{"id":"d29ede41-4ef7-4a7f-8adf-07b2169acc15","arxiv_id":"2608.03046","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CAPE-T2V fine-tunes a prompt enhancer on captioner-generated targets, then uses that same enhancer to write both the video model's fine-tuning captions and the inference-time prompt rewrites, reducing the distribution gap between training and inference conditioning.","lead":"This paper presents CAPE-T2V, a two-step method that makes the text conditioning a video generation model match the text the model sees at inference by using the same prompt-enhancer model on both sides. It reports consistent gains over a schema-matched baseline on StoryEval, VBench-2.0, and T2V-CompBench for both Wan2.2 and LTX-2.3.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The matched control does not isolate two-sided alignment: missing 2x2 cells leave open that gains come from Anchored-PE caption quality rather than from using the deployed PE on both sides.","rationale":"The reader's weakest assumption correctly identifies that the control does not uniquely support the two-sided-alignment interpretation, and the missing 2x2 cells make this precise. The current design holds the inference PE fixed at the Anchored PE and varies only the DiT fine-tuning captions; this cannot distinguish 'the deployed PE wrote the training captions' from 'the Anchored PE happens to produce better captions.' The proposed 2x2 test directly settles that distinction by swapping the inference-time PE and checking whether the advantage follows the matched operator. The MMD2 analysis, while a useful diagnostic, is not decisive because same-model outputs are expected to be closer in embedding space; it supports the existence of a smaller measured gap but not its causal role in video scores. The small magnitude of some gains and the lack of uncertainty quantification reinforce the need for the additional condition, but they do not warrant rejection. The paper is otherwise well structured, with clear matched settings and honest limitations, so the conditional verdict remains appropriate.","tokens_in":26617,"tokens_out":8285,"duration_ms":82485,"concrete_test":"Run the two missing cells of the 2x2 comparison: use the prompted rewriter (Qwen3.5-397B-A17B with the six-part instruction) as the inference-time PE, and evaluate both the Schema-Aligned DiT (trained on prompted-rewriter captions) and the PE-Aligned DiT (trained on Anchored-PE captions) on the same benchmarks and seed grids. If the DiT trained on prompted-rewriter captions wins under the prompted rewriter by a comparable margin, two-sided operator matching is the operative mechanism; if the Anchored-PE-trained DiT still wins, the gains are attributable to caption quality rather than to alignment. Report paired bootstrap confidence intervals over prompts for the matched-versus-cross differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that using the deployed PE to construct DiT fine-tuning captions yields benefits beyond schema matching alone. The evidence compares two caption operators: the Anchored PE (Qwen3.5-9B, SFT) and the prompted rewriter (Qwen3.5-397B-A17B, fixed instruction). These differ in model scale, pretraining, and fine-tuning status, so the control is not matched on operator identity. More importantly, the design is a 2x2 factorial with only two of four cells populated: both evaluated DiTs use the Anchored PE at inference, while only the training captions vary. Without the two cells in which the prompted rewriter is the inference-time PE, the observed gains could arise either from two-sided operator matching or simply because Anchored-PE captions are better conditioning text independent of the inference-time PE. The MMD2 evidence in Section 4.3 is consistent with the weaker explanation that outputs of the same model are closer to each other in an embedding space than outputs of a different model; it does not establish that the video-score gains are caused by distributional alignment. The benchmark deltas are also small (0.30 pp on Wan2.2 T2V-CompBench) and each DiT configuration is fine-tuned once, with no confidence intervals or significance tests for the headline gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CAPE-T2V, a two-step method for aligning the conditioning text distributions in text-to-video DiTs. In Step 1, a Qwen3.5-9B prompt enhancer is fine-tuned to map concise captions, detailed captions, and pseudo user prompts to captioner-generated six-part MiraData-style targets, yielding an 'Anchored PE.' In Step 2, the same per-video dense captions are rewritten either by the Anchored PE (for the PE-Aligned DiT) or by a prompted Qwen3.5-397B-A17B rewriter (for the Schema-Aligned DiT), and each DiT is fine-tuned on its caption set. At inference, both