{"id":"1b3c4bf6-4dbb-414a-9615-1e8a4df34822","arxiv_id":"2607.27380","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Using executable Blender code as an intermediate simulation draft improves physical consistency in text-to-video generation, lifting OmniWeaving from 0.475 to 0.558 on PhyGenBench and from 52.18% to 77.88% on VBench-2.0.","lead":"VideoCoCo makes text-to-video models generate more physically correct videos by first having an AI write and run a Blender simulation, then turning that simulation into a realistic video. The system beats current models on physics benchmarks, showing that executable code can act as a reliable 'chain of thought' for video dynamics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own ablation undercuts the 'draft carries physics' claim: LoRA tuning adds more score than the draft does (Table 3), and LLM-judge validity is unexamined.","rationale":"The reader's concern about LLM-judge validity is legitimate, but the most direct problem is internal: the ablation table contradicts the summary that drafting supplies most of the gain. This matters because the paper's claimed contribution is the Code-as-CoT mechanism, not merely another fine-tuned editor. The system-level scores could still be correct, so rejection is not warranted; conditional acceptance remains the right stance. The proposed human-rating study would simultaneously test judge validity and, if combined with trajectory/flow comparisons, the draft-preservation mechanism. Secondary limitations—Blender's expressiveness ceiling and the absence of error bars—do not change this assessment.","tokens_in":12942,"tokens_out":7437,"duration_ms":74790,"concrete_test":"Human-rating validation: sample 100 prompts from PhyGenBench (and release the exact VBench-2.0 subset), generate videos with OmniWeaving, VideoCoCo-TuneFree, and VideoCoCo-LoRA, and have multiple physics-aware human raters score physical consistency blind to condition. If human ordering does not reproduce the GPT-4o ordering (TuneFree > base and LoRA > TuneFree), the headline claim is not established. Also report inter-rater agreement and judge–human correlation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing concern: the central decomposition claim is not supported by the paper's own evidence. In Table 3, VideoCoCo Tune-Free improves PhyGenBench average over OmniWeaving by ~0.031 (0.475→0.506), while LoRA-Tune over Tune-Free improves by ~0.052 (0.506→0.558). Thus editor adaptation contributes roughly 63% of the total gain, yet the text says 'Drafting supplies most of the physical gain.' The benchmarks rely on GPT-4o (PhyGenBench) and an undefined 'pre-registered subset' (VBench-2.0); no human correlation for the judge is provided. Since the editor changes appearance/style in addition to motion, the end-task improvements could reflect the judge's preference for cleaner restyled videos rather than the draft's physical fidelity. There is also no direct check that final videos preserve the Blender draft's spatiotemporal trajectories, so 'the editor only restyles' is an assumption, not a measured fact. Missing code/data/checkpoints make these claims unverifiable from the preprint.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VideoCoCo, an agentic dual-engine framework for physically consistent text-to-video generation. A coding agent writes a Blender program from a text prompt; the program is executed in a sandbox to render a deterministic, low-fidelity 'white-clay' spatiotemporal draft. An instruction agent composes an appearance-focused editing instruction, and a video editor is conditioned jointly on the draft and instruction to produce the final photorealistic video. To make off-the-shelf editors consume simulated drafts, the authors construct VideoCoCo-3K, a dataset of draft–instruction–target triplets generated by a teacher editor (Seedance 2.0), and fine-tune the editor with a conditional denoising objective (Eq. 7). Experiments on PhyGenBench and VBench-2.0 report substantial gains over the OmniWeaving base (0.475 to 0.558 and 52.18% to 77.88%, respectively), with ablations separating tuning-free and fine-tuned variants. The central claim is that executable code serves as a process-level chain of thought, decoupling physical dynamics from appearance synthesis.","tokens_in":13208,"tokens_out":2568,"duration_ms":26395,"significance":"If the claims hold, VideoCoCo is a meaningful step toward controllable and inspectable physical video generation: unlike textual plans or keyframes, executable code can be run, inspected, and revised, and the dense simulated draft provides explicit temporal grounding. The paper includes a useful ablation isolating the draft's contribution from editor adaptation, reports per-category scores on two physics benchmarks, and