoughtful review and for recognizing the unified conditioning pathway and the evaluation of both rendering and complete-system performance. We have added OCR verification and provide more detail on text fitting below.

Weakness 1: Reproducibility of the planner and human evaluation.

Appendices F and G describe the prompt-enhancement workflow, baseline configurations, and human evaluation protocol. The system experiment includes the GPT-6 planner as part of the pipeline, while Section 5.4.2 and Fig. 4(a) compare generators using the same enhanced prompts. The shared-prompt comparison therefore gives a common input for assessing rendering across systems. We will make the planner dependency and the distinction between these two comparisons clearer in the presentation.

Weakness 2: Scope of the benchmarks.

LLTB covers English and Chinese at three text lengths, and RLA contains 61 requests spanning posters, infographics, and UI designs. We chose these settings to study dense text and layout generation across different forms of visual design. Extending the evaluation to more languages and document types is a useful next step, and we appreciate the suggestion.

Question 1: Text fit and line wrapping at different aspect ratios.

There are two parts to this: how the renderer handles the requested canvas and how the planner allocates space for text.

DiT. As described in Section 3.3, we preserve image aspect ratios during training. In additional tests, the model generates at ratios ranging approximately from 1:8 to 8:1. The coordinate representation also helps with layout transfer: Appendix K, Fig. 10 illustrates how center-relative pixel coordinates preserve the layout's spatial scale at an unseen resolution better than normalized coordinates in that example.

PE . The requested output width and height are used to calculate the physical dimensions of each bounding box. The planner considers those dimensions together with the intended line breaks and typography hierarchy when deciding whether the text will fit, as described in Appendices F.2–F.3. These checks are performed by the planner; the pipeline has no additional hard constraint solver for text fitting or line wrapping, so the planner can occasionally propose an unsuitable layout.

We have also observed an interesting behavior when the composition implied by a prompt differs from the canvas shape. For example, when asked for a portrait composition on a square canvas, the model sometimes keeps the portrait and adds space on either side. Our data policy retains and captions such compositions, which we believe may help the model learn this behavior.

Question 2: Human verification of the OCR evaluator.

Thank you for suggesting this check. We manually verified OCR transcriptions and recomputed coverage using the corrected text. After applying the evaluator's normalization rules, the English word-level edit rate was 0.244%, and the Chinese character-level edit rate was 0.287%.

LanguageOCR coverageCoverage after human correction
English99.418%99.294%
Chinese97.522%97.677%

Coverage changes by less than 0.16 percentage points in both languages on the audited samples from high-scoring models. In low-scoring outputs, we found frequent malformed or merged glyphs. OCR can sometimes interpret these as intended characters and give them coverage credit, yet these models still score substantially below the high-scoring models.

We also measured agreement at the individual-image level. For each audited image, we computed coverage from the original OCR transcription and from the human-corrected transcription against the same ground-truth text, then performed a paired correlation analysis:

Audited samplesPearson rSpearman ρ
All0.9999250.988022
English0.9999420.993611
Chinese0.9999290.983516

These results show very strong linear and rank correlations between the two coverage measurements on the audited images, complementing the small changes in aggregate coverage reported above.

Controlled tests with rendered text. A language-based OCR evaluator may use its familiarity with a passage to restore the original wording when the image contains different text. To test transcription fidelity and this potential source of bias, we evaluated Qwen2.5-VL-7B-Instruct on 720 programmatically rendered images: 240 with the original text, 240 with substitutions, and 240 with random text of the same length. These controlled images let us compare the OCR output with exactly what was rendered. Coverage is computed against the actual rendered text using a multiset (Counter) intersection that preserves duplicate counts and ignores order, with English words and Chinese characters as matching units. The following scores are weighted by the number of target units:

AuthorsOpenReview ↗

Response to Reviewer dQih ( 2/2 )

Rendered contentEnglishChinese
Original text99.89%99.32%
Selected words/characters replaced99.90%99.26%
Random text of the same length99.96%99.29%

Each condition contains 240 images. Pooling target units across English and Chinese gives overall coverage of 99.45%, 99.41%, and 99.44% for original, substituted, and random text, respectively; these pooled values are not simple averages of the two language scores.

