{"id":"87ec8614-5e61-4049-b893-7536815f70ef","arxiv_id":"2605.19320","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"TextAlign uses a hierarchical VLM reward for preference alignment to boost text accuracy in generative models like FLUX.1-dev.","lead":"TextAlign is a framework that improves text rendering in text-to-image models by using a hierarchical vision-language model reward to guide preference alignment without changing the model architecture. Smart generalists might read it because it suggests a way to fix a common flaw in AI image generation using existing models rather than redesigning them.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"VLM binary defect judgments lack reported validation against human labels before conversion to scalar reward","rationale":"The reader's weakest assumption directly identifies the same unvalidated conversion step. Because the query provides only the abstract and states full text is elsewhere, no additional internal inconsistency or stronger evidence can be checked; the concern therefore stands as load-bearing and the low-confidence UNVERDICTED verdict is appropriate.","tokens_in":1730,"tokens_out":297,"duration_ms":14478,"concrete_test":"Sample 300 images from the FLUX.1-dev evaluation set; obtain independent human binary defect labels at the three hierarchy levels; compute Cohen's kappa between VLM and human judgments; if any level falls below 0.65, re-run the GRPO/DPO stage with human-derived preferences and measure OCR accuracy delta.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that binary VLM outputs at global/word/glyph levels can be turned into a scalar preference signal that drives effective GRPO/DPO. The abstract states this conversion occurs but supplies no quantitative evidence (e.g., inter-annotator agreement, precision/recall of VLM vs. human glyph-level labels, or ablation of the aggregation rule). If VLM judgments contain systematic biases (common in current VLMs on fine-grained text), the resulting preference pairs will be noisy or misaligned, undermining the claim that reward design alone suffices.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes TextAlign, a non-invasive post-training framework for improving text rendering in text-to-image models. It introduces a hierarchical VLM-based reward that decomposes errors into global, word, and glyph levels, converts binary defect judgments into scalar preferences, and applies these to GRPO and DPO optimization on models such as FLUX.1-dev and Z-Image-Turbo. Experiments claim consistent OCR accuracy gains without degrading general generation quality, outperforming baselines including SD3.5, Qwen-Image, AnyText, and TextDiffuser, positioning reward design as a scalable alternative to architectural changes.","tokens_in":1822,"tokens_out":586,"duration_ms":21781,"significance":"If the results hold after addressing validation gaps, the work would be significant as it provides empirical evidence that preference alignment via carefully designed hierarchical rewards can improve a persistent weakness in foundation models without modifying their architecture. This could enable broader deployment across existing generators and contribute to the growing literature on reward modeling for generative tasks.","major_comments":[{"comment":"Abstract and Experiments section: The central claim of consistent gains in OCR-based text accuracy relies on the hierarchical VLM reward producing reliable preference signals, yet no quantitative validation (e.g., precision/recall of VLM glyph-level judgments vs. human labels or inter-annotator agreement) is reported. This directly affects whether the scalar preference signal supports effective GRPO/DPO as asserted.","section":"Abstract"},{"comment":"Reward construction (hierarchical VLM section): The conversion rule from three-level binary defect judgments to scalar preference is load-bearing for the optimization claims, but the manuscript supplies no ablation of the aggregation function or evidence that VLM judgments avoid systematic biases on fine-grained text, which is a known issue with current VLMs.","section":"Reward Model"},{"comment":"Experiments and baselines: Comparisons to AnyText and TextDiffuser require explicit details on implementation, prompt sets, and statistical significance testing of the reported OCR gains; without these, it is unclear whether the improvements are attributable to the proposed reward or to differences in evaluation protocol.","section":"Experiments"}],"minor_comments":[{"comment":"Clarify the exact mathematical form of the scalar reward aggregation from the three binary levels in the methods section.","section":"Methods"},{"comment":"Add error bars or multiple-run statistics to the OCR accuracy tables to support the 'consistent gains' claim.","section":"Results"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of a computer vision journal focused on generative models. The absence of human validation for the core reward component is the primary concern; if addressed, the contribution on scalable alignment would be stronger."