identifies clear failure cases (e.g., turbulent fluids). However, the strength of the central decomposition claim depends on untested assumptions about the LLM judges used for evaluation and on the attribution of gains in Table 3, which currently appears to contradict the text. The absence of released code, data, and evaluation details further limits verification. The idea is promising and the empirical framework is coherent, but the evidence as presented does not yet fully support the paper's strongest claims.","major_comments":[{"comment":"The text states that 'Drafting supplies most of the physical gain,' but the Table 3 numbers show the opposite attribution. OmniWeaving→Tune-Free improves the average by 0.506−0.475 = 0.031, while Tune-Free→LoRA-Tune improves by 0.558−0.506 = 0.052; thus editor adaptation contributes roughly 63% of the total gain. This is a load-bearing inconsistency for the central decomposition claim. Please correct the claim, or provide a controlled edit-only baseline (e.g., omitting the draft but keeping the instruction and LoRA adaptation) that would isolate the draft's contribution.","section":"§4.3, Table 3; §4.2 'Takeaway'"},{"comment":"PhyGenBench scores rely entirely on GPT-4o as the MLLM judge, and VBench-2.0 uses a 'pre-registered subset' that is never defined. No human agreement, judge discriminability analysis, or complete benchmark results are reported. Since the editor restyles appearance (lighting, texture, camera), the observed gains could reflect judge preference for cleaner restyled videos rather than physical fidelity. Please report judge–human correlation (or at least a human-rated subset), specify the exact VBench-2.0 prompts/suite used, and justify why the subset is representative.","section":"§4.1 'Datasets and Metrics'"},{"comment":"The training setup is underspecified. Eq. (7) is a generic conditional denoising loss, but the architecture changes needed to condition the editor on the draft video d are not described: is d injected as concatenated frames, via cross-attention, or through an adapter? What LoRA rank, learning rate, training steps, and compute were used? Without these details the 'LoRA-Tune outperforms Full-Tune' result cannot be reproduced or interpreted. Additionally, the VideoCoCo-3K targets are generated by Seedance 2.0 using the same draft-and-instruction pipeline; there is no verification that the teacher outputs actually preserve the draft's spatiotemporal trajectories, which is critical because the editor learns to imitate the teacher.","section":"§3.4, Eq. (7)"},{"comment":"The project page (https://github.com/micky-li-hd/VideoCoCo) is listed but no code, data, or checkpoints are provided in the manuscript. Given the paper's reliance on a new dataset and a tuned editor, and its claim of reproducibility ('fully automatic, inspectable, and reproducible'), releasing at least the dataset metadata, Blender programs, and evaluation scripts is necessary for the claims to be verifiable. Please clarify the release plan.","section":"Reproducibility"}],"minor_comments":[{"comment":"The abstract and text use 0.475/0.558 while Table 1 rounds to 0.48/0.56; please standardize to avoid apparent inconsistency.","section":"Table 1 vs. §4.2"},{"comment":"The figure contains placeholder symbols ('??') in the draft rendering chain, which make the pipeline unclear. Use actual rendered draft frames or remove the placeholders.","section":"Figure 2"},{"comment":"'VideoRL VR' reads as a typo; clarify the method name and citation.","section":"§2.3"},{"comment":"The term 'pre-registered subset' appears only here; define it in the text or in an appendix, including how it was selected from VBench-2.0.","section":"§4.1"},{"comment":"The loss is written as plain L2 on the noise residual; specify whether any reweighting (e.g., SNR weighting) is used, and whether the model is trained with classifier-free guidance on the draft condition.","section":"§3.4, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is appealing and the benchmark results are striking, but the attribution of gains to drafting versus editor adaptation is internally inconsistent, and the evaluation rests on LLM judges with no demonstrated validity for physical consistency. The self-aligned construction of VideoCoCo-3K using the same teacher and pipeline also deserves careful scrutiny, though this is not circular in the strict sense. With the current evidence I cannot recommend acceptance; the authors should be asked to supply controlled ablations, judge-validation, and full training/reproducibility details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new way to inject physical priors into video generation — write Blender code, render a deterministic draft, and condition a video editor on it. That's worth taking seriously. But the paper claims 'drafting supplies most of the physical gain' in a way that its own Table 3 does not support: Tune-Free adds 0.03 over OmniWeaving, LoRA tuning adds another 0.05. So editor adaptation contributes a larger chunk of the score, and without a direct measure of whether final videos preserve the draft's trajectories, the central decomposition is more asserted than established.