DiTs are conditioned on Anchored-PE rewrites of user prompts. The paper reports that CAPE-T2V outperforms the Schema-Aligned control on StoryEval, VBench-2.0, and T2V-CompBench for both Wan2.2 and LTX-2.3, with gains of 1.6, 1.12, and 0.30 percentage points (Wan2.2) and 1.4, 0.92, and 0.65 points (LTX-2.3). It also reports lower MMD2 between Anchored-PE fine-tuning captions and Anchored-PE inference rewrites than between prompted-rewriter captions and those same rewrites, and it shows that Anchored-PE captions do not simply copy the dense sources.","tokens_in":26908,"tokens_out":11344,"duration_ms":99426,"significance":"The proposed recipe is practically relevant: it requires no access to a deployed generator's native captioning recipe, and it is evaluated under a carefully matched protocol in which videos, dense-caption sources, schema, optimization settings, training budget, and inference-time PE are held fixed. The use of three external benchmarks across two different model families, the inclusion of paired bootstrap/permutation intervals for the MMD2 diagnostics in Appendix E, and the unusually candid Section 5 limitations paragraph are all strengths. If the central claim is accepted, the work provides a simple and transferable way to reduce training–inference conditioning mismatch for T2V models. However, the current experimental design does not uniquely support the 'two-sided alignment' mechanism, and the benchmark gains lack statistical characterization.","major_comments":[{"comment":"The matched comparison is not matched on operator identity. The Anchored PE is Qwen3.5-9B after SFT on about 735K pairs, while the prompted rewriter is Qwen3.5-397B-A17B with a fixed six-part instruction; these differ in scale, base training, fine-tuning status, and prompting. The paper's statement that 'the two paths differ only in which caption operator produces the DiT fine-tuning captions' (Section 3.2) is therefore true only in a narrow sense that bundles several independent attributes into the word 'operator.' The observed gains on StoryEval, VBench-2.0, and T2V-CompBench could equally arise from the Anchored PE producing captions that are better conditioning text on their own, independent of any matching with the inference-time PE. Moreover, the 2x2 factorial design that would separate these explanations—prompted rewriter as inference PE with both caption sets, and Anchored PE as inference PE with both caption sets—is never run; only the two Anchored-PE-inference cells are evaluated. I recommend adding the missing cells or explicitly reframing the contribution as 'captions written by the deployed prompt enhancer improve DiT fine-tuning,' with a discussion of why the residual operator confound does not affect that claim.","section":"Section 3.2 and Table 1"},{"comment":"The MMD2 diagnostic cannot distinguish the paper's alignment mechanism from a same-model proximity effect. The reference distribution P^user_PE consists of Anchored-PE rewrites, so the comparison P^user_PE versus P^dense_PE involves two sets of texts produced by the same fine-tuned model, whereas P^user_PE versus P^dense_rewriter involves texts from different models. Lower MMD2 for the first pair is expected from shared model-specific phrasing and generation habits even if the Anchored-PE captions have no video-grounded advantage. Appendix E.1 correctly cautions that the analysis does not imply that the underlying text distributions are identical or explain the video-score gains, but Section 5 draws the stronger inference that the smaller discrepancy 'reflects the caption operator's realization rather than increased copying.' To make the diagnostic informative, the authors should add the contrast against rewrites produced by the prompted rewriter (the inference-time text under the alternative mechanism), and ideally connect caption-set identity to per-prompt benchmark scores through a randomization or regression analysis.","section":"Section 4.3 and Appendix E.1"},{"comment":"The headline benchmark gains are single-run observations without confidence intervals or significance tests. Each DiT configuration is fine-tuned once, as the paper acknowledges in Section 5, and the evaluation protocol averages nine videos per prompt. The smallest reported gains (0.30 percentage points on Wan2.2 T2V-CompBench and 0.92 points on Wan2.2 VBench-2.0) are small relative to the likely variance from PE sampling seeds, latent-noise seeds, and judge stochasticity; the StoryEval judge is an LLM and the VBench-2.0 Diversity protocol is separate. The claim of 'a consistent advantage' across all six model–benchmark pairs would be much stronger with a bootstrap over prompts for the aggregate scores or with at least one additional fine-tuning seed. Without such