To measure restoration of the original wording directly, we replaced 1%, 2%, 5%, or 10% of the body text. Each rate was assigned to 60 images (30 English and 30 Chinese), and each source passage was assigned to one rate. Replacement positions were sampled without repetition and could be adjacent; titles and authors were unchanged. We define the original-text restoration rate as the number of substituted positions unambiguously aligned to the original word or character in the OCR output, divided by all substituted positions.

Substitution rateImagesEnglish restoration rateChinese restoration rate
1%600/80 (0%)6/272 (2.21%)
2%601/166 (0.60%)13/548 (2.37%)
5%600/397 (0%)11/1380 (0.80%)
10%600/811 (0%)21/2700 (0.78%)
Total2401/1454 (0.069%)51/4900 (1.04%)

Across 6,354 substituted positions, 52 were read back as the original text (0.82%). The remaining positions comprised 6,225 correctly recognized substitutions, 44 other recognition errors, 28 omissions, and 5 cases with uncertain alignment. The latter three categories were excluded from the restoration numerator but retained in its denominator.

The evaluator maintains high coverage across all three conditions, including random text. The substitution test also reveals occasional restoration of familiar Chinese passages. We will report this behavior alongside the coverage results and the human audit to give a fuller account of the evaluator’s reliability.

We follow the OCR-based evaluation protocol of LongText-Bench, an established benchmark for bilingual text rendering used by recent image-generation models such as Qwen-Image (Table 10) and Z-Image (Table 6). As described in Section 5.1 and Appendix G.1, we use Qwen2.5-VL-7B-Instruct as the OCR evaluator, which also reports strong OCR results in its technical report, Table 5: OCRBench 864 and CC-OCR 77.8. We will include the additional verification results in the paper. Thank you again for the helpful feedback.

AuthorsOpenReview ↗

To Reviewer dQih

Dear Reviewer dQih,

We hope this message finds you well. As the discussion period draws to a close, we would like to kindly follow up to ensure that we have fully addressed your concerns. If you have any further questions or feedback you would like us to consider, we would be grateful if you could let us know. We particularly appreciate your thoughtful feedback on evaluation reproducibility and reliability, and sincerely welcome any further suggestions to help us improve the paper. Thank you very much for the time and effort you have dedicated to reviewing our paper.

Best regards,

The Authors

Reviewer_zYMBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MORTIS targets information-dense graphics — infographics, knowledge cards, UI mockups — where a plausible global appearance hides illegible text and incoherent layout. The authors decompose the gap into rendering (reproducing long text, obeying spatial constraints) and specification (turning a ten-word request into content and layout decisions), and address each with one component. The renderer is a 5B single-stream DiT (42 blocks, joint self-attention, SiTU-GLU FFN, 4D RoPE) conditioned on hidden states from a frozen Qwen3.5-9B encoder and a frozen FLUX.2 VAE; natural-language prompts and bbox-JSON layout specifications go through the same text-conditioning path, so one model follows either. Pretraining uses ~3B images after SSCD dedup and AI-image removal, deliberately no quality filtering (artifacts like watermarks and blur are instead described in the caption), gemma4 recaptions at three semantic granularities plus structured layout JSON with verbatim text and preserved line breaks, and three progressive-resolution stages (~4.5M steps total) with Muon–AdamW. At inference, an agent built on GPT-6 routes the request to task modules, retrieves knowledge and visual references, and emits an executable JSON specification. Evaluation is split: the bare backbone on LongTextBench (near-saturated, 96.05–98.85), the new LongLongTextBench (LLTB, 240 bilingual tasks averaging 1,339.8 characters, ~11× LongTextBench), and 7Bench for layout; the full system on the new RealLayoutArena (RLA, 61 short design requests, blind pairwise human votes, Bradley–Terry Elo). Reported results: 96.3 Mean-6 on LLTB (best among open-weight), 0.88 mIoU_det / 0.70 Layout AUC on 7Bench, 1698.70 Elo end-to-end on RLA (second behind GPT Image 2), and 1807.34 Elo for the backbone under shared enhanced prompts (first).