\n\nWhat's good: the framework is clean and well-motivated, the idea of executable code as a process-level chain of thought is a real extension beyond DraCo/CoCo for images, and the ablation structure (tuning-free vs LoRA vs full) is the right way to try to isolate the draft contribution. The benchmark results, if trustworthy, are strong — best average on PhyGenBench and VBench-2.0 physical dimensions, with particularly large gains on material and thermal. The paper also honestly notes its own limits (latency, Blender expressiveness).\n\nWhere the softness is: first, the ablation interpretation. The text says the draft supplies most of the physical gain, but the numbers show the opposite in terms of raw improvement. You could defend it as 'the draft enabled the rest,' but that's not what's written. Second, both benchmarks depend on LLM judges (GPT-4o for PhyGenBench; an undefined 'pre-registered subset' for VBench-2.0) with no human correlation reported. That's a field-wide issue, but here it's load-bearing because the headline gains are large and the judge could be rewarding restyled appearance over physical fidelity. Third, there's no check — automatic or human — that the final video preserves the simulated spatiotemporal structure. That's the core claim, and it's unmeasured. Fourth, no code/data/checkpoints are provided, so none of this is independently checkable, and the dataset is self-built with the same pipeline components, including a Seedance teacher, adding a real but not fatal circularity.\n\nThis paper is for people working on physics-aware video generation who want an alternative to pure reward-based post-training. It deserves a serious referee, but that referee should ask for the missing artifacts and an honest re-analysis of the ablation before the strong claims are accepted.","headline":"Clever code-as-CoT video pipeline with solid benchmark gains, but the central decomposition claim is not actually proven by the ablation, and missing code/data makes results unverifiable.","tokens_in":13801,"tokens_out":3730,"would_cite":true,"duration_ms":36016,"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":"Executable Blender code as a process-level chain of thought enforces physical consistency in text-to-video generation.","keywords":["video generation","physical consistency","chain-of-thought","executable code","Blender simulation","draft-conditioned editing","text-to-video","world modeling"],"falsifier":"Run the same PhyGenBench and VBench-2.0 prompt suites through VideoCoCo and the base generator and have human annotators with physics training rate which video better matches the requested dynamics; if human ratings do not reproduce a substantial gap favoring VideoCoCo, the central claim that the draft carries the physics is not established.","tokens_in":12836,"feed_emoji":"🎬","tokens_out":4164,"duration_ms":36971,"temperature":0.7,"pith_summary":"The paper tries to establish that physically consistent video generation can be achieved by separating the physics from the pixels: have an agent write an executable Blender program specifying the scene and its temporal evolution, run it in a sandbox to get a deterministic low-fidelity draft, then condition a video editor on that draft to add photorealistic appearance. The claim is that this executable, frame-dense intermediate—a 'process-level chain of thought'—carries the physical dynamics, so the generative model only needs to restyle, not reason about how objects move. The authors report that this raises the average physical-consistency score from 0.475 to 0.558 on PhyGenBench and from 52.18% to 77.88% on VBench-2.0, the best average on both benchmarks, with the largest gains in thermal and material dynamics where appearance-driven priors are weakest. A sympathetic reader would care because it offers a controllable, inspectable way to inject physics into video generation without hand-authored rules per phenomenon.","feed_headline":"Drafting video physics in Blender code lifts consistency scores","feed_subtitle":"VideoCoCo writes a runnable simulation, renders the draft, and lets the video model only restyle its physics","key_machinery":"The central object is the 'process-level chain of thought': an executable Blender program that explicitly declares objects, physical properties, and temporal evolution, rendered into a deterministic spatiotemporal draft. Named in the paper as