{"id":"5c445ad3-7f65-4d8b-8f5b-c2fc9acd16b1","arxiv_id":"2411.18159","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Type-R automatically detects and corrects typographic errors in text-to-image outputs using OCR, inpainting, layout planning, and text editing, improving OCR-measured text accuracy on MARIO-Eval while keeping image quality.","lead":"Type-R is a post-processing system that finds and fixes misspelled or missing words in images made by text-to-image generators like Stable Diffusion and Flux. It uses OCR to spot the errors and then erases, repositions, and rewrites the text boxes automatically, so designers can get accurate text without retraining the generator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim rests on an unnamed OCR evaluator while the pipeline is tuned to the same OCR family, so 'highest text-rendering accuracy' is not independently established.","rationale":"The reader's conditional verdict is well supported, and my stress-test converges on the same load-bearing point: the pipeline's decisions and the headline accuracy metric both depend on OCR, and the evaluation OCR is never named in the main text. This is not a claim of dishonesty; it is a straightforward identifiability problem. Type-R uses OCR as ground truth in Sec. 3.1, stops correcting when its own OCR reads a word as correct in Alg. 1, and Sec. 4.7 shows that OCR accuracy is highest when the text editor and the recognizer are compatible (Paddle). If the evaluator is the same recognizer or a close relative, the reported OCR gains could overstate human-readable text accuracy. The user study validates graphic quality and prompt matching, but not word-level typo correctness, so it does not close this gap. I also noticed that Eq. (1) writes 'max' where a Levenshtein-distance matching objective would normally be 'min'; this is likely a typographical slip because the reported experiments would not succeed under the literal max objective, but it should be corrected. Since the reader has already marked the paper CONDITIONAL and the unresolved issue is exactly the reason for that condition, I would keep the verdict unchanged rather than escalate to rejection.","tokens_in":16944,"tokens_out":7730,"duration_ms":71540,"concrete_test":"Name the evaluation OCR in Sec. 4.2 and release the evaluation script. Then rerun the Table 1 OCR comparison on the full 5,414 MARIO-Eval prompts with two independent recognizers that are not used anywhere in Type-R (e.g., Google Cloud Vision and TrOCR without Paddle), plus a human word-level accuracy read on a 200-image stratified sample. If Type-R w/ Flux no longer ranks first on either independent OCR, or if its Paddle-based OCR score exceeds the independent OCR scores by more than 5 points, the headline claim must be restated as OCR-specific rather than general text-rendering accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim is that Type-R gives the highest text-rendering accuracy while maintaining image quality. The evidence for 'text-rendering accuracy' is entirely OCR-based, and the same kind of OCR is the internal ground truth that decides what gets erased, inserted, and when correction stops (Sec. 3.1, Alg. 1). The main text never names the evaluation OCR; only Fig. 8 distinguishes an internal OCR from an evaluation OCR, and Sec. 4.7 explicitly shows the pipeline is tuned to be compatible with Paddle, the recognizer used internally. This leaves open two concrete failure modes: (i) if the evaluation OCR is Paddle or a close relative, the reported OCR gains partly measure the pipeline's ability to produce Paddle-readable text rather than human-legible text; (ii) even with a different evaluator, a single OCR is a weak proxy for human reading, and the user study (Table 2) asks about graphic quality and prompt matching, not about word-level legibility. The paper's own Appendix F calls the internal counts 'pseudo-errors,' conceding that the ground truth for detection is OCR-dependent. Thus the central accuracy claim is not independently established until the evaluator is specified and checked against a recognizer and a human read that are not used anywhere in Type-R.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Type-R is a training-free post-processing pipeline that detects and repairs typographic errors in images generated by arbitrary text-to-image models. It operates in four stages: word-level error detection via scene-text detection/recognition matched to prompt words by optimal transport (Eq. 1), erasure of unintended words with inpainting, layout regeneration of missing words using GPT-4o, and iterative typo correction via a text-editing model (Algorithm 1). The paper evaluates Type-R combined with SD3, Flux, SD3.5, and DALL-E 3 on the MARIO-Eval benchmark, reporting OCR accuracy, GPT-based graphic quality/content alignment scores, FID, CLIP score, and a human user study. The central claim is that Type-R with Flux achieves the best trade-off between text rendering accuracy and image quality.","tokens_in":17199,"tokens_out":2250,"duration_ms":21595,"significance":"If the claims hold, Type-R is a practically useful contribution: it offers a plug-and-play improvement over existing text-to-image generators without fine-tuning, and the ablation studies (Tables 3, 4, 5, 9, 11, 12) give a fairly complete picture of how each module and each backend choice matters. The paper is also transparent about limitations, including OCR failures and the need for better backend models. The main weak point is that the headline 'text rendering accuracy' rests entirely on an OCR evaluator that is never named in the main text, while the pipeline itself is driven by an OCR model from the same family (Paddle). This creates a real risk that the reported gains partly measure Paddle-readability rather than human-legible correctness. The human study (Table 2) asks about graphic quality and content matching, not word-level legibility, so it does not resolve this concern.","major_comments":[{"comment":"The evaluation OCR model used to compute the headline OCR accuracy in Tables 1, 3, 4, 5, 9, 11, and 12 is never specified in the main text. The Fig. 8 caption only says that OCR* is computed by the OCR models in Type-R and is distinct from the OCR model for evaluation, but the main text does not name the evaluator. Given that the pipeline internally uses Paddle (Sec. 4.3) and Sec. 4.7 (Table 4) explicitly shows that AnyText's compatibility with Paddle affects the measured OCR accuracy, the reported gains could partly reflect the pipeline's ability to produce Paddle-readable text rather than generally legible text. The authors should specify the exact evaluation OCR model and rerun the main comparisons with at least one recognizer that is not used anywhere in Type-R, plus a human word-level legibility check.","section":"§4.2 and Fig. 8 caption"},{"comment":"The pipeline treats OCR output as ground truth for what text is actually in the image. In Eq. (1) and Sec. 3.1, detected words are matched to prompt words using Levenshtein distance; in Algorithm 1, the same OCR readings decide when to stop retyping. Appendix F explicitly calls the internal error counts 'pseudo-errors.' This makes the method vulnerable to OCR misreads: a correct word misread as wrong can be erased or retyped, and a wrong word misread as correct will be left in place. The paper acknowledges in the Limitation section that OCR models sometimes fail, but it does not quantify how often this happens in the benchmark or how it affects final image correctness. A sensitivity analysis with a second, independent recognizer would help establish that the correction decisions are not systematically biased by the internal OCR's blind spots.","section":"§3.1, Algorithm 1, and Appendix F"},{"comment":"The claim of 'highest text rendering accuracy' is supported only by OCR accuracy. The GPT-based graphic quality scores show substantial disagreement with human ratings (Supplementary A reports, e.g., 56% of GPT scores above 8.0 versus 3% for humans for Type-R w/ Flux), and the user study in Table 2 asks participants to rate graphic design quality and text-image matching, not word-level legibility. Thus the evaluation does not directly measure whether a human reader can correctly read the retouched words. The authors should add a small human reading test (e.g., forced-choice word identification) on a sample of images, or at least report per-word human legibility agreement for the main comparison.","section":"Supplementary A and Table 2"},{"comment":"The main comparison reports single OCR accuracy numbers and GPT scores without error bars, confidence intervals, or per-prompt statistical tests. The user study provides p-values for the matching score only, not for OCR accuracy or graphic quality. Because the benchmark has 5,414 prompts, bootstrapping or per-prompt mean differences with paired tests would be straightforward to add and would make the headline 'significantly outperforms' claim more robust.","section":"Table 1 and Sec. 4.5"}],"minor_comments":[{"comment":"The detector is called 'Deepsolo' in the implementation details but 'DeepSolo' in Table 4 and the references; please use consistent capitalization.","section":"§4.3"},{"comment":"The symbol OCR* is used in the caption and main text without a formal definition; define it where it first appears.","section":"Fig. 8"},{"comment":"The citation '[15?]' contains a stray question mark; replace it with the proper reference.","section":"Appendix B"},{"comment":"The phrase 'compute the similarity between each image and its prompt-based on the CLIP' should read 'based on CLIP' or 'using CLIP.'","section":"§4.2"},{"comment":"The grid search selects θ on a validation set, but the main results then use the selected value on the benchmark; a sentence noting that this selection did not use the benchmark would help clarify the protocol.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the engineering is solid, but the evaluation's dependence on an unnamed OCR model creates a load-bearing weakness. I would advise the editor that the revision must identify the evaluation OCR, add an independent recognizer and a human legibility check, and report variability in the main metrics. If the authors can do that, the paper would be a solid accept; without it, the headline claim is not independently established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Type-R is a well-built systems paper. The contribution is a full post-processing pipeline—prompt-word matching with padding tokens, text erasing, VLM layout regeneration, and iterative typo correction—that works with any base text-to-image model. That combination is genuinely new relative to TextDiffuser and Li et al., and the experiments are solid: ablations for each module, OCR/recognizer swaps, eraser swaps, multiple T2I backends, and a user study. The paper also does the honest thing of reporting FID and CLIP and explaining why they don't favor Type-R, and appendix A gives concrete examples of GPT-human disagreement. This is a real engineering contribution, not a toy.