{"id":"4a433014-5815-4472-93f5-2065e1817f0b","arxiv_id":"2601.02211","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Scaling text-condition hidden states by 1.5 in a small, attribute-specific set of MMDiT blocks improves text-image alignment, editing, and speed on SD3.5, FLUX, and Qwen Image with no training.","lead":"MMDiT image models like SD3.5, FLUX, and Qwen-Image can follow prompts more faithfully when the text signal is amplified inside a few carefully chosen internal blocks — no retraining required. The same block-level control also improves text-guided editing and modestly speeds up inference.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Probe-based block selection is not shown to be stable or to outperform random block amplification; the central 'selective block' claim may rest on a small, unvalidated selection criterion.","rationale":"The reader identified the probe set reliability as the weakest assumption. I agree that the probe is under-validated, but I extend the concern to a more specific threat: the method's improvement may not be attributable to the selected blocks at all, since Table 6 shows random 5-block enhancement also lifts performance. The permutation test would directly settle whether the selection criteria add value. Additionally, the abstract's T2I-CompBench aggregate appears inconsistent with Table 2, which further undermines confidence in the reported gains. These issues do not disprove the central claim, but they require additional evidence to be convincing, so a conditional verdict remains appropriate.","tokens_in":14143,"tokens_out":4234,"duration_ms":47217,"concrete_test":"Run a permutation test: evaluate GenEval with the authors' selected block set and with 20 random block sets of the same size, all with λ=1.5, and report mean and std of overall GenEval. If the selected set's improvement over baseline is within 2 std of the random sets' distribution, the block selection is not the cause of the gain.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The block sets in Table 1 are selected from the Sec. 2 probe (333 prompts, 5 seeds) using a loosely defined criterion ('response magnitudes significantly above baseline') plus heuristics (uniform depth distribution, avoid adjacent layers). No stability analysis is given: does the selection change if the probe prompts are subsampled or the VLM judge is changed? The probe is also the same type of manipulation (enhancing) as the final method, so it is a form of in-sample tuning; the fact that the method helps on GenEval/T2I-CompBench++ could be due to the λ scaling providing a general boost (random blocks also improve per Table 6) rather than to the specific block targeting. The abstract's T2I-CompBench aggregate (56.92→63.00) is not reproducible from Table 2's categories (the simple average of the eight categories is ~49.8% for baseline), suggesting possible selective reporting or a different aggregation that is not described. Without error bars on the benchmark tables and without distributional comparison between probe and benchmark prompts, the central claim that 'selective block enhancement' is the active ingredient is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes block-wise behavior of multimodal diffusion transformers (MMDiT) by removing, disabling, and enhancing textual hidden states at individual blocks across SD3.5-Large, FLUX.1-Dev, and Qwen Image. Based on a 333-prompt probe with Qwen2.5-VL-72B and CountGD evaluation, it identifies attribute-specific block sets, then proposes TexTailor, a training-free method that multiplies text hidden states by lambda=1.5 in selected blocks during inference. The method is applied to text-to-image generation (reported gains on GenEval and T2I-CompBench++), instruction-based editing (CLIPtxt gain and human-preference majority), and inference acceleration by skipping blocks 20-40 in FLUX. The paper claims consistent gains across all three models while preserving synthesis quality.","tokens_in":14331,"tokens_out":5699,"duration_ms":61484,"significance":"If the reported effects are robust, the paper makes a useful contribution: it provides a block-wise interpretability map of MMDiT models and a simple, training-free inference-time intervention that improves prompt adherence across multiple architectures and public benchmarks. The paper ships a clear pipeline, includes ablations against random block selection, and evaluates on two standard public benchmarks, which are strengths. However, the central quantitative claims currently lack variance information, the block-selection protocol is under-specified, and the headline T2I-CompBench++ number is not reconstructable from the main table. These issues are load-bearing because the method's value rests on the specificity of the selected blocks and the reliability of the benchmark gains.","major_comments":[{"comment":"The abstract's headline gain on T2I-CompBench++ (56.92% to 63.00% for SD3.5) is not reproducible from Table 2. The simple average of the eight listed categories for SD3.5 is approximately 49.8% (baseline) and 55.1% with Ours. No aggregate row is reported and no weighting or subset formula is given. Please state the exact aggregation rule, include the aggregate in the table, and reconcile the numbers; as written, the central quantitative claim cannot be audited.","section":"Abstract; Table 2"},{"comment":"Alignment scores and editing metrics are reported as single numbers without error bars or seed counts. This is especially problematic in Table 6, where the random-5-block ablation is close to Ours on several categories (Color 0.7624 vs 0.7804; Shape 0.5072 vs 0.5482; 2D Spatial 0.3119 vs 0.3280). Without repeated-seed variance and multiple random block draws, the claim that the selected blocks are superior to generic amplification is not established. Please report mean±std over at least five seeds and the number of random draws.","section":"Tables 2, 3, 4, 6"},{"comment":"The pivotal-block selection criterion is under-specified: 'response magnitudes significantly above baseline' is not defined (no statistic, threshold, or multiple-testing correction), and the subsequent heuristics (uniform depth, avoid adjacent layers) are informal. No stability analysis is shown for subsampling the 333 probe prompts or changing the VLM judge, and no distributional comparison is made between the probe prompts and GenEval/T2I-CompBench++. Because the selection uses the same enhancement operation as the final method, the pipeline is an in-sample tuning loop. Section 6's limitation that the method 'depends on automatic block-wise analysis' makes this the central risk. Please add a precise selection rule and validation of its stability.","section":"Section 4.1; Section 2.2; Section 6"},{"comment":"The enhancement scale lambda=1.5 is chosen from the sweep in Fig. 7(a), which is shown only for FLUX, and is then applied to all models and all tasks. The paper does not state whether this choice was made on a development set before benchmarking or tuned on the same benchmark tables. This introduces a free parameter into the central result. Please report lambda sensitivity per model and clearly separate any development-set choices from the reported benchmark evaluations.","section":"Section 4.1; Fig. 7(a)"},{"comment":"The acceleration results report single timing numbers over 400 prompts without variance or details on warm-up, batching, or GPU state, and the quality metrics show small decreases (e.g., HPSv2 29.0533 to 28.8408 on FLUX) that are described as 'preserved' without statistical support. Please provide repeated-run statistics and a comparison treating the quality deltas as measurements subject to noise.","section":"Section 3.4; Table 5"}],"minor_comments":[{"comment":"The benchmark name appears as 'T2I-Combench' in the abstract and 'T2I-CompBench++' in the main text; use one consistent name.","section":"Abstract"},{"comment":"The manuscript header uses 'TexTailor', while the title in the full text is 'Unraveling MMDiT Blocks...'. Please make this consistent.","section":"Title/header"},{"comment":"The statement that '5 fixed seeds produces consistent and reliable results' is asserted but no variance or agreement statistics are reported. Please add the actual repeated-seed numbers.","section":"Section 2.2"},{"comment":"Equation (4) uses the symbol \\odot but defines a scalar multiplication; align notation with the text.","section":"Eq. (4)"},{"comment":"The '*' marker on Amount is not explained in the caption. Clarify that token-level enhancement from Eq. (5) is applied for amount.","section":"Table 2"},{"comment":"The 'Other Dimensions' block sets are not used in the main results. State their role or remove them from the table.","section":"Table 1"},{"comment":"Human evaluation details are minimal: describe the instruction given to participants, the selection of images, and report inter-annotator agreement.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The mismatch between the abstract's T2I-CompBench++ aggregate and the category scores in Table 2 is the most urgent issue; the editor should require the authors to supply the exact aggregation formula and a reproducible aggregate row. The lack of error bars around the random-block ablation is also a serious concern because the central claim depends on the selected blocks being more than a generic scaling effect. I believe these issues are fixable within a revision, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper is worth a serious referee. The core finding—multiplying text hidden states by λ=1.5 in a small set of MMDiT blocks improves alignment—has consistent directional support across three open models (SD3.5, FLUX, Qwen Image) and two public benchmarks. That is not nothing, especially since it is training-free and the ablation in Table 6 shows dedicated blocks beat random blocks. The attribute-block preference map (early blocks for color/spatial, later blocks for amount) is genuinely new as far as I know, and the acceleration and editing spin-offs are reasonable extra evidence that the authors understand what their intervention does.