Code-as-CoT, this representation is explicit (every object and motion is declared), executable (the program actually runs and produces every frame), and inspectable (the code can be read, edited, and re-run). Its work in the argument is to commit the system to a concrete physical realization before any pixel is synthesized, so that the downstream generative video engine is asked only to restyle an already-instantiated process; the VideoCoCo-3K dataset s","core_discovery":"The central claim is that executable code, not textual plans or keyframes, should be the intermediate reasoning representation for video generation. VideoCoCo externalizes the process hidden in a text prompt as a self-contained Blender Python program; running that program in a sandbox yields a deterministic, temporally dense draft of the event. A second engine, a video editor fine-tuned on the VideoCoCo-3K dataset of draft–instruction–target triplets, takes this draft as a structural condition and produces a photorealistic video. The paper argues that this decomposition lets the simulation engine commit to what happens and when, while the editor decides only how it looks, and presents benchm","pith_inferences":["The same code-as-draft idea could be pushed one step further: instead of a video editor conditioned on a draft, a video model could be trained with synthetic intermediate targets so that at inference the simulation is not needed, as the authors themselves suggest for future work—but the paper does not demonstrate this.","Because the staged construction is automatic, a natural extension is to close the loop with a verifier that checks whether the rendered draft matches the prompt before editing; the paper does not implement such a rejection loop.","The benchmark gains are only as trustworthy as the judges; a human rating study on the same prompt set would test whether the gap reflects genuine physical plausibility rather than a stylistic preference for the editor's outputs."],"forward_implications":["If the claim holds, any prompt whose physics can be expressed in a simulator can be made physically consistent by drafting first, independent of the video model's own physical priors.","The largest gains in thermal and material dynamics suggest that appearance-driven generators are weakest exactly where an explicit simulator adds the most, pointing to a division of labor for future systems.","VideoCoCo-3K-style aligned triplets become a general training resource: any video editor can be adapted to consume simulated drafts, so improvements in the editor transfer to the whole pipeline.","Because the intermediate artifacts are code and rendered video, failures can be debugged by editing the program, not by re-prompting the generator; this makes physical errors inspectable and correctable.","LoRA adaptation outperforming full fine-tuning indicates that the needed skill—restyling drafts while preserving motion—is narrow and can be learned with modest data, lowering the cost of adapting the pipeline to new editors."],"fun_headline_variants":["VideoCoCo: code as chain-of-thought for video physics","Blender code drafts video dynamics, boosts consistency benchmarks","Executable code as video reasoning beats text prompts","Agentic dual-engine generates physically consistent videos from code","VideoCoCo renders simulation drafts to steady video physics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the LLM judges used by the benchmarks score genuine physical consistency, so the reported gains really reflect better physics rather than a bias toward the edited videos' appearance; the paper offers no human comparison to confirm this.","fun_headline_variants_meta":{"raw":{"variants":["VideoCoCo: code as chain-of-thought for video physics","Blender code drafts video dynamics, boosts consistency benchmarks","Executable code as video reasoning beats text prompts","Agentic dual-engine generates physically consistent videos from code","VideoCoCo renders simulation drafts to steady video physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000557,"raw_usage":{"total_tokens":2513,"prompt_tokens":798,"completion_tokens":1715,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1637}},"tokens_in":542,"tokens_out":1715,"duration_ms":11892,"temperature":1.0,"reasoning_tokens":1637,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T08:23:34.002878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same PhyGenBench and VBench-2.0 prompt suites through VideoCoCo and the base generator and have human annotators with physics training rate which video better matches the requested dynamics; if human ratings do not reproduce a substantial gap favoring VideoCoCo, the central claim that the draft carries the physics is not established.","supporting_citations":[],"review_version":1}