{"id":"3771ed72-4246-4601-9052-a6639908c2d5","arxiv_id":"2505.03730","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FlexiAct transfers actions from a reference video to an arbitrary target image, allowing changes in layout, skeleton, and viewpoint while keeping the target subject's appearance.","lead":"This paper presents a video generation method that copies the motion from one clip onto a different person or animal shown in a single still image, even when body shape, camera angle, or art style differ. It makes AI video editing more flexible: one action clip could drive many different characters, though each clip still needs a separate training step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FAE's frequency-motion separation is supported only by qualitative attention maps, and the fixed W_bias schedule is calibrated without RefAdapter but applied with it.","rationale":"The reader's weakest assumption is essentially correct: the FAE mechanism is supported only by visual attention maps and a hand-set scalar schedule. I agree that this is the most load-bearing point, because FAE is the claimed novelty and the ablation only demonstrates that some attention bias helps, not that the frequency interpretation is right. I would add one aggravating detail: FAE is trained without RefAdapter, yet deployed with it, so the attention statistics used to motivate Eq. (1) may not even describe the inference configuration. This is a concrete train/inference mismatch that the paper does not address. The paper still has genuine independent support: the components are ablated, an adapted CogVideoX-I2V baseline is included, human preference results are reported, and code/weights are claimed to be released. My concern is not that the system fails, but that the stated reason it works is underdetermined; an early-timestep attention boost would plausibly help coarse motion control in many diffusion pipelines regardless of any 'frequency' separation. Therefore the reader's CONDITIONAL verdict should stand, with the additional recommendation that the authors provide either quantitative frequency analysis or causal schedule interventions before the mechanism is taken as established.","tokens_in":12597,"tokens_out":9668,"duration_ms":109216,"concrete_test":"On the fixed 250-pair benchmark, rerun the Ours pipeline with the W_bias schedule inverted: apply the same positive bias at timesteps t < 100 instead of t in [700, 800], keeping alpha and all other settings unchanged, and report Motion Fidelity and Appearance Consistency alongside the published 'w/o FAE' and 'Ours' rows. If the inverted schedule preserves the Motion Fidelity gain, or if the forward schedule does not significantly degrade when the bias is shifted to late (high-frequency) timesteps, then the directional frequency-motion claim in Section 3.4 is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the per-video frequency-aware embedding actually separates low-frequency motion from high-frequency appearance across denoising timesteps, so that adding W_bias (Eq. 1; alpha=1, t_h=700, t_l=800) to attention logits during t in [700, 800] increases action transfer without leaking reference appearance. Section 3.4 supports this premise only with the attention-map montage in Figure 2; there is no quantitative frequency decomposition, no per-layer/head/timestep statistics, and no sensitivity analysis of alpha/t_h/t_l. Moreover, Section 3.5 states that FAE is trained without RefAdapter loaded, 'to protect its conditioning ability,' while inference loads RefAdapter into the same MMDiT. The Figure 2 maps therefore characterize the training configuration, not the deployed one; if RefAdapter's LoRA shifts the attention distribution, the fixed bias may under-amplify motion or leak reference appearance. The ablation 'w/o FAE vs. Ours' shows only that some positive bias helps on their benchmark; it does not test whether the schedule direction or the frequency interpretation is correct. If this premise fails, the claimed 'first denoising process-based action extraction framework' is reduced to an unexplained, hyperparameter-sensitive recipe even if the qualitative results survive.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlexiAct, an action-transfer framework built on CogVideoX-I2V that takes a reference video and an arbitrary target image and generates a video of the target subject performing the reference action, without requiring alignment of layout, skeleton, or viewpoint. The method has two novel components: RefAdapter, a lightweight LoRA-based image-conditioned adapter trained with randomly sampled condition frames to support spatial adaptation and appearance consistency, and FAE (Frequency-aware Action Extraction), which trains per-video learnable embeddings and, at inference, adds a hand-crafted attention bias W_bias (Eq. 1) at early denoising timesteps based on qualitative observations of attention maps. The paper evaluates on a self-constructed benchmark of 250 reference-video/target-image pairs with 25 action categories, comparing against a reimplemented MD-I2V baseline and a BaseModel, using automatic metrics and a human preference study. The authors claim state-of-the-art flexible action transfer, the first denoising process-based action extraction framework, and release code and model weights.","tokens_in":12856,"tokens_out":5856,"duration_ms":55536,"significance":"If the central claim holds, FlexiAct would be a practically useful