Critical strengths

1). The rendering/specification decomposition is a genuinely useful framing, and the shared-prompt cross-backbone experiment is the right way to test it. Most text-rendering work optimizes the generator and ignores that a user request never specifies what the text should say. Separating the two and then evaluating both jointly (end-to-end) and independently (Fig. 4a: nine generators, one fixed enhanced prompt) is a cleaner experimental design than the field typically offers. Fig. 4(b) additionally isolates prose vs. JSON layout specifications on a shared task set. This is the methodological core of the paper and it holds up. 2). The "annotate rather than filter" recipe is a real, contrarian, evidence-backed finding. Retaining watermarks, blur, and compression artifacts and describing them in the caption is the opposite of standard practice. The supporting measurement — applying a quality filter retains 43% of the corpus and drops the OneIG text score of the retained data from 0.52 to 0.12 — is a concrete demonstration that conventional quality filtering preferentially discards exactly the text-rich, layout-dense images this task needs. One metric is thin evidence for a global pretraining decision, but the observation itself is valuable and, to my knowledge, not previously documented this directly. 3). Benchmark construction is careful where it counts. LLTB's length bins are constructed with DejaVu-Sans metrics on a fixed 1024² canvas with 88px margins, wrapping at 32→8px, so that "long" means actually dense on the page rather than "long but set in 40pt". Preserving line breaks in the captions is a small insight with real consequences ("line breaks are also a form of layout"). Using four samples per task with no best-of-4 selection, and documenting the refusal-handling convention for both numerator and denominator (Table 16), is more rigorous than most benchmark papers.

Critical weaknesses

1). The OCR judge may be reciting the ground truth rather than reading the image. 2). The 7Bench comparison disagrees with a published baseline by 3× and the asymmetry is unresolved. The authors measure Ideogram 4 at mIoU 0.2433 while the vendor reports ≈0.69, and attribute the difference to "model variant, prompt format, and evaluation implementation". That is exactly the set of causes that could also inflate MORTIS's 0.6788. Two asymmetries are visible and uncontrolled. 3). The first contribution — the pretraining recipe — is never isolated. Caption granularity, the no-filtering decision, and line-break preservation have no quantitative ablation — the line-break claim, which the paper itself promotes to a principle, is supported only by a qualitative case study (Appendix J). Meanwhile the most likely single cause of the text-rendering gains — a 9B LLM text encoder, which is excluded from the "5B" label and is much larger than the DiT — is never varied. 4). RLA statistics are too thin, and too clustered, to support the ranking claims. 5). Baseline parity is uneven across all three evaluations, and several table rows look behavioral rather than capability-driven. In the RLA end-to-end group, every participant uses its own prompt pipeline (MORTIS + Layout PE vs. Qwen 3.0 Pro + PE vs. Ideogram + local Qwen3VL-8B PE), so the comparison mixes backbone quality with enhancer quality — only Fig. 4(a) is clean, and it is buried in an ablation.

Questions for authors
  1. Validate the OCR judge.
  2. Reproduce a published 7Bench number before claiming a 3× win.
  3. Decompose the attribution. At 1B scale: encoder size (small vs. 9B), line breaks on/off, caption granularity on/off, quality filtering on/off — all scored on LLTB and 7Bench. This converts contribution #1 from a description into a finding.
  4. Strengthen RLA or soften its claims.
  5. Diagnose the anomalous baseline rows.
  6. Report cost. Steps, batch size, hardware, GPU-hours, and wall-clock for the 4.5M-step 5B run, plus inference latency for the GPT-6-based enhancer. The paper claims to provide "a practical foundation"; practitioners need the price.
  7. Restore a main-text Related Work.
4 comments
AuthorsOpenReview ↗

Response to Reviewer zYMB ( 1/3 )

Thank you for the detailed review and for recognizing the value of separating specification from rendering. To address your questions, we ran additional OCR checks, tested the Ideogram 4 API, and extended the text-rendering comparisons. We report these results below, along with the relevant details from the submission.

Weakness 1 & Question 1: Reliability of the OCR judge.

We follow the OCR-based evaluation protocol of LongText-Bench, an established benchmark for bilingual text rendering used by recent image-generation models such as Qwen-Image (Table 10) and Z-Image (Table 6). We use Qwen2.5-VL-7B-Instruct as the OCR evaluator. Its official results include 864 on OCRBench, 77.8 on CC-OCR, and 56.3/57.2 on the English/Chinese subsets of OCRBench v2 (Qwen2.5-VL Technical Report, Section 3.3.2, Table 5). To examine its behavior in our evaluation, we carried out two additional checks.