analysis, the central quantitative result is a directional pattern rather than an established effect.","section":"Table 1 and Section 5"}],"minor_comments":[{"comment":"The column headings for the StoryEval and VBench-2.0 sub-scores are difficult to parse because the main-text table appears to merge several heading levels; please reformat the table so each numeric column is clearly labeled and verify that the StoryEval Overall column reproduces the gains quoted in Section 4.2.","section":"Table 1"},{"comment":"The metric labeled simply 'BLEU-4' in Table 3 should specify that it is corpus-level BLEU-4 with the SacreBLEU signature, consistent with Appendix Table 9, so that readers do not confuse it with the mean sentence-level BLEU reported in the appendix.","section":"Table 3 and Appendix Table 9"},{"comment":"The caption uses markers such as 'O1', 'D1', 'G1', and 'P1' without defining them; please add a sentence explaining the annotation scheme or remove the markers.","section":"Figure 3"},{"comment":"The notation 'captions→PE' and 'PE→captions' for the coupling direction is introduced without definition; please state explicitly that the arrow indicates which distribution defines the target form for the other.","section":"Section 2"},{"comment":"The sentence 'we report no numbers for published prompt enhancers' is slightly overstated, since the released Official PE baselines are themselves published configurations; consider rephrasing to 'third-party prompt enhancers beyond the released system PEs.'","section":"Section 5"},{"comment":"Please note in the main text that the StoryEval judge is GPT-5.5 rather than the judge used in the official StoryEval protocol, and state how this substitution affects comparability with published StoryEval scores; the appendix already says the judge backend is runtime configuration, but the main text should be equally explicit.","section":"Section 4.1 and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest, clearly written, and offers a practical recipe that could be useful to the T2V community. The main obstacle is that the central mechanistic claim of 'two-sided conditioning alignment' is not uniquely identified by the experiments: the Anchored-PE and prompted-rewriter conditions differ in model scale and fine-tuning, and the missing 2x2 cells leave open the simpler explanation that Anchored-PE captions are simply better conditioning text. The benchmark results also lack any significance testing or repeated runs. I would support publication after the authors either add the missing inference cells, soften the attribution claim, and add at least prompt-level bootstrap intervals, or provide a clear argument for why the operator confound does not affect the headline conclusion. I see no citation or novelty-disclosure problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'd read this if you do T2V conditioning or prompt engineering. CAPE-T2V does something useful: it anchors a small prompt enhancer (Qwen3.5-9B, SFT) to captioner-generated targets under a fixed six-part schema, then uses that same Anchored PE to rewrite the dense captions used to fine-tune the DiT. The comparison against a schema-aligned control is genuinely well-built: same videos, same dense captions, same schema, same optimization and budget; the only intended difference is which operator writes the fine-tuning captions. Across Wan2.2 and LTX-2.3, the PE-captioned DiT beats the control on all three benchmarks (StoryEval +1.6/+1.4, VBench-2.0 +1.12/+0.92, T2V-CompBench +0.30/+0.65). The MMD2 analysis is length-matched and bootstrap CI'd; the surface-overlap control rules out lexical copying. Code and prompts are released, and the limitations section is unusually plain-spoken.\n\nThe soft spot is attribution, not execution. The control uses a prompted Qwen3.5-397B-A17B as rewriter; the Anchored PE is a 9B SFT model, so the two caption sets differ in operator identity plus scale and fine-tuning. The design fills only two of the four 2x2 cells (both evaluated DiTs use the Anchored PE at inference). Without the cells where the prompted rewriter is the inference PE, you cannot distinguish 'using the deployed PE on both sides' from 'Anchored PE captions are better training text regardless of the inference rewriter.' The MMD2 diagnostic is particularly ambiguous: same-model outputs are naturally closer in embedding space. The authors are careful in the appendix — they call it a diagnostic and say it doesn't explain video-score gains — but the abstract and title still hang on two-sided alignment, and the benchmark deltas are small, with no CIs or significance tests on the headline numbers (each DiT is fine-tuned once; some per-dimension scores actually drop).