{"id":"1c8035ea-eebe-46f5-b01e-02c02f4872d0","arxiv_id":"2606.27741","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Self-imagination fine-tuning from pure noise plus motion-classifier supervision breaks reconstruction shortcuts and reduces camera-object motion entanglement in video diffusion models.","lead":"Video diffusion models often entangle camera and object motions because training shortcuts let them copy residual cues from noisy inputs instead of reasoning from text. SIFT fine-tunes them by generating from pure noise under motion-classifier feedback, improving kinematic disentanglement without needing special video data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Classifier idiosyncrasies may drive the reported disentanglement gains more than genuine kinematic inference.","rationale":"The reader correctly isolates the weakest link: the quality and unbiasedness of the four-way classifier signal. All other design choices (pure-noise start, progressive replay, retained MSE) are secondary; if that signal is merely a noisy proxy for appearance statistics, the self-imagination loop cannot be said to teach kinematic reasoning. The empirical gains are real and consistent across backbones, so the paper remains a solid engineering contribution worth conditional acceptance and further stress-testing; the concern does not overturn the verdict but confirms why it must stay CONDITIONAL rather than unconditional ACCEPT. No stronger internal inconsistency or data-fabrication issue is present.","tokens_in":16553,"tokens_out":510,"duration_ms":14466,"concrete_test":"Hold out a disjoint 500-video motion-labeled set never seen by either classifier; train a third architecture (e.g., VideoMAE or a simple optical-flow + MLP) to the same four-way task. Score the final SIFT-generated videos from the 100-prompt test set with this independent classifier. If its accuracy on the intended labels falls below ~65 % while human/VLM SA-PC remain high, the supervision is idiosyncratic and the strongest claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on motion-aware discriminative supervision from two four-way classifiers (R3D 78.4 %, SlowFast 82.8 % on noisy one-step reconstructions of a 4 000-video set; §3.4). Because the generator is optimized solely against these classifiers’ cross-entropy (alternating, with hard-case replay of their failures), any systematic bias or residual correlation they retain can be exploited without the model ever learning independent reference frames. The ablation in Table 4 shows single-classifier variants degrade, yet that only demonstrates that the two classifiers disagree; it does not prove that their consensus equals true kinematic disentanglement. The coarse taxonomy itself (camera-only / object-only / both / static) supplies no signal on direction, magnitude or multi-entity trajectories, exactly the axes where residual entanglement and the failure case in Fig. S2 still appear. Thus the SA/PC lifts and human preferences could largely reflect successful classifier matching rather than the claimed physically grounded motion prior.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper identifies “motion entanglement” (unintended coupling of camera and object motion) as a kinematic failure mode of video diffusion models, attributing it to data bias and residual motion cues that create a reconstruction shortcut under pixel-level MSE training. Diagnostic experiments (Fig. 2) show that pretrained models largely ignore prompt and temporal-order perturbations when reconstructing from noisy inputs. The authors propose Self-Imagination Fine-Tuning (SIFT): generation from pure noise under freely generated text prompts, motion-aware discriminative supervision from two four-way classifiers (R3D and SlowFast) that label camera-only / object-only / both / static, and progressive hard-case replay. A lightweight MSE term on real pairs is retained for visual quality. Experiments on Wan2.1-1.3B and CogVideoX report gains in Semantic Adherence and Physical Commonsense under both VLM and human evaluation, with ablations (Table 4) isolating each component and additional tests on multi-object, articulated, and long-horizon settings.","tokens_in":16850,"tokens_out":1171,"duration_ms":11587,"significance":"If the gains reflect genuine kinematic disentanglement rather than classifier matching, the work supplies a practical, data-efficient route to improve motion priors in text-to-video models without collecting motion-decoupled video-text pairs. The diagnostic of residual-motion shortcuts is clean and reusable, the self-imagination framing is a clear conceptual shift from reconstruction-based SFT, and the method is shown on two open backbones with both automatic and human preference evidence. The free-prompt coverage of rare relative-motion combinations is a genuine practical advantage. These strengths make the paper a useful contribution to physics-aware video generation even if the supervisory taxonomy remains coarse.","major_comments":[{"comment":"§3.4 and Table 4: The central claim that SIFT teaches “kinematically-grounded” independent reference frames rests almost entirely on cross-entropy against two four-way classifiers (R3D 78.4 %, SlowFast 82.8 % on noisy one-step reconstructions of a 4 000-video set). The ablation shows that single-classifier variants degrade and that alternating helps, but this only demonstrates that the two classifiers disagree; it does not establish that their consensus equals true kinematic disentanglement rather than shared