Human verification of generated images. We manually checked the transcriptions, distinguishing OCR mistakes from errors in the generated glyphs. After applying the evaluator's normalization rules, the English word-level edit rate was 0.244%, and the Chinese character-level edit rate was 0.287%. Recomputing coverage with the corrected transcriptions gave:

LanguageOCR coverageCoverage after human correction
English99.418%99.294%
Chinese97.522%97.677%

The change is below 0.16 percentage points in both languages, showing that OCR errors have little effect on coverage in the audited samples from high-scoring models. We also inspected low-scoring outputs, where malformed and merged glyphs were common. OCR can sometimes recover an similar character from an illegible glyph, which may raise these models' coverage scores; even with this effect, their scores remain substantially lower.

We also measured agreement at the individual-image level. For each audited image, we computed coverage from the original OCR transcription and from the human-corrected transcription against the same ground-truth text, then performed a paired correlation analysis:

Audited samplesPearson rSpearman ρ
All0.9999250.988022
English0.9999420.993611
Chinese0.9999290.983516

These results show very strong linear and rank correlations between the two coverage measurements on the audited images, complementing the small changes in aggregate coverage reported above.

Controlled tests with rendered text. A language-based OCR evaluator may use its familiarity with a passage to restore the original wording when the image contains different text. To test transcription fidelity and this potential source of bias, we evaluated Qwen2.5-VL-7B-Instruct on 720 programmatically rendered images: 240 with the original text, 240 with substitutions, and 240 with random text of the same length. These controlled images let us compare the OCR output with exactly what was rendered. Coverage is computed against the actual rendered text using a multiset (Counter) intersection that preserves duplicate counts and ignores order, with English words and Chinese characters as matching units. The following scores are weighted by the number of target units:

Rendered contentEnglishChinese
Original text99.89%99.32%
Selected words/characters replaced99.90%99.26%
Random text of the same length99.96%99.29%

Each condition contains 240 images. Pooling target units across English and Chinese gives overall coverage of 99.45%, 99.41%, and 99.44% for original, substituted, and random text, respectively; these pooled values are not simple averages of the two language scores.

To measure restoration of the original wording directly, we replaced 1%, 2%, 5%, or 10% of the body text. Each rate was assigned to 60 images (30 English and 30 Chinese), and each source passage was assigned to one rate. Replacement positions were sampled without repetition and could be adjacent; titles and authors were unchanged. We define the original-text restoration rate as the number of substituted positions unambiguously aligned to the original word or character in the OCR output, divided by all substituted positions.

Substitution rateImagesEnglish restoration rateChinese restoration rate
1%600/80 (0%)6/272 (2.21%)
2%601/166 (0.60%)13/548 (2.37%)
5%600/397 (0%)11/1380 (0.80%)
10%600/811 (0%)21/2700 (0.78%)
Total2401/1454 (0.069%)51/4900 (1.04%)
AuthorsOpenReview ↗

Response to Reviewer zYMB ( 2/3 )

Across 6,354 substituted positions, 52 were read back as the original text (0.82%). The remaining positions comprised 6,225 correctly recognized substitutions, 44 other recognition errors, 28 omissions, and 5 cases with uncertain alignment. The latter three categories were excluded from the restoration numerator but retained in its denominator.

The evaluator maintains high coverage across all three conditions, including random text. The substitution test also reveals occasional restoration of familiar Chinese passages. We will report this behavior alongside the coverage results and the human audit to give a fuller account of the evaluator’s reliability.

Weakness 2 & Question 2: The discrepancy with the published 7Bench baseline.

Thank you for pointing out this discrepancy. The 0.2433 in the submission was obtained with the publicly released FP8 weights, whereas the Ideogram 4 API evaluated under the same protocol yields an mIoU of 0.64, close to the vendor-reported value of about 0.69. As noted by Artificial Analysis, its Ideogram 4.0 evaluation uses the first-party API, which differs from the public weights because the latter include additional safety training and quantization; this likely accounts for the observed gap. We will distinguish the public-weight and API versions explicitly in the table, and the approximately 3× margin noted in the review is relative to the public-weight baseline used in the submission.

Weakness 3 & Question 3: Ablations of the training recipe and the role of the text encoder.