\n\nThat said, the recipe stands whether or not the mechanism is uniquely identified. It is practical, model-agnostic, and reproducible, and the matched comparison already rules out a lot of confounds. I would send this to a serious referee with a request to either add one cross cell or soften the causal language. The paper deserves a round of revision rather than a desk reject.","headline":"Useful two-step recipe for closing the PE-caption gap, honestly reported — but the two-sided-alignment mechanism is stronger than the missing 2x2 cells can support.","tokens_in":27430,"tokens_out":3965,"would_cite":true,"duration_ms":38206,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Text-to-video generation improves when the same prompt enhancer both writes the captions used to fine-tune the generator and rewrites user prompts at inference, even when a shared caption schema is already in place.","keywords":["text-to-video generation","prompt enhancement","conditioning alignment","diffusion transformer","recaptioning","caption schema","PE-Caption gap","maximum mean discrepancy"],"falsifier":"A decisive test would compare a third DiT fine-tuned on captions from a same-size, same-data rewriter that is not the deployed PE: if that third DiT matches CAPE-T2V's scores and MMD2, the mechanism is not the two-sided use of the deployed PE.","tokens_in":26427,"feed_emoji":"🎬","tokens_out":10171,"duration_ms":92803,"temperature":0.7,"pith_summary":"Text-to-video diffusion transformers (DiTs) generate video from text and are trained on detailed captions, but at deployment they are conditioned on user prompts rewritten by a prompt enhancer (PE). This paper argues that even when training captions and inference prompts follow the same caption template, they still differ in detail selection, organization, granularity, and phrasing—a residual mismatch it calls the PE–Caption gap. CAPE-T2V attacks the gap on both sides: it first fine-tunes a PE to map concise captions, detailed captions, and pseudo user prompts into captioner-generated six-part targets, then uses that same Anchored PE to rewrite the dense captions used for DiT fine-tuning. At inference the same Anchored PE rewrites user prompts, so the text that trains the model and the text that conditions it come from one operator. In a matched comparison against a schema-aligned control, CAPE-T2V scores higher on StoryEval, VBench-2.0, and T2V-CompBench for both Wan2.2 and LTX-2.3, and its training captions sit closer to inference-time outputs as measured by embedding-based squared maximum mean discrepancy.","feed_headline":"Same prompt rewriter on both sides lifts all six benchmark scores","feed_subtitle":"CAPE-T2V closes the residual prompt-caption gap, outscoring the schema-matched baseline on Wan2.2 and LTX-2.3.","key_machinery":"The load-bearing object is the Anchored PE: a prompt enhancer fine-tuned so that three kinds of inputs—concise captions, detailed captions, and pseudo user prompts—all map to captioner-generated targets in a fixed six-part schema (Short Caption, Dense Caption, Main Object, Background, Camera, Style). This same frozen PE is then reused in both directions: it rewrites user prompts at inference and rewrites the video-derived dense captions used to fine-tune the DiT. The argument is carried by the inequality $D(P^{\\mathrm{user}}_{\\mathrm{PE}}, P^{\\mathrm{dense}}_{\\mathrm{PE}}) < D(P^{\\mathrm{user}}_{\\mathrm{PE}}, P^{\\mathrm{dense}}_{\\mathrm{rewriter}})$, where $D$ is instantiated as squared maximum mean discrepancy ($\\mathrm{MMD}^2$) in a frozen text-embedding space. The schema alone fixes which parts appear and their order; the Anchored PE is what fixes their realization, while the DiT architecture and training objective remain unchanged.","core_discovery":"The paper's central claim is that the identity of the caption writer matters beyond the schema it follows. Using the deployed PE to build DiT fine-tuning captions reduces the measured PE–Caption gap relative to a separately prompted schema-aligned rewriter: embedding-based MMD2 is 0.0682 versus 0.0764 in the token-length-matched comparison, and 0.0673 versus 0.0758 on the full set. CAPE-T2V beats the matched Schema-Aligned DiT on all six model–benchmark pairs, with gains of 1.6, 1.12, and 0.30 percentage points on StoryEval, VBench-2.0, and T2V-CompBench for Wan2.2, and 1.4, 0.92, and 0.65 for LTX-2.3. The anchoring step alone already improves both original generators, and the DiT adaptation adds further gains, so the two steps contribute separately. The paper concludes that two-sided conditioning alignment—not merely sharing a caption schema—is what narrows the residual mismatch.","pith_inferences":["A sharper test of the paper's mechanism would equalize model size and training data between the two caption operators; because the prompted rewriter