\n\nNow the soft spots, in proportion. The main benchmark tables do not give error bars or seed counts for the alignment metrics, so the +5-7 point gains could be within run-to-run noise; the GenEval position improvement on SD3.5 (+0.16) is the kind of number that makes you want to see variance. The abstract's T2I-CompBench++ aggregate (56.92→63.00) does not reproduce from Table 2's eight categories—simple averaging gives roughly 49.8 for the baseline. That is either a different aggregation or a selective-reporting problem, and the paper does not say which. The \"without sacrificing synthesis quality\" claim is also not supported by the paper's own Tables 2 and 3: SD3.5 shows consistent small drops in HPSv2 and Aesthetics. The drops are small, but they are in the wrong direction, and the abstract is flatly too strong. Finally, the block-selection criterion is loosely defined (\"response magnitudes significantly above baseline\") and there is no stability analysis showing the selected block sets survive changes to the probe prompts or the VLM judge. That matters because the whole method hangs on the probe; random blocks also help in Table 6, so the specific block targets are not shown to be load-bearing.\n\nThese are all fixable with more transparency, not fatal flaws. The reasoning is clear, the experiments are honest enough in what they show, and the limitation section concedes the dependence on automatic block-wise analysis. I would send this to peer review and ask the authors to add error bars, explain the aggregate metric, tone down the quality claim, and include a probe-stability check. Until then, I would not build on the exact block sets, but the phenomenon itself looks real.","headline":"Serious empirical paper with a plausible core finding, but the headline numbers are over-sold and the block-selection procedure needs more validation before the central claim is fully established.","tokens_in":14934,"tokens_out":2115,"would_cite":false,"duration_ms":22707,"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 argues that multimodal diffusion transformers assign different semantic attributes to different internal blocks, and that amplifying the text hidden states in just a few of those blocks—by a constant factor, with no retraining—su","keywords":["multimodal diffusion transformer","text-to-image generation","block-wise analysis","inference-time guidance","training-free enhancement","image editing","inference acceleration","text alignment"],"falsifier":"A concrete test: rerun the block-selection procedure using a different judge (e.g., human raters or another vision-language model) on the same 333 prompts, then compare GenEval and T2I-CompBench++ gains using the new block sets. If the 'significantly above baseline' blocks shift substantially, the gains are judge artifacts. More directly, evaluate the paper's fixed block sets on a held-out set of novel compositional prompts that are not in GenEval or T2I-CompBench++; if the enhancement fails to improve alignment there, the map does not generalize beyond the two benchmark distributions.","tokens_in":13938,"feed_emoji":"🖼️","tokens_out":3687,"duration_ms":42773,"temperature":0.7,"pith_summary":"The paper asks whether text-to-image alignment in modern transformer-based diffusion models can be improved without retraining. It first probes how each internal block of three MMDiT models (SD3.5, FLUX, Qwen Image) contributes to attributes like color, spatial relations, and object count, by removing, disabling, and enhancing textual hidden states block by block. It finds that early blocks carry core semantic information, later blocks refine fine details, and that specific blocks are attribute-specialized. Building on that, it shows that multiplying the textual hidden states by 1.5 in a small, carefully selected set of blocks improves alignment scores (GenEval +6.3, T2I-CompBench++ +6.1 on SD3.5) while preserving image quality, and enables more precise editing and faster inference by skipping low-impact blocks. A sympathetic reader would care because this suggests a general, training-free lever for improving text adherence in any MMDiT-based generator.","feed_headline":"Boosting text states in a few blocks adds ~6 points to image-prompt match","feed_subtitle":"Training-free: amplifying textual conditioning in selective MMDiT blocks improves generation, editing, and speed.","key_machinery":"The key machinery is the block-wise probe pipeline and the resulting 'enhance-at-selected-blocks' rule. The probe operates on 333 GPT-5-generated prompts spanning color, spatial, and amount attributes, judged by a