Thank you for these suggestions. The effect of additional quality filtering is studied in Appendix B.2. Within the rebuttal period, we added two analyses that do not require retraining, a fixed-model line-break test and a cross-model text-encoder comparison. A controlled ablation of caption granularity remains future work.

Data filtering and captioning. Appendix B.2 and Table 10 compare models trained with our baseline data policy and with additional heuristic filtering. At 1.6M steps, the OneIG text score drops from 0.5243 to 0.1180 with the additional filtering. Table 10 also reports GenEval, overall OneIG, and DPG-Bench results. Appendix C describes the multi-granularity caption construction and sampling. These details explain the motivation for retaining varied images and describing their content, text, and layout in the conditioning data.

Line-break conditioning. Training conditions include text both with and without explicit line breaks, each with equally high probability, so both input forms are in-distribution for the model; we will state this explicitly in the revision. We evaluated MORTIS on LLTB with the input line breaks (computed from fixed font metrics, Appendix G.1) either retained or removed. The coverage results are as follows.

SubsetWith line breaksWithout line breaks
English, short98.091.6
English, medium97.190.5
English, long97.185.3
Chinese, short96.392.4
Chinese, medium96.788.1
Chinese, long92.482.3
Mean-696.388.3

As shown in the table, explicit line breaks improve Mean-6 coverage by 8.0 points, with the largest gains on long passages (11.8 points in English and 10.1 in Chinese). This gap reflects the benefit of explicit line structure itself.

Text encoder. We use a 9B text encoder to support language understanding and generalization across natural-language and structured inputs. The 5B designation refers to the DiT, with the frozen encoder specified separately. To put encoder scale in context, we additionally evaluated Qwen Image 2.1 and MiniMax H3 Image, whose text encoders have 8B and 32B parameters, respectively:

Model/configurationEN-SEN-MEN-LZH-SZH-MZH-LMean-6
Qwen Image 2.192.9491.6790.9595.4095.0191.6292.93
MiniMax H3 Image (4096)95.7696.5295.2097.9895.7191.3995.43
MiniMax H3 Image (2048)96.9397.1794.0698.8294.5780.6093.69
MORTIS98.097.197.196.396.792.496.3

Resolution note. To accommodate MiniMax H3 Image's generation characteristics, (4096) and (2048) denote images generated at 4096 × 4096 and 2048 × 2048, respectively. Both sets of images are resized to 1024 × 1024 for evaluation.

MORTIS remains competitive across the six subsets and has the highest average in this comparison, including against a model with a substantially larger text encoder. This suggests that encoder size alone is insufficient to explain the rendering results.

AuthorsOpenReview ↗

Response to Reviewer zYMB ( 3/3 )

Weakness 4 & Question 4: Statistical support for the RLA rankings.

The submission reports 95% confidence intervals in Fig. 4 and Sections 5.3 and 5.4.2. Appendix G.2 describes the Bradley–Terry fitting and interval calculation, and Table 17 gives the comparison coverage. The system evaluation contains 1,420 votes across all 55 model pairs, and the shared-prompt generator evaluation contains 576 votes across all 45 model pairs.

MORTIS's system-level Elo is 1698.70 [1655.96, 1741.43], and its shared-prompt generator Elo is 1807.34 [1727.44, 1887.24]. The ranks summarize the fitted scores for these RLA settings, with uncertainty shown alongside them. We use them to describe the observed ordering; significance between individual neighboring models should be assessed using the corresponding uncertainty.

We would be grateful if you could point us to the particular comparison or aspect of the sampling that you find too thin or clustered. This would help us address your concern directly. We will also bring the vote coverage and confidence intervals closer to the ranking discussion for easier reference.

Weakness 5 & Question 5: System comparisons and the shared-prompt experiment.

RLA evaluates how well a complete system turns a brief user request into a finished image. For this comparison, each model uses its recommended prompt-enhancement setup where available, including the corresponding workflow for closed systems. Planning the content and layout is part of the capability being measured.

The generator comparison in Section 5.4.2 and Fig. 4(a) holds the enhanced natural-language prompts shared across models and applies no further model-specific enhancement. Fig. 4(b) then compares specification formats with the rendering backbone fixed. These experiments are in the main paper and provide complementary views of system quality, rendering capability, and specification quality. Thank you for suggesting that we give the shared-prompt results greater prominence; we will connect them more directly to the main RLA comparison.