inductive biases. Because the generator is optimized solely against these classifiers (plus a small MSE term), residual correlations or artifacts they retain can be exploited without learning independent frames. A load-bearing check is missing: e.g., held-out human or optical-flow-based verification that the improved videos actually separate camera and object trajectories on axes th","section":null},{"comment":"§4.1–4.2, Tables 1–3 and Fig. 4: Evaluation uses a 100-prompt test set (plus 50-prompt generalization subsets) scored on 1–5 Likert SA/PC by InternVideo2.5 and 20 human raters, with no error bars, confidence intervals, or inter-rater reliability. The absolute numbers and preference rates are therefore hard to interpret as robust evidence of “substantial” improvement. At minimum the paper should report variance across prompts or raters and clarify how many unique videos underlie the human preference percentages.","section":null},{"comment":"Limitations and Fig. S2: The four-category taxonomy supplies no signal on direction, magnitude, or temporal transitions. The supplementary failure case (clockwise prompt yielding counter-clockwise camera motion) and residual entanglement visible in qualitative figures show that the claimed physically grounded prior remains incomplete precisely on the axes the supervision never constrains. This does not invalidate the reported gains, but it does limit the strength of the claim that SIFT endows models with “internalized physical reasoning” beyond the coarse camera/object dichotomy.","section":null}],"minor_comments":[{"comment":"§3.3 and Algorithm 1: The precise schedule of the few denoising steps (t = 1000, 980, 960) and the weight λ = 0.01 are stated only in the implementation details; a short sensitivity note would help readers judge free-parameter dependence.","section":null},{"comment":"Fig. 1 caption and main text: “red indicates physically implausible relative motion” is clear in the figure but the trajectory visualization method itself is never described; a one-sentence explanation would improve reproducibility.","section":null},{"comment":"Table 4 caption correctly warns that scores are relative within the ablation group and not comparable to Table 1; the same caveat should be repeated in the main text when the ablation is discussed.","section":null},{"comment":"Related Work: The distinction between kinematic entanglement and dynamics-focused physics methods is useful; a brief pointer to any concurrent work on residual-noise shortcuts in video diffusion would strengthen the positioning.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core diagnostic (Fig. 2) and the self-imagination idea are solid and publishable. The main risk is over-claiming “physical reasoning” when the only supervisory signal is a coarse four-way classifier pair. If the authors can add even a modest independent trajectory or human verification that the gains survive axes the classifiers never see, the paper becomes much stronger; otherwise the claims should be dialed back to “improved camera/object category consistency under self-imagined prompts.” Fit for a solid CV venue after revision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful part of this paper is the diagnosis plus the training shift. They show, with a simple four-setting MSE experiment on Wan, that the model mostly rides residual motion cues in the noisy input rather than the prompt. That explains why ordinary SFT on motion-decoupled clips barely helps. Their fix is to drop real video entirely, generate from pure noise under LLM prompts, and supervise only with two alternating four-way motion classifiers (R3D + SlowFast) plus a progressive hard-case buffer. That combination is new relative to standard SFT and to VideoREPA-style alignment, and the gains show up consistently on Wan and CogVideoX under both InternVideo2.5 and human preference.\n\nWhat they do well: the ablations isolate self-imagination, alternating classifiers, and the hard-case schedule; the method needs no paired motion-decoupled video data; and they are honest about the coarse taxonomy and the remaining direction failures (Fig. S2). The free-parameter list is short and the free-prompt generation is a genuine practical advantage.\n\nThe soft spot is exactly the one the stress-test flags, but it is not load-bearing enough to sink the claim. The generator is optimized against classifiers that sit at 78–83 % accuracy on noisy one-step reconstructions. Alternating them reduces single-classifier overfitting, yet it does not prove the consensus equals true kinematic reference frames. Direction, magnitude, and multi-entity trajectories are outside the four-way label set, so residual entanglement can still hide. The 100-prompt test set and Likert scores without error bars keep the evaluation modest. Still, the paper never pretends the classifiers are perfect physics oracles; it treats them as a practical, data-efficient signal that works better than reconstruction. That is fair.