\n\nThe soft spot is exactly where the stress-test points: the headline OCR numbers are computed by an evaluator that the main text never names. Figure 8's caption says the evaluation OCR is distinct from the internal OCR, so the authors are aware of the issue, but they don't tell us which recognizer it is or show that it agrees with human reading at the word level. Given Section 4.7 shows the whole pipeline is effectively tuned to be Paddle-compatible (AnyText is trained with Paddle loss, and Paddle gives the highest OCR), an unnamed evaluator leaves the central accuracy claim underdetermined. The user study asks about graphic quality and prompt matching, not word-level legibility, so it doesn't validate the OCR proxy either. This is a moderate issue, not a fatal one: the qualitative examples show real corrections, and the pipeline improves even the alternative recognizers tested in Table 4. Still, the authors should name the evaluator, add a second independent OCR (and ideally a human word-reading check), and give error bars for the main table.\n\nMinor: no code URL appears in the text despite a footnote claiming code availability, and the comparison to Refining T2I is skipped because of the different setting—reasonable, but worth one sentence in the final paper.\n\nWho is this for? Anyone working on text-to-image post-processing or graphic-design generation. A careful referee can get the paper into good shape. I'd send it to review.","headline":"A solid post-processing pipeline for typo correction that deserves refereeing, but the unnamed OCR evaluator leaves the headline accuracy claim underdetermined.","tokens_in":17755,"tokens_out":2463,"would_cite":true,"duration_ms":22825,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a post-processing pipeline called Type-R can locate, erase, and rewrite misspelled words in images from any text-to-image model, improving text accuracy without sacrificing image quality.","keywords":["text-to-image generation","typography correction","post-processing","OCR","text rendering","image inpainting","text editing","layout regeneration"],"falsifier":"Take a generated image in which the requested word is rendered correctly but in a font the OCR model reads incorrectly; if Type-R then erases or rewrites that correct word, it shows the method's accuracy is bounded by its internal recognizer. A cleaner test is to evaluate Type-R's output with an independent OCR model from a different family; if the reported accuracy gain shrinks or disappears, the gain is partly tuning the image to the pipeline's own recognizer.","tokens_in":16765,"feed_emoji":"🔤","tokens_out":6358,"duration_ms":55104,"temperature":0.7,"pith_summary":"Type-R is a post-processing approach to the persistent problem of misspelled or missing words in text-to-image generation. Instead of retraining or conditioning the image generator, it runs four automatic stages on the already generated image: detect words with OCR, erase unintended text, plan boxes for missing words, and iteratively rewrite misspelled words. The paper argues this corrects typos while preserving visual design, and that pairing Type-R with current generators such as Stable Diffusion or Flux delivers the best balance of text accuracy and image quality among the methods tested. A sympathetic reader would take away that typo-free design images can be obtained by retouching rather than by building better text renderers.","feed_headline":"Retouching pipeline fixes typos in AI-generated images","feed_subtitle":"Type-R erases bad text, replans missing words, and rewrites misspellings while preserving image quality.","key_machinery":"The mechanism that carries the argument is a four-stage pipeline whose pivot is the word-matching step. Detection and recognition turn the image into a set of OCR words; an optimal matching with Levenshtein distance pairs them against the prompt words, with padding tokens for missing or extra words. This match decides which regions to erase, where to insert missing words, and which words to send to a text editor. The text editor is applied iteratively, and only successfully corrected regions are composited back, so the image changes as little as possible.","core_discovery":"The paper's central claim is that typographic errors in text-to-image outputs are better fixed after generation than by changing the generator. Type-R formulates the task as a matching problem between the words requested in the prompt and the words an OCR model reads in the image: an optimal assignment with Levenshtein edit distances decides which rendered words are correct, which are unwanted, which are missing, and which are misspelled. It then erases surplus text, asks a vision-language model to place missing words, and repeatedly applies a text editor until the OCR no longer reports an error. The reported experiments show that Type-R paired with the Flux generator reaches the highest OCR accuracy among the compared methods while keeping graphic design quality near that of the raw generator, and similar gains appear when it is attached to other base generators.","pith_inferences":["If OCR accuracy is the bottleneck, Type-R effectively converts 'render the text correctly' into 'render the text legibly to a chosen recognizer'; the paper's finding that the text editor works best with the recognition model it was trained with suggests the reported metric can be sensitive to model pairing.","The same retouching pipeline could be repurposed as an automatic data-cleaning tool: instead of discarding images with typos from training sets, one could retouch them and keep the image, which may be cheaper than regenerating.","Because Type-R depends on OCR, a text editor, and a vision-language model, its ceiling is set by those components, so improvements in any one should transfer directly to the pipeline without changing its structure.","The pipeline decomposes text by spaces, so it needs explicit handling for scripts without spaces or for reading-order-sensitive captions before it generalizes