{"id":"8fddf140-f0d7-4d51-91af-407a723d3b41","arxiv_id":"2508.07603","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LaVieID improves identity-preserving text-to-video by routing local facial parts into early DiT blocks and autoregressively refining denoised video tokens in temporal chunks.","lead":"LaVieID adds a local facial router and a chunk-wise temporal autoregressive module to a diffusion transformer video generator, improving identity preservation when synthesizing videos of a specific person from a single reference photo. The paper reports higher face-similarity and temporal-consistency scores than four open-source identity-preserving video baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No training objective is defined for the temporal autoregressive module: Eq. (6) contains only the diffusion loss and the router loss, so as written the TAM parameters receive no gradient and its claimed benefit is unsupported.","rationale":"The reader's weakest assumption focused on post-hoc bias correction pushing denoised tokens outside the decoder's trained manifold (Eq. 4). That is a legitimate concern, but I find a more fundamental issue: as written, the training objective in Eq. (6) provides no gradient to TAM at all. The diffusion loss is computed on the noise prediction, not on the final refined tokens; the router loss is unrelated to TAM. This omission makes the claimed contribution of TAM unsupported by the paper's own formalism. It is a correctness and reproducibility issue, not an external-consensus disagreement. The released code could resolve the ambiguity by revealing an undocumented training loss; hence conditional acceptance is appropriate. If the code confirms no TAM training, the central claim collapses and the verdict should move to reject. Since the reader and I both target TAM but identify different weaknesses, I marked partial agreement.","tokens_in":14923,"tokens_out":5454,"duration_ms":61808,"concrete_test":"Inspect the released repository (https://github.com/ssugarwh/LaVieID): (1) locate the training loss computation; (2) verify whether TAM parameters are included in the optimizer and receive non-zero gradients; (3) identify the exact loss term that depends on c*_k or b_k. If no such term exists, re-run the w/ TAM ablation with TAM parameters frozen (random init) and compare; if the improvement persists, the Table 2 gains are not due to learned temporal modeling.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The temporal autoregressive module (TAM) is described in §3.3 as taking the denoised latent tokens z0, splitting them into chunks, and producing refined tokens c*_k = c_k + β b_k, with b_k computed using cross-attention conditioned on the previous chunk. The paper states TAM is 'trained' (§4.1: 'fine-tune it alongside the local router and the temporal autoregressive module'), but the only training losses are L_diff (Eq. 1) and L_route (Eq. 5), combined in Eq. (6) as L_total = λ_diff L_diff + λ_route L_route. L_diff is a noise-prediction loss evaluated at arbitrary timesteps t on the DiT output (epsilon_theta), not on the post-denoising enhanced tokens. L_route depends only on router weights w_m and ground-truth masks. Neither term is a function of b_k or c*_k. Therefore, if Eq. (6) is the complete objective, TAM parameters receive zero gradient and the module cannot learn. This is load-bearing because Table 2 attributes substantial gains to TAM (e.g., FaceSim-Curricular 0.312→0.358) and the abstract credits TAM with 'significantly enhancing inter-frame identity consistency.' Without a specified training signal for TAM, the method description is internally inconsistent and the contribution is unsupported. The paper does not mention any additional loss (e.g., reconstruction or perceptual loss on decoded video), so either an undocumented loss exists or TAM is untrained.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LaVieID, an identity-preserving text-to-video framework built on a CogVideoX-style diffusion transformer. It adds two modules: a local router that uses fine-grained facial-component tokens to reweight latent tokens spatially, and a temporal autoregressive module (TAM) that splits denoised latent tokens into chunks and refines each chunk by predicting biases conditioned on the previous chunk. The reported training objective is a diffusion loss plus a router cross-entropy loss. Experiments compare LaVieID against four open-source baselines on a 49-frame 720x480 benchmark, reporting improved FaceSim, subject/background consistency, and user-study identity-similarity scores. Code and models are promised.","tokens_in":15308,"tokens_out":6076,"duration_ms":69341,"significance":"If the central claim holds, the paper offers a simple, low-cost recipe (single A100 GPU, 10K steps) for improving identity preservation in DiT video models. The local-router idea and the post-denoising bias-refinement scheme are interesting and are presented with clear equations and external baselines. However, the manuscript has a load-bearing internal inconsistency: the TAM appears to receive no gradient from the stated training objective. The empirical support also lacks error bars and uses different evaluation subsets across tables. These issues must be addressed before the proposed contributions can be accepted.","major_comments":[{"comment":"The total objective in Eq. (6) contains only L_diff (Eq. 1) and L_route (Eq. 5). L_diff is a noise-prediction loss on ε_θ at arbitrary timestep t, before TAM is applied, and L_route depends only on the router weights w_m. Neither term is a function of the TAM outputs b_k or c*_k defined in Eq. (4). Consequently, if Eq. (6) is the complete training objective, the TAM parameters (ψ and φ) receive zero gradient and cannot be learned. This contradicts the statement in Sec. 3.2 that the local encoder is fine-tuned 'alongside the local router and the temporal autoregressive module' and