is a much larger model than the fine-tuned Anchored PE, part of the reported gain could come from operator strength rather than from two-sided alignment.","The pseudo-user-prompt and detailed-source input types are not ablated individually, so an extension is to measure how much each construction contributes to closing the gap; the paper itself flags that this remains open.","If the mechanism generalizes, any conditional generative model with an inference-time text rewriter—such as image or audio generators—should benefit from the same two-sided recipe, a prediction this paper does not test.","The choice of caption schema is a free parameter; varying which parts it includes could reveal what makes a schema a good anchor and whether simpler schemas leave more residual gap."],"forward_implications":["Deployed text-to-video systems can narrow the conditioning gap by recaptioning their post-training data with the same prompt enhancer they use at inference, without touching the generator architecture or objective.","A shared caption schema is not enough; the caption operator that produces fine-tuning captions must match the operator that rewrites user prompts.","The recipe works without access to a generator's native captioning pipeline, making it applicable across models whose internal captioning recipes are unknown.","Both steps pay off separately, so systems that already use a strong prompt enhancer can still gain from adapting the generator to that enhancer's output distribution.","The measured reduction in embedding-space discrepancy indicates the residual mismatch is quantifiable and can be tracked without video generation, as a cheaper diagnostic."],"supporting_citations":[{"why":"Defines structured caption sharing between training and inference, the schema-aligned approach that CAPE-T2V's control is designed to beat.","marker":"Fan et al. 2025"},{"why":"Shows recaptioning training data reduces train–inference text skew, the prior idea CAPE-T2V extends by anchoring to captioner targets.","marker":"Segalis et al. 2023"},{"why":"Couples the prompt enhancer toward training-derived descriptions, the captions-to-PE direction that CAPE-T2V reverses into PE-to-captions.","marker":"Yao et al. 2024"},{"why":"Provides the MiraData captioner-generated targets and the six-part caption schema used for PE anchoring and DiT captions.","marker":"Ju et al. 2024"},{"why":"Supplies a gallery of natural user prompts used as few-shot examples to generate pseudo user prompts.","marker":"Wang and Yang 2024"},{"why":"Provides the unbiased squared maximum mean discrepancy estimator used to quantify the PE–Caption gap.","marker":"Gretton et al. 2012"},{"why":"Defines the Wan generator family, one of the two target DiTs on which the matched comparison is run.","marker":"Wan Team 2025a"},{"why":"Defines the LTX-2 generator family, the second target DiT on which the matched comparison is run.","marker":"HaCohen et al. 2026"}],"fun_headline_variants":["Same prompt enhancer on both sides closes T2V caption gap","CAPE-T2V anchors PE to cut caption gap, beats all six benchmarks","One rewriter for training and inference lifts every benchmark score","Two-sided alignment: shared PE outperforms schema-matched baseline","CAPE-T2V slashes PE-Caption gap, wins on Wan2.2 and LTX-2.3"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The matched control assumes the two fine-tuning paths differ only in which caption operator writes the captions, but the operators also differ in model size, fine-tuning, and prompt design, and the benchmark differences are reported without significance tests, so the gains cannot be uniquely pinned on two-sided alignment.","fun_headline_variants_meta":{"raw":{"variants":["Same prompt enhancer on both sides closes T2V caption gap","CAPE-T2V anchors PE to cut caption gap, beats all six benchmarks","One rewriter for training and inference lifts every benchmark score","Two-sided alignment: shared PE outperforms schema-matched baseline","CAPE-T2V slashes PE-Caption gap, wins on Wan2.2 and LTX-2.3"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000307,"raw_usage":{"total_tokens":1859,"prompt_tokens":1148,"completion_tokens":711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":764,"completion_tokens_details":{"reasoning_tokens":609}},"tokens_in":764,"tokens_out":711,"duration_ms":8023,"temperature":1.0,"reasoning_tokens":609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:58:32.747688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would compare a third DiT fine-tuned on captions from a same-size, same-data rewriter that is not the deployed PE: if that third DiT matches CAPE-T2V's scores and MMD2, the mechanism is not the two-sided use of the deployed PE.","supporting_citations":[],"review_version":1}