beyond the English-like cases tested; the paper notes that reading order is not guaranteed."],"forward_implications":["Any current or future text-to-image generator can be paired with Type-R without fine-tuning, since the pipeline treats the generator as a black box.","The main comparison reports that Type-R with Flux beats text-focused generators TextDiffuser and TextDiffuser-2 on OCR accuracy and on the quality-accuracy trade-off, implying dedicated layout-conditioned generators are not required for typo-free output.","The pipeline's gains stack: erasing surplus words, regenerating missing layouts, and correcting typos each add accuracy, with layout correction contributing most for Flux.","Iterative typo correction concentrates its benefit early, reaching roughly 80 percent of total improvement within four attempts across base models."],"supporting_citations":[{"why":"Serves as the primary base text-to-image generator whose combination with Type-R achieves the best reported accuracy-quality trade-off.","marker":"[3]"},{"why":"Supplies the vision-language model used for layout regeneration and for the GPT-based quality evaluation.","marker":"[4]"},{"why":"Provides the scene text recognition model that reads detected words and drives the error-detection matching.","marker":"[5]"},{"why":"Provides the benchmark dataset and one of the main text-focused baselines that Type-R is compared against.","marker":"[9]"},{"why":"Provides the state-of-the-art text-focused baseline and the evaluation protocol that the paper adapts for rating-based scoring.","marker":"[10]"},{"why":"Serves as the second base text-to-image generator in the main comparison, showing Type-R works across generators.","marker":"[11]"},{"why":"Supplies the inpainting model used to erase unintended text regions before layout regeneration.","marker":"[37]"},{"why":"Supplies the text editing model used for iterative typo correction, the final stage of the pipeline.","marker":"[39]"},{"why":"Supplies the scene text detection model that localizes words in the generated image, with the best reported detection performance in the ablation.","marker":"[48]"}],"fun_headline_variants":["Fix AI image typos via post-hoc retouching","Post-generation pipeline erases and rewrites typos","Type-R retouches typos post-hoc without hurting image quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline treats the OCR model's reading as the truth about which words in the image are wrong, so a misread can make it erase or rewrite a correct word or leave a real typo untouched.","fun_headline_variants_meta":{"raw":{"variants":["Fix AI image typos via post-hoc retouching","Post-generation pipeline erases and rewrites typos","Type-R retouches typos post-hoc without hurting image quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2838,"prompt_tokens":826,"completion_tokens":2012,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":1958}},"tokens_in":442,"tokens_out":2012,"duration_ms":12410,"temperature":1.0,"reasoning_tokens":1958,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:27:15.908911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a generated image in which the requested word is rendered correctly but in a font the OCR model reads incorrectly; if Type-R then erases or rewrites that correct word, it shows the method's accuracy is bounded by its internal recognizer. A cleaner test is to evaluate Type-R's output with an independent OCR model from a different family; if the reported accuracy gain shrinks or disappears, the gain is partly tuning the image to the pipeline's own recognizer.","supporting_citations":[{"cited_title":"Textdiffuser: Diffusion models as text painters","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark dataset and one of the main text-focused baselines that Type-R is compared against."},{"cited_title":"https://blackforestlabs.ai/ announcing - black - forest - labs/","cited_arxiv_id":null,"evidence_quote":"Serves as the primary base text-to-image generator whose combination with Type-R achieves the best reported accuracy-quality trade-off."},{"cited_title":"https://openai.com/index/gpt- 4o- system-card/","cited_arxiv_id":null,"evidence_quote":"Supplies the vision-language model used for layout regeneration and for the GPT-based quality evaluation."},{"cited_title":"https : / / github","cited_arxiv_id":null,"evidence_quote":"Provides the scene text recognition model that reads detected words and drives the error-detection matching."},{"cited_title":"Textdiffuser-2: Unleashing the power of language models for text rendering","cited_arxiv_id":null,"evidence_quote":"Provides the state-of-the-art text-focused baseline and the evaluation protocol that the paper adapts for rating-based scoring."},{"cited_title":"Scaling recti- fied flow transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"Serves as the second base text-to-image generator in the main comparison, showing Type-R works across generators."},{"cited_title":"Resolution-robust large mask inpainting with fourier convolutions","cited_arxiv_id":null,"evidence_quote":"Supplies the inpainting model used to erase unintended text regions before layout regeneration."},{"cited_title":"Anytext: Multilingual visual text gener- ation and editing","cited_arxiv_id":null,"evidence_quote":"Supplies the text editing model used for iterative typo correction, the final stage of the pipeline."},{"cited_title":"DeepSolo: Let transformer decoder with explicit points solo for text spot- ting","cited_arxiv_id":null,"evidence_quote":"Supplies the scene text detection model that localizes words in the generated image, with the best reported detection performance in the ablation."}],"review_version":1}