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback highlighting the need for stronger validation of the reward model, ablations on aggregation, and clearer experimental protocols. We address each major comment below and will incorporate revisions to improve the manuscript.","responses":[{"response":"We agree that explicit quantitative validation of the VLM judgments is a gap in the current manuscript. While end-to-end OCR gains provide supporting evidence, we will add a new validation subsection in the revised Experiments section. This will include a human study on a sampled set of generations, reporting precision/recall of VLM glyph-level judgments against human labels and inter-annotator agreement statistics.","revision_made":"yes","referee_comment":"[Abstract] Abstract and Experiments section: The central claim of consistent gains in OCR-based text accuracy relies on the hierarchical VLM reward producing reliable preference signals, yet no quantitative validation (e.g., precision/recall of VLM glyph-level judgments vs. human labels or inter-annotator agreement) is reported. This directly affects whether the scalar preference signal supports effective GRPO/DPO as asserted."},{"response":"We will add an ablation study in the revised Reward Model section comparing alternative aggregation functions for converting the three-level binary judgments into scalar preferences. We will also include qualitative and quantitative analysis of potential VLM biases on fine-grained text, with examples contrasting VLM outputs against human judgments to address known limitations.","revision_made":"yes","referee_comment":"[Reward Model] Reward construction (hierarchical VLM section): The conversion rule from three-level binary defect judgments to scalar preference is load-bearing for the optimization claims, but the manuscript supplies no ablation of the aggregation function or evidence that VLM judgments avoid systematic biases on fine-grained text, which is a known issue with current VLMs."},{"response":"We will expand the Experiments section with full implementation details for AnyText and TextDiffuser (including code references or hyperparameters), the exact prompt sets and evaluation protocols, and statistical significance testing (e.g., paired t-tests) on the OCR accuracy improvements to confirm the gains are robust and not due to protocol differences.","revision_made":"yes","referee_comment":"[Experiments] Experiments and baselines: Comparisons to AnyText and TextDiffuser require explicit details on implementation, prompt sets, and statistical significance testing of the reported OCR gains; without these, it is unclear whether the improvements are attributable to the proposed reward or to differences in evaluation protocol."}],"tokens_in":1406,"tokens_out":536,"duration_ms":27100,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that this paper treats text rendering as a preference-alignment task rather than an architecture problem. It introduces a hierarchical VLM reward that scores at global, word, and glyph levels, turns the binary defect calls into a scalar signal, and runs GRPO or DPO on FLUX.1-dev and Z-Image-Turbo. The reported outcome is higher OCR accuracy with no drop in general image quality, and it beats several text-specific baselines.\n\nWhat the work does cleanly is keep the generator unchanged. That makes the method portable across foundation models, which is the practical advantage over adding encoder modules or special text layers. The three-level decomposition is a sensible way to handle the different granularities of text errors.\n\nThe soft spot is exactly where the stress-test flagged it. The whole pipeline rests on the VLM producing reliable binary judgments that can be aggregated into a preference signal. The abstract gives no numbers on how those VLM outputs align with human glyph-level labels, no inter-annotator agreement, and no ablation of the aggregation rule. If the VLM is systematically off on fine text details, the preference pairs fed to GRPO and DPO will be noisy, and the claimed gains become harder to trust. The experiments compare against SD3.5, Qwen-Image, AnyText, and TextDiffuser, but without that validation step the central claim that reward design alone is enough stays under-supported.