the Table 2 attribution of gains to TAM (e.g., FaceSim-Curricular 0.312→0.358). Please specify the actual loss used to train TAM (e.g., a reconstruction or perceptual loss on decoded frames), or state clearly that TAM is not trained; in the latter case the performance attribution is unsupported.","section":"Sec. 3.3–3.4, Eq. (6)"},{"comment":"The ablation study uses 'a subset of the test dataset' and reports FID=168.714 for LaVieID, while Table 1 reports FID=174.121 for the same model on the full test set. Because the numbers in Tables 1 and 2 come from different evaluation sets, the ablation cannot be used to claim improvements that transfer to the main comparison. Please report results on the full test set, or at least list the subset size and composition and provide the baselines on the same subset, so readers can calibrate the gains.","section":"Sec. 4.3, Table 2"},{"comment":"All metrics in Table 1 are single-run point estimates with no error bars, confidence intervals, or multi-seed variance. The margins over the strongest baseline ConsisID are small: +0.013 FaceSim-Curricular, +0.012 FaceSim-ArcFace, +0.016 subject consistency, +0.008 background consistency, and the FID is 6.3 worse than CogVideoX+IPA. Without variance estimates or significance tests, the claimed 'state-of-the-art' improvement may be within run-to-run noise. Please add bootstrapped confidence intervals or multiple-seed runs.","section":"Sec. 4.2, Table 1"},{"comment":"The user study uses 30 participants and 50 questions, but no confidence intervals, significance tests, or agreement measures are reported. Since the IDS preference for LaVieID (0.504) is a key subjective claim, please provide at least a binomial confidence interval and a description of how ties and random ordering were handled.","section":"Sec. 4.4"}],"minor_comments":[{"comment":"The notation z0 is used both for the clean latent in Eq. (1) and for the fully denoised latent tokens in Sec. 3.3 (e.g., Eq. (4)). Please disambiguate these two meanings.","section":"Eq. (1) vs Sec. 3.3"},{"comment":"The mapping from the ground-truth segmentation mask y_m to latent-token positions is not described. How are token-level labels obtained for L'=17750 latent tokens? Please clarify.","section":"Sec. 3.2, Eq. (5)"},{"comment":"The claim that TAM 'will not alter the latent space of the baseline DiT' is only true if the decoder is used as-is; adding β b_k changes the input distribution to the decoder. The stability argument would benefit from a sensitivity analysis over β or a distributional check on the corrected latents.","section":"Sec. 3.3, Discussion"},{"comment":"There are duplicated VideoPoet references ([30] and [31]) and some garbled text in the Figure 3 caption; please clean these up before publication.","section":"References and figures"},{"comment":"The text first says the local encoder architecture is adopted from [66] and later says it is fine-tuned. Please state explicitly which parameters are initialized from [66] and which are frozen/unfrozen during training.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The missing TAM training signal is the key technical issue. It may be an omission in the write-up rather than a fundamental flaw, but it must be resolved with a clearly specified training loss or a corrected claim. The empirical gaps (error bars, ablation subset) are manageable revisions. I would not reject outright, but the paper cannot be accepted in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the local-router idea is worth a look, but the temporal autoregressive module (TAM) has no training loss in the paper's equations, and the paper's headline gains rest on it. That is the main thing to know.\n\nWhat is actually new: a local router that weights fine-grained facial component tokens and injects them into the first DiT block, plus post-denoising chunk-wise bias prediction. The local router is a sensible way to reduce feature interference, and the ablation suggests it helps (FaceSim-Curricular from 0.312 to 0.361). The writing is clear, the method is well positioned against prior work, and the main comparison is against external baselines on a shared benchmark, so no circular fitting.\n\nThe soft spot is load-bearing. Equation (6) defines the total loss as λ_diff L_diff + λ_route L_route. L_diff is the diffusion noise-prediction loss on the DiT output; L_route is the router mask loss. Neither depends on the TAM parameters (the bias predictor b_k or the transformer ψ). So as written, TAM receives zero gradient. The text says the TAM is 'fine-tuned', but no loss for it appears anywhere in the paper. This is not a minor omission: Table 2 attributes a substantial gain to TAM (FaceSim-Curricular from 0.312 to 0.358 for TAM alone; 0.428 for the full model), and the abstract credits it with 'significantly enhancing inter-frame identity consistency.' Unless there is an undocumented loss—reconstruction, perceptual, or identity on decoded video—the module cannot learn. The authors need to state the actual objective and re-run the ablations.\n\nOther issues are secondary but real: the ablation FID is computed on a different test subset, so those numbers are not comparable to Table 1; there are no error bars or multi-seed runs; and hyperparameters α, β, K, N are hand-set with no sensitivity analysis. The user study is a fine qualitative check but small.