vision-language model (Qwen2.5-VL QA) and a counting model (CountGD) across five seeds. For every block, the paper removes, disables, or enhances the textual hidden states and measures attribute accuracy plus perceptual/semantic similarity. The resulting response curves identify which blocks are 'vital' for each attribute, forming small block sets (e.g., {3,9,15,20} for color in SD3.5, {32,37,45,49,54} for amount in FLUX). The enhancement operation itself is a straightforward scali","core_discovery":"The central discovery is that MMDiT blocks are attribute-specialized: early blocks handle color and spatial relationships, while later blocks influence object amount. Amplifying the textual hidden states at these attribute-matched blocks—via the simple operation c_enh = λ·c with λ=1.5, optionally mask-selected per token—improves text alignment across all three tested models without degrading aesthetic quality. The paper reports concrete gains: on SD3.5, GenEval rises from 66.42% to 71.63% and T2I-CompBench++ from 56.92% to 63.00%. The same enhancement scheme, applied to the target prompt during parallel-generation editing, improves instruction-following (CLIPtxt +0.94) while preserving sourc","pith_inferences":["A natural next step is to make block selection per-prompt dynamic, choosing blocks based on the attribute words present in the prompt rather than a fixed global set; the probe pipeline could precompute attribute-block maps and route at inference.","The robustness of the multiplier up to λ=2.0 hints that text conditioning in these blocks is under-saturated, suggesting headroom to pair this with training-time conditioning strengthening—though this is untested.","The depth-attribution view (semantics early, details late) predicts that classifier-free guidance or other conditioning interventions should be concentrated on early blocks; that design change is a concrete, testable consequence beyond what the paper demonstrates.","The same enhancement scheme may transfer to video or 3D MMDiTs where strong text alignment is also hard, but this remains speculative since the paper only tests image models."],"forward_implications":["If the block-attribute map transfers, any MMDiT-based generator could be aligned better by rerunning the probe on its own blocks and scaling text states—no retraining or architecture changes needed.","Attribute-specific editing becomes tractable: color edits only need to amplify early blocks, amount edits late blocks, promising more surgical instruction-following.","Skipping low-impact middle blocks offers a practical speedup (about 14–16% on FLUX) with negligible quality loss, and stacks with existing inference accelerators.","Token-level enhancement means a single prompt can weight specific nouns or adjectives, enabling fine-grained compositional control from the text side alone."],"fun_headline_variants":["Boosting key text blocks lifts image match by 6 points","Tune text states at the right blocks to sharpen image fidelity","No training: amplify select text features for better alignment","Inference-time text tweak: +5–6 points on alignment benchmarks","Targeted text-state boost improves generation, editing, and speed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The block-attribute map inferred from 333 machine-judged prompts is stable and transfers to the benchmark prompts and all three models—if the VLM judge is biased or the probe prompts are unrepresentative, the selected block sets are mistargeted and the reported gains would not replicate outside the paper's testbed.","fun_headline_variants_meta":{"raw":{"variants":["Boosting key text blocks lifts image match by 6 points","Tune text states at the right blocks to sharpen image fidelity","No training: amplify select text features for better alignment","Inference-time text tweak: +5–6 points on alignment benchmarks","Targeted text-state boost improves generation, editing, and speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1179,"prompt_tokens":861,"completion_tokens":318,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":245}},"tokens_in":605,"tokens_out":318,"duration_ms":4273,"temperature":1.0,"reasoning_tokens":245,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:36:01.184497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: rerun the block-selection procedure using a different judge (e.g., human raters or another vision-language model) on the same 333 prompts, then compare GenEval and T2I-CompBench++ gains using the new block sets. If the 'significantly above baseline' blocks shift substantially, the gains are judge artifacts. More directly, evaluate the paper's fixed block sets on a held-out set of novel compositional prompts that are not in GenEval or T2I-CompBench++; if the enhancement fails to improve alignment there, the map does not generalize beyond the two benchmark distributions.","supporting_citations":[],"review_version":1}