contribution to action customization in heterogeneous scenarios, since it relaxes the spatial-alignment constraints of pose-based and global-motion methods while maintaining identity. The paper has clear strengths: it includes separate ablations for both proposed components, uses human preference evaluation, reports qualitative results across diverse subjects including animals and cross-domain characters, and promises release of code and model weights. However, the central 'frequency-aware' mechanism is supported only by qualitative attention-map inspection, the automatic metrics lack error bars and significance tests, and the human study is small; these issues make the significance conditional on additional evidence rather than established by the current manuscript.","major_comments":[{"comment":"The automatic metrics are reported as point estimates with no error bars, confidence intervals, or significance tests. For example, the Motion Fidelity gap between w/o FAE (0.3614) and Ours (0.4103), and the Appearance Consistency gap between w/o RefAdapter (0.9021) and Ours (0.9162), could be within run-to-run or seed-to-seed variation. Please report variance over seeds or paired per-item statistics with a significance test. The evaluation set is also self-constructed and not released, which limits reproducibility of these numbers.","section":"Section 4.2, Table 1"},{"comment":"The load-bearing premise that the frequency-aware embeddings separate low-frequency motion from high-frequency appearance is supported only by a qualitative attention-map montage. There is no quantitative frequency decomposition, no per-layer/head/timestep statistics, and no sensitivity analysis for the parameters alpha, t_h, and t_l. Moreover, FAE is trained without RefAdapter (Section 3.5) but deployed with it, so the Figure 2 attention maps characterize the training configuration rather than the deployed configuration; if RefAdapter's LoRA shifts the attention distribution, the fixed bias may not behave as claimed. I recommend adding a quantitative frequency analysis and a control experiment that varies the schedule direction (e.g., applying the bias at late timesteps or at random timesteps) to show that the specific early-timestep schedule, and not any positive bias, is what drives the gain.","section":"Section 3.4, Eq. (1), Figure 2"},{"comment":"The human study uses 5 raters, each comparing 50 randomly selected pairs, and the paper reports preference percentages without confidence intervals, significance tests, or inter-rater agreement. Additionally, the statement that '5 raters assessed each generated video' is inconsistent with the described protocol in which each rater compares 50 random pairs. Please clarify the protocol and provide statistical support for the preference results, e.g., bootstrap confidence intervals or a paired significance test, along with details of rater instructions.","section":"Section 4.2, Human Evaluation"},{"comment":"The baseline MD-I2V is a reimplementation of MotionDirector on CogVideoX-I2V, but no evidence is provided that the reimplementation matches MotionDirector's original behavior or that hyperparameters were tuned for the I2V setting. The text says 'identical training settings to our methods' while also separately reporting MotionDirector's original 3,000/300 LoRA steps, which is confusing. To substantiate the claim of surpassing existing methods, please include at least one additional global-motion baseline and, on the human subset where pose is available, a pose-based baseline.","section":"Section 4.1, Comparison Methods"}],"minor_comments":[{"comment":"The cosine transition function uses the variable x in its argument but t elsewhere; please use t consistently for the denoising timestep.","section":"Eq. (1)"},{"comment":"The table header is ambiguous: 'Appearance Consistency↑' appears before the human-evaluation columns 'Motion Consistency Appearance Consistency', making it unclear which columns belong to automatic versus human evaluation. Also, 'are prefer over' should be 'are preferred over'.","section":"Table 1"},{"comment":"The citation 'MotionDirector [Tu et al. 2024a]' is inconsistent with the later 'MD-I2V [Zhao et al. 2023]'; in the reference list, Tu et al. 2024a is MotionEditor. Please correct the citation.","section":"Section 4.1"},{"comment":"The notation 'N Gaussian noise' is unclear; please specify that N is the noise latent and state its shape or refer explicitly to the latent noise described in Section 3.2.","section":"Figure 3"},{"comment":"The limitation that FAE requires per-video optimization is acknowledged, but please add a brief discussion of computational cost in practice, since this affects the practical utility of the method.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the venue's scope, and the empirical ablations indicate a potentially useful system. The main concern is that the central mechanism claim—frequency-aware separation during denoising—needs quantitative support and a control experiment before the paper can be accepted. I would also encourage the editor to weigh whether the small human study and lack of statistical tests meet the bar for the claims made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Terry, quick take on FlexiAct: it's a genuinely useful engineering contribution to action transfer, with a nice trick for loosening I2V's first-frame constraint, but the mechanistic claim about frequency-aware embeddings is supported only by eyeballed attention maps. The paper deserves a serious referee, but the next revision should tighten the evidence.