For the baseline results described as behavioral rather than capability-driven, could you identify the models or table entries you have in mind? We would be happy to examine those cases and discuss the behavior behind them.

Question 6: Training cost and prompt-enhancement latency.

Table 1 reports 2.1M, 2.1M, and 0.31M training steps for the three stages, totaling approximately 4.51M steps. Appendix A gives the packing setup: up to 8,576 image-and-text tokens per device per microbatch, 512 devices, and two gradient-accumulation steps, corresponding to a nominal capacity of 8,781,824 tokens per update. The image count per update varies with the packed image and text lengths.

The overall development period, including the various training attempts, was approximately 3–4 months. The main training path consumed approximately 335K H200-equivalent GPU-hours. On our internal testing platform, the GPT-6-based prompt enhancer took 24.29 seconds per request on average for the PE stage. We will add these cost and latency measurements to the practical details of the system.

Question 7: Related Work in the main text.

Thank you for the suggestion. We placed Related Work in Appendix L to meet the nine-page main-text limit. We will restore this Related Work to the main text in the rebuttal revision so that the connections to prior work are easier to follow.

Thank you again for the careful review. Your questions helped us strengthen the empirical checks and explain the evaluation setup more clearly.

AuthorsOpenReview ↗

To Reviewer zYMB

Dear Reviewer zYMB,

We hope this message finds you well. As the discussion period draws to a close, we would like to kindly follow up to ensure that we have fully addressed your concerns. If you have any further questions or feedback you would like us to consider, we would be grateful if you could let us know. Your insights are invaluable to us, and we sincerely welcome any additional suggestions you may have to help us improve the paper.

Thank you very much for the time and effort you have dedicated to reviewing our paper.

Best regards,

The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Responses(1/3)

Dear AC/SAC/PC,

We sincerely thank you and the reviewers for the time and effort devoted to evaluating our submission. We posted our responses on November 12 and have received follow-up feedback from one reviewer on November 19. We summarize below the strengths identified in the original reviews, the main concerns, and the experiments and revisions provided in response, for your consideration in assessing our submission. Section, figure, and table references below refer to the revised manuscript.

Key Strengths Recognized by the Reviewers

We appreciate the reviewers' recognition of the practical value, methodological design, and empirical findings of MORTIS. Their reviews highlight the following strengths. Below, S, W, and Q refer to the numbered strengths, weaknesses, and questions in the original reviews.

  1. A 5B backbone pretrained from scratch for bilingual long-text rendering and layout control (dQih [S1]; R2Rb [S2]). Reviewer dQih highlights the well-documented 5B single-stream DiT that jointly supports bilingual long-form text rendering and explicit layout control through a unified conditioning pathway. Reviewer R2Rb recognizes the substantial effort involved in training the model from scratch and its promising rendering results.

  2. An evidence-backed finding on data retention (zYMB [S2]). Reviewer zYMB specifically recognizes the value of retaining images with watermarks, blur, and compression artifacts while describing these properties in captions. The review highlights the filtering result as a concrete, nonstandard empirical observation relevant to dense-text generation.

  3. A meaningful separation of specification and rendering (zYMB [S1]). Reviewer zYMB identifies this decomposition and its experimental validation as the methodological core of the paper. In particular, the review endorses comparing generators under shared enhanced prompts and comparing prose versus structured specifications with the renderer fixed.

  4. An agentic system that turns brief requests into executable layouts (dQih [S2]). Reviewer dQih recognizes the pairing of the renderer with an intention-aware prompt enhancer, which converts short requests into content and layout specifications, and notes the supporting LLTB, RLA, and prompt/backbone experiments.

  5. Careful benchmark construction and detailed experimental documentation (zYMB [S3]; R2Rb [S3]). Reviewer zYMB specifically highlights LLTB's font-metric-based density construction, preserved line breaks, four samples per task without best-of-four selection, and documented refusal handling. Reviewer R2Rb recognizes the detailed training recipe and evaluation of both the generator and complete system.

  6. A useful problem with practical relevance (R2Rb [S1]). Reviewer R2Rb recognizes the value of generating text and layouts for practical images such as posters and infographics.