\n\nThis is for people who train or fine-tune open video diffusion models and care about camera/object decoupling for simulation or control. It is not a new theory of physics, just a solid engineering recipe. I would send it to peer review; the diagnosis is clean, the method is reproducible in principle, and the caveats are already stated. Worth reading and trying.","headline":"Clean diagnosis of motion entanglement plus a practical pure-noise fine-tuning recipe that beats SFT and VideoREPA on two backbones; classifier-driven supervision is the real soft spot, not a fatal flaw.","tokens_in":17434,"tokens_out":550,"would_cite":true,"duration_ms":5882,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Self-Imagination Fine-Tuning forces video diffusion models to invent motion from text alone, ending the habit of copying residual cues from noisy inputs and producing disentangled, physically plausible camera and object trajectories.","keywords":["text-to-video generation","motion entanglement","video diffusion models","self-imagination fine-tuning","physical plausibility","motion disentanglement","kinematic reasoning"],"falsifier":"If human raters or an independent physics checker score a large held-out set of novel camera-versus-object prompts and find no statistically significant improvement in relative-trajectory correctness or physical-commonsense scores for SIFT-tuned models over the untuned baseline, the central claim is false.","tokens_in":17464,"feed_emoji":"🎬","tokens_out":744,"duration_ms":15941,"temperature":0.7,"pith_summary":"Video diffusion models produce visually convincing frames yet routinely entangle independent motions: a camera that should orbit a stationary object also drags the object, or a fixed camera tracks a moving subject. The authors show this kinematic failure stems from two sources—real videos almost always mix camera and object motion, and the standard reconstruction objective lets the model simply copy residual motion that survives heavy noise rather than reason from the text prompt. Their Self-Imagination Fine-Tuning (SIFT) discards real video inputs, starts every sample from pure Gaussian noise, and trains the model solely on its own generations under feedback from motion classifiers that label camera-only, object-only, both-moving or both-static. Progressive hard-case replay gradually re-introduces failures so learning stays stable. Because prompts can be generated freely by language models, the method densely covers rare or finely disentangled motion combinations that would be expensive to film. The result is measurably better physical realism, relative-motion control and prompt adherence without any curated motion-decoupled video dataset.","feed_headline":"Video models learn real motion by imagining it from pure noise","feed_subtitle":"SIFT breaks the reconstruction shortcut so camera and object paths finally stay independent","key_machinery":"Self-Imagination Fine-Tuning (SIFT): the generator is fine-tuned by sampling pure noise, running a few high-noise denoising steps under an LLM-generated prompt that specifies a known motion class, then receiving cross-entropy loss from alternating R3D and SlowFast classifiers that judge whether the result is camera-only, object-only, both-moving or both-static; a hard-case buffer replays misclassified samples with probability that rises linearly during warm-up.","core_discovery":"Motion entanglement—the unintended coupling of independent kinematic sources such as camera movement and object motion—is a pervasive failure of current video diffusion models. It arises because training data rarely separate the two motions and because denoising objectives allow the model to reconstruct residual motion cues still present in noisy inputs instead of inferring dynamics from language. Self-Imagination Fine-Tuning removes real-video inputs, generates from pure noise under freely generated prompts, and replaces pixel reconstruction with motion-aware discriminative supervision plus progressive hard-case replay. Experiments on two open-source backbones show substantial gains in phys","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Self-imagination fine-tuning untangles camera and object motion","Video models learn kinematics by training on pure-noise generations","SIFT breaks reconstruction bias so motions stay independent","Fine-tune diffusion on self-generated clips for physical plausibility","Motion entanglement fixed via self-imagined free-prompt training"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The four-way motion classifiers, trained on only four thousand noisy one-step reconstructions and reaching roughly eighty percent accuracy, supply a clean enough signal that the generator learns true kinematic disentanglement rather than merely matching classifier quirks.","fun_headline_variants_meta":{"raw":{"variants":["Self-imagination fine-tuning untangles camera and object motion","Video models learn kinematics by training on pure-noise generations","SIFT breaks reconstruction bias so motions stay independent","Fine-tune diffusion on self-generated clips for physical plausibility","Motion entanglement fixed via self-imagined free-prompt training"]},"model":"grok-4.5","effort":"low","cost_usd":0.00495,"raw_usage":{"total_tokens":1415,"prompt_tokens":789,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":49500000,"prompt_tokens_details":{"text_tokens":789,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":555,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":789,"tokens_out":71,"duration_ms":6858,"temperature":1.0,"reasoning_tokens":555,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T10:27:46.299638+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"If human raters or an independent physics checker score a large held-out set of novel camera-versus-object prompts and find no statistically significant improvement in relative-trajectory correctness or physical-commonsense scores for SIFT-tuned models over the untuned baseline, the central claim is false.","supporting_citations":[],"review_version":2}