\n\nWhat's new: unlike pose-guided or global-motion methods, FlexiAct transfers an action from a reference video to an arbitrary target image without requiring matching skeletons, layouts, or viewpoints. The two components are RefAdapter, a LoRA-based image conditioner trained with a randomly selected frame (not the first frame) as the condition and with the first latent embedding replaced by the condition embedding, and FAE, a per-video learnable embedding plus an inference-time timestep-dependent bias on attention weights (Eq. 1) that boosts early denoising steps to emphasize motion. The random-frame conditioning is a simple, sensible departure from standard I2V training and likely generalizes. The qualitative results, especially cross-domain transfers (human to animal, real to animated), look convincing, and the authors release code and weights.\n\nSoft spots: the frequency interpretation is the weakest link. Figure 2 shows attention maps at three timesteps, but there is no quantitative frequency decomposition, no per-layer/head statistics, and no causal test that the bias actually amplifies low-frequency motion. More importantly, FAE is trained without RefAdapter loaded (Section 3.4), yet at inference RefAdapter is loaded; the attention maps may not describe the deployed configuration. The schedule (alpha=1, t_h=700, t_l=800) is hand-set and justified by artifact analysis in Figure 11 on the same examples, so there is some circularity. Evaluation is also thinner than I'd like: 250 self-constructed pairs, no released benchmark, no error bars on the automatic metrics, and a 5-rater human study. The baseline set is just one reimplemented global-motion method and a base model; recent methods like MotionClone or Motion Inversion are not compared. There's also a citation slip in Section 4.1 where MotionDirector is attributed to Tu et al. 2024a instead of Zhao et al. 2023.\n\nNone of this kills the central capability claim—the ablations show both components matter, and the qualitative evidence is strong. But the 'frequency-aware' framing is more of an interpretation than a demonstrated mechanism. For a paper at this venue, that's acceptable if the engineering works; for a journal version, I'd want the frequency claim quantified.\n\nBottom line: send it to review. The ideas are clear, the method is reproducible (code released), and the results are interesting to AIGC and video-generation folks. With a bit more evidence on the mechanism and a stronger evaluation, it could be a solid reference.","headline":"A practical heterogeneous action-transfer method with a genuine RefAdapter trick, but the FAE frequency mechanism is under-evidenced; worth reviewing, not yet fully convincing.","tokens_in":13402,"tokens_out":3331,"would_cite":true,"duration_ms":32294,"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":"FlexiAct transfers actions from a reference video to an arbitrary target image without requiring alignment of layout, shape, or viewpoint, preserving both motion and appearance.","keywords":["action transfer","video customization","image-to-video diffusion","frequency-aware action extraction","attention reweighting","heterogeneous scenarios","RefAdapter","CogVideoX-I2V"],"falsifier":"Run the FAE inference with the bias schedule inverted—high bias in late steps, low in early steps—and measure motion fidelity and appearance consistency on the paper's 250-pair benchmark; if motion fidelity does not drop substantially and appearance does not leak, the timestep-frequency story is not what is carrying the result.","tokens_in":12394,"feed_emoji":"🎬","tokens_out":5273,"duration_ms":46009,"temperature":0.7,"pith_summary":"The paper proposes FlexiAct, a system that takes a reference video and an arbitrary target image and generates a new video in which the target subject performs the reference action, even when the two subjects differ in pose, skeleton, layout, viewpoint, or even species and artistic style. The authors argue this is the first action-transfer framework that does not need pose maps, depth, or matched first frames, and that it does so on a standard image-to-video diffusion backbone with two lightweight additions. If correct, FlexiAct would lower the cost of animation and video customization by replacing motion capture or hand animation with a single reference clip and one image.","feed_headline":"FlexiAct transplants any action onto any subject","feed_subtitle":"Reference video supplies the motion; a single image supplies the identity, with no skeleton or layout matching.","key_machinery":"The mechanism is a timestep-dependent attention bias $W_{\\text{bias}}$ (Eq. 1) applied to the attention weights between video tokens and a learnable frequency-aware embedding during denoising. The embedding is trained on a single reference video with random crops to prevent layout memorization; at inference, the bias is $\\alpha=1$ for $t \\geq 800$, a cosine ramp between $t_h=700$ and $t_l=800$, and zero below $700$, on the assumption that early denoising steps carry low-frequency motion and later steps carry high-frequency appearance. RefAdapter supplies the spatial flexibility by randomly sampling the conditioning frame from the full video during training and replacing the first temporal embedding with the conditioned image, breaking the strict first-frame constraint of I2V