AuthorsOpenReview ↗

General Responses(2/3)

Main Concerns and Our Responses

Q1. What is the research contribution beyond existing practices, and how does MORTIS differ from prior systems? (R2Rb [W1, main question]; zYMB [Q7])

Response. We strengthened the discussion of two complementary contributions: the training methodology and empirical findings for dense-text and layout generation, and the resulting 5B pretrained backbone.

Training methodology and empirical findings. MORTIS jointly trains natural-language captions and structured layout specifications through a shared conditioning pathway, enabling the same renderer to support both input forms. Section 7 clarifies this distinction from Ideogram 4, which reports training exclusively on structured JSON captions, while Appendix C documents layout annotation construction and augmentation through multiple schemas, field dropout, element-order randomization, and coordinate conventions. Our controlled filtering study shows that additional heuristic filtering reduces the OneIG text score from 0.5243 to 0.1180 at 1.6M training steps (Appendix B.2). Separately, a fixed-model experiment shows that explicit line breaks improve LLTB coverage by 7.92 percentage points at inference (Appendix J), and the specification-format comparison evaluates structured layouts against prose descriptions with the renderer held fixed (Fig. 4(b)). These experiments provide concrete evidence for broad data retention and explicit text and layout conditioning.

A capable 5B backbone for further research. Pretrained from scratch, MORTIS combines bilingual long-text rendering with explicit layout control, achieving 96.3% LLTB coverage and the highest fitted score in the shared-prompt renderer comparison (Table 4; Fig. 4(a)). Its open weights provide a foundation for studying text rendering, layout conditioning, and alternative planning methods without repeating the full pretraining effort.

Q2. What evidence supports the training recipe, line-break conditioning, and the role of the text encoder? (zYMB [W3, Q3]; R2Rb [W1])

Response. We clarified the controlled early-stage filtering study in Appendix B.2: at 1.6M training steps, additional heuristic filtering reduces the trained model's OneIG text score from 0.5243 to 0.1180. Separately, the new fixed-model experiment changes only the input specification: explicit line breaks raise LLTB Mean-6 coverage from 88.35% to 96.26%, a gain of 7.92 percentage points computed before rounding (Appendix J; Table 26). Gains on long English and Chinese text are 11.75 and 10.16 points. This experiment measures the benefit of explicit line structure at inference. The 7.92-point mean gain and 10.16-point long-Chinese gain correct the rounded figures of 8.0 and 10.1 in our November 12 responses. Additional LLTB comparisons give 96.3% for MORTIS, 92.9% for Qwen Image 2.1, and 95.4% for MiniMax H3 Image (4096) (Table 4; Appendix G.1). We separately specify the 5B DiT and frozen 9B text encoder. The cross-model comparisons provide context for encoder scale; controlled encoder-size and caption-granularity training ablations remain open.

Q3. Is the OCR evaluator reliable, or could it restore familiar text instead of reading the image? (zYMB [W1, Q1]; dQih [Q2])

Response. Appendix G.1.1 (Tables 17–20) adds two complementary checks. First, a human audit covers 25,130 characters, excluding spaces and line breaks, in outputs from six models: MORTIS, GPT Image 2, Qwen Image 2512, ERNIE Image, Krea 2 Raw, and FLUX.2. The English word-level and Chinese character-level edit rates are 0.244% and 0.287%; correcting the OCR transcriptions changes coverage by less than 0.16 percentage points in either language. Per-image coverage agreement is high (Pearson r > 0.9999; Spearman ρ = 0.988). Second, controlled tests on 720 programmatically rendered images yield pooled coverage above 99.4% for original, substituted, and random text. Restoration errors occur at 52 of 6,354 substituted positions (0.82%), comprising 1 of 1,454 English positions and 51 of 4,900 Chinese positions. Together, these checks strengthen confidence in the reported coverage on the audited samples and directly quantify the evaluator's tendency to restore familiar text.

Q4. Why does the Ideogram 4 baseline differ from its published result? (zYMB [W2, Q2])

Response. We evaluated the official Ideogram 4 API on all 224 7Bench tasks under our protocol. It obtains 0.6445 mIoU, closer to the vendor-reported value of approximately 0.69 than the public FP8 result of 0.2433. Tables 3 and 23 now distinguish the API and public FP8 versions. Against the API, MORTIS achieves 0.6788 vs. 0.6445 mIoU and 0.7031 vs. 0.6580 Layout AUC, with similar mIoU over detected objects (0.8786 vs. 0.8776). This additional comparison directly addresses the baseline-version discrepancy raised in the review.