\n\nThis is for people who work on post-training fixes for text-to-image models and want something that does not require retraining the base weights. A reader who cares about alignment techniques or practical deployment would find the reward decomposition worth looking at. The paper is coherent on its own terms and engages the literature, so it deserves a serious referee even though the VLM validation needs to be addressed.","headline":"TextAlign shows a post-training hierarchical VLM reward can lift text rendering on FLUX and Z-Image-Turbo, but the conversion from VLM binary labels to scalar preference lacks any reported human validation.","tokens_in":2312,"tokens_out":456,"would_cite":false,"duration_ms":18520,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A hierarchical VLM reward converts binary defect judgments into scalar signals that align text-to-image models for accurate rendering via GRPO or DPO.","keywords":["text rendering","preference alignment","hierarchical reward","text-to-image","VLM","GRPO","DPO","OCR accuracy"],"falsifier":"Training FLUX.1-dev with the hierarchical reward and then measuring OCR accuracy on a fixed prompt set yields no gain over the unaligned baseline, or general image quality drops.","tokens_in":2631,"feed_emoji":"🔤","tokens_out":591,"duration_ms":20729,"temperature":0.7,"pith_summary":"The paper frames text rendering as a post-training preference-alignment task instead of an architecture redesign problem. It builds TextAlign around a vision-language model that checks errors at global, word, and glyph levels and turns those checks into a single preference score. The score then drives Group Relative Policy Optimization or Direct Preference Optimization on the unchanged generator. Tests on FLUX.1-dev and Z-Image-Turbo produce higher OCR accuracy while preserving general image quality. The results position reward design as a model-agnostic way to fix text rendering across foundation models.","feed_headline":"Hierarchical VLM rewards fix text errors in image generators","feed_subtitle":"A three-level reward turns defect checks into GRPO and DPO signals that raise OCR accuracy without model changes or quality loss.","key_machinery":"Hierarchical VLM-based reward that decomposes rendering errors into global, word, and glyph levels then converts the binary judgments into a scalar preference signal for GRPO or DPO.","core_discovery":"TextAlign keeps the generator architecture fixed and instead supplies a hierarchical VLM-based reward that decomposes text-rendering quality into global, word, and glyph levels, converts the binary defect judgments into a scalar preference signal, and feeds that signal to either GRPO or DPO optimization.","pith_inferences":["The same three-level judgment structure could be reused for other fine-grained visual control tasks.","If the reward proves robust, specialized text encoders may become unnecessary for many applications.","Extending the hierarchy to video or layout-constrained generation is a direct next test."],"forward_implications":["OCR text accuracy rises on both FLUX.1-dev and Z-Image-Turbo.","General generation quality stays comparable to the original models.","The same reward works with both GRPO and DPO.","The method outperforms several strong baselines including SD3.5, Qwen-Image, AnyText, and TextDiffuser."],"fun_headline_variants":["TextAlign supplies hierarchical VLM rewards for text rendering alignment","Three-level rewards turn binary judgments into scalar preference signals","VLM decomposes rendering errors across global word and glyph levels","GRPO and DPO optimize text rendering via hierarchical vision language rewards","TextAlign keeps architecture unchanged using VLM-based preference rewards"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The VLM's binary defect judgments at the three levels can be turned into a reliable scalar preference signal that supports effective optimization.","fun_headline_variants_meta":{"raw":{"variants":["TextAlign supplies hierarchical VLM rewards for text rendering alignment","Three-level rewards turn binary judgments into scalar preference signals","VLM decomposes rendering errors across global word and glyph levels","GRPO and DPO optimize text rendering via hierarchical vision language rewards","TextAlign keeps architecture unchanged using VLM-based preference rewards"]},"model":"grok-4.3","cost_usd":0.005572,"raw_usage":{"total_tokens":2645,"prompt_tokens":617,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":55724500,"prompt_tokens_details":{"text_tokens":617,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1947,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":617,"tokens_out":81,"duration_ms":15691,"temperature":1.0,"reasoning_tokens":1947,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T18:48:49.655376+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Training FLUX.1-dev with the hierarchical reward and then measuring OCR accuracy on a fixed prompt set yields no gain over the unaligned baseline, or general image quality drops.","supporting_citations":[],"review_version":2}