models.","core_discovery":"The central claim is that action transfer in heterogeneous scenarios reduces to two problems—spatial structure adaptation and action extraction—and that FlexiAct solves both: RefAdapter conditions generation on an arbitrary frame of the reference video with a small LoRA, enabling the model to adapt motion to new layouts and shapes, and FAE trains per-video frequency-aware embeddings whose attention to video tokens is reweighted by timestep during inference, amplifying motion in early denoising steps. The paper reports that this pipeline outperforms a MotionDirector baseline on CogVideoX-I2V and its own ablations on motion fidelity and appearance consistency, with human raters preferring the outputs.","pith_inferences":["Editorial inference: if the frequency separation is real and robust, the same $W_{\\text{bias}}$ trick could be extended to video editing beyond action transfer—e.g., transferring camera motion or object trajectories from one clip to another, since these are also low-frequency components.","Editorial inference: the hand-set hyperparameters ($t_h=700$, $t_l=800$, $\\alpha=1$) are likely backbone-specific; applying FAE to a different diffusion model would require re-measuring where the motion-to-appearance transition occurs, suggesting a per-model calibration step.","Editorial inference: a cleaner test of the frequency story would be to compute the spectral content of the attention maps directly (e.g., via a spatial or temporal Fourier transform) and correlate it with timestep; the paper's visual inspection of attention maps supports but does not prove the claim."],"forward_implications":["Pose-based action customization becomes unnecessary for many tasks: users can transfer an action from a single clip to any image of a human, animal, or drawn character, without extracting skeletons or depth.","Adding the same attention-bias mechanism to other image-to-video diffusion models could give them heterogeneous action transfer without retraining the base model.","Because FAE requires per-reference-video optimization (1,500 to 3,000 training steps), the method is not feed-forward; a single reference clip costs time to embed, which the authors identify as a limitation.","The 250-pair benchmark with 25 action categories and cross-domain targets gives the community a common test bed, but the paper's reported numbers depend on CLIP-based metrics and preference votes, not on direct action recognition."],"supporting_citations":[{"why":"Supplies the base image-to-video diffusion model (CogVideoX-I2V) with its MMDiT backbone on which RefAdapter and FAE are trained.","marker":"[Yang et al. 2024]"},{"why":"MotionDirector is the global-motion baseline that the paper reimplements on CogVideoX-I2V as MD-I2V for quantitative and qualitative comparison.","marker":"[Zhao et al. 2023]"},{"why":"Miradata provides the 42,000 videos used to train RefAdapter.","marker":"[Ju et al. 2024]"},{"why":"IP-Adapter represents the cross-attention condition-injection approach that RefAdapter contrasts with and draws flexibility from.","marker":"[Ye et al. 2023a]"},{"why":"Genie is the motion-embedding approach that FAE's design initially follows and then modifies with timestep-aware reweighting.","marker":"[Bruce et al. 2024]"},{"why":"Provides the Motion Fidelity metric used to evaluate how closely generated motion matches the reference video.","marker":"[Yatim et al. 2023]"},{"why":"CoTracker supplies the tracklets that the Motion Fidelity metric relies on to compare trajectories in unaligned videos.","marker":"[Karaev et al. 2023]"},{"why":"Textual inversion is the basis for training per-reference embeddings, which FAE adapts from image to video tokens.","marker":"[Gal et al. 2022]"}],"fun_headline_variants":["Action transfer without skeleton or layout limits","Drop the skeleton, keep the motion: FlexiAct","Motion from video, identity from image: FlexiAct","No skeleton match? FlexiAct still transfers any action","Freely transplant actions across scenes with FlexiAct"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the claim that the trained frequency-aware embedding's attention actually separates motion (low frequency) from appearance (high frequency) in the way the timestep schedule assumes, so that adding bias during early steps transfers motion without leaking the reference video's appearance; that claim rests on visual inspection of attention maps and ablations, not on a quantitative frequency analysis.","fun_headline_variants_meta":{"raw":{"variants":["Action transfer without skeleton or layout limits","Drop the skeleton, keep the motion: FlexiAct","Motion from video, identity from image: FlexiAct","No skeleton match? FlexiAct still transfers any action","Freely transplant actions across scenes with FlexiAct"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1227,"prompt_tokens":924,"completion_tokens":303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":540,"tokens_out":303,"duration_ms":3405,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:43:31.689795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the FAE inference with the bias schedule inverted—high bias in late steps, low in early steps—and measure motion fidelity and appearance consistency on the paper's 250-pair benchmark; if motion fidelity does not drop substantially and appearance does not leak, the timestep-frequency story is not what is carrying the result.","supporting_citations":[],"review_version":1}