AuthorsOpenReview ↗

General Responses(3/3)

Q5. Do the system comparisons confound planner quality with renderer quality? (zYMB [W5, Q5]; dQih [W1])

Response. We made the distinction among the three evaluations more prominent: RLA evaluates complete systems with their respective prompt pipelines; Fig. 4(a) uses shared enhanced natural-language prompts with no further model-specific enhancement; and Fig. 4(b) varies specification format with the renderer fixed. The shared-prompt experiment isolates renderer comparisons from differences among prompt-enhancement pipelines. We also clarified the proprietary planner dependency and baseline configurations. Appendix G.1 documents 61 refusal images among the 160 EN-short generations from Ideogram 4 at CFG 4; excluding refusals raises its Mean-6 coverage from 63.47% to 68.05% (Table 16). Appendix G.4 documents 22 safety-blocked GPT Image 2 requests among the 224 7Bench tasks and their treatment as zero-IoU outputs.

Q6. Are the RLA rankings sufficiently supported, and how broadly do the benchmarks generalize? (zYMB [W4, Q4]; dQih [W2])

Response. Section 5 now reports comparison coverage alongside the rankings: the system evaluation contains 1,420 votes across all 55 model pairs, and the shared-prompt evaluation contains 576 votes across all 45 pairs (Table 21). The rankings describe fitted Bradley–Terry scores under these evaluation settings; the reported 95% intervals are asymptotic estimates and do not account for clustering of votes by task (Appendix G.2). The benchmark scope is aligned with the paper's focus on dense-text rendering and complex layout generation. LLTB evaluates English and Chinese at three text lengths, using literary passages to test faithful reproduction of long, continuous text. RLA complements this with 61 poster, infographic, and UI requests that require coherent organization of text and visual elements. Together, they evaluate the two capabilities central to MORTIS: rendering specified text accurately and turning brief requests into complete, layout-rich designs. We also report results on the public LongTextBench and 7Bench benchmarks (Table 22; Tables 3 and 23).

Q7. Which resources will be released, and how can the supervision be reconstructed? (R2Rb [W2])

Response. We clarified the scope of the resources supporting reproduction: the model weights, training recipe, and evaluation framework, together with reference captioning prompts and annotation-model inference settings that complement the processing and annotation procedures in Appendices B and C. The model weights are available, and the accompanying resources are specified in the Reproducibility Statement. The annotated training corpus, including generated captions and structured layout annotations, is outside the release scope.

Q8. How are text fit and line breaks handled at different canvas aspect ratios? (dQih [Q1])

Response. Appendix F.3 explains how the planner uses canvas dimensions, bounding-box sizes, intended line breaks, and typography hierarchy to allocate text. These are planner judgments, with no additional hard constraint solver or guaranteed fit. The final training stage preserves source aspect ratios (Section 3.3), and Appendix F.3 describes additional generation tests spanning approximately 1:8 to 8:1, without establishing uniform text-fitting reliability across that range. Appendix K illustrates coordinate transfer to an unseen resolution.

Q9. What are the training cost and prompt-enhancement latency? (zYMB [Q6])

Response. The main training path used approximately 335K H200-equivalent GPU-hours over 4.51M training steps (Table 1; Appendix A.2). The packing configuration of up to 8,781,824 image-and-text tokens per update across 512 devices applies to Stage 3. The reported 3–4 months cover model development and training, including the different training attempts during this period, but exclude earlier preparation and preliminary experiments. This is the development period rather than the duration of a single uninterrupted training run. The average latency of 24.29 seconds per request measures prompt enhancement only and excludes image generation (Appendix F.1).

The revised manuscript incorporates these experiments and clarifications, with additions and substantive revisions marked in blue. The additional evidence strengthens both the case for MORTIS as a capable 5B backbone for future research and the empirical findings underlying its training and conditioning design. We respectfully ask that these responses and revisions be considered alongside the original reviews in the final assessment. Thank you again for your time and consideration.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.