\n\nWho this is for: people working on identity-preserving video generation may find the local router useful and worth citing later. The TAM claim needs fixing before the paper can be taken at face value. This is exactly the kind of internal inconsistency that peer review should catch, so it deserves a serious referee—but not acceptance in current form.","headline":"The local router is a plausible contribution, but the temporal autoregressive module has no training objective in the paper's own equations, so its claimed gains are unsupported as written.","tokens_in":15780,"tokens_out":2165,"would_cite":false,"duration_ms":24710,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Identity-preserving video generation is improved by routing local facial structure into a diffusion transformer and correcting denoised latent tokens with chunk-wise autoregressive biases.","keywords":["identity-preserving video generation","diffusion transformers","text-to-video synthesis","local facial components","temporal autoregressive module","latent bias correction","personalized video generation","spatio-temporal consistency"],"falsifier":"Sweep $\\beta$ in Eq. (4) from 0 to 1 on the same 90-prompt benchmark while watching FaceSim-Curricular and FID together, and rerun the ablation with the route loss removed. If face-identity scores keep climbing as FID worsens, or if identity gains disappear when the segmentation supervision is dropped, the reported advantage depends on correction scale or router supervision rather than being a stable property of the architecture.","tokens_in":14833,"feed_emoji":"🎬","tokens_out":9039,"duration_ms":102922,"temperature":0.7,"pith_summary":"This paper tries to establish that identity loss in diffusion-transformer video generation is not inevitable: with two targeted additions, a standard text-to-video diffusion transformer can keep a single person's face recognizable across frames while following a text prompt. The first addition, a local router, makes the model attend to fine facial parts (eyebrows, eyes, mouth, nose, skin, hair) rather than only global tokens, reweighting latent tokens by what those parts look like. The second, a temporal autoregressive module, splits fully denoised latent tokens into chunks and predicts a small bias for each chunk from the previous chunk, making frame-to-frame identity relationships explicit after denoising. On the benchmark used in this work, the full model reports the highest face-similarity, subject-consistency, background-consistency, and user-study identity scores among the compared methods. The practical stake is that identity-preserving video may need structured spatial and temporal corrections on top of an existing diffusion transformer, rather than a new generative paradigm.","feed_headline":"Local router and temporal biases keep a face consistent across video frames","feed_subtitle":"On the paper's benchmark, face-similarity and consistency scores beat four prior identity-preserving video generators.","key_machinery":"The paper's argument is carried by two mechanisms. The local router is a learned weighting layer that routes latent tokens through local facial component tokens, expressing global face identity as a weighted combination of eyebrows, eyes, mouth, nose, skin, and hair; it is supervised by a segmentation-mask cross-entropy loss. The temporal autoregressive module is a post-denoising bias predictor: it groups latent tokens into temporal chunks and applies $c_k^* = c_k + \\beta b_k$, with $b_k$ produced by cross-attention over the previously enhanced chunk; causal self-attention with rotary position embeddings and teacher forcing keep the refinement temporally ordered. Their work is to add spatial","core_discovery":"The central claim is that two modules, used together on a CogVideoX-style diffusion transformer, overcome the identity loss caused by global unstructured attention. The local router injects spatial structure: it extracts six facial components with an off-the-shelf segmentation network, encodes them into token sequences, computes per-token weights $w_m$ that score how much each component should influence the joint latent tokens, and refines the latents as $z^* = z + \\alpha \\sum_m w_m^\\top \\odot \\varphi(l_m, z)$, where $\\varphi$ is a cross-attention reconstruction using local facial tokens as keys and values. A route loss supervises these weights with the ground-truth segmentation mask, so the","pith_inferences":["The paper does not report how sensitive the result is to $\\beta$ or how the corrected latents compare with the decoder's training distribution; a beta sweep with FID and FaceSim would show whether the identity gain is a genuine latent-space improvement or a metric-side effect.","Because the router's supervision signal comes from a face-segmentation network, the method inherits that network's limits: on occluded, heavily angled, or stylized faces where segmentation fails, identity preservation should degrade, and removing the segmentation loss would isolate that dependency.","The benchmark uses 30 evaluation subjects and 90 prompts; the claim of state-of-the-art identity preservation should be read as demonstrated on that distribution until tested across more identities, longer videos, and more varied motion.","The chunk-wise bias formulation suggests a direct extension to longer videos by keeping the chunk architecture and increasing the number of chunks, but then bias corrections accumulate, so the stability of teacher forcing under longer rollout remains an open test."],"forward_implications":["With the local router alone, FaceSim-Curricular/ArcFace rise from 0.312/0.291 to 0.361/0.334 on the ablation subset; with the temporal module alone they reach 0.358/0.327; together they reach 0.428/0.398.","In the user study, LaVieID wins the largest share of preference votes on all four criteria, with identity-similarity preference 0.504 versus 0.155 for the next-best identity baseline.","The temporal module changes motion behavior: it produces a 'softly gazes downward' action that the baseline and the local-router-only variant miss, suggesting text can drive subject motion when temporal dependencies are explicit.","Because the autoregressive module acts on denoised latents without modifying the base model's latent space, the whole enhancement can be trained on a single GPU (about 60 hours for 10K steps), making it a cheap plug-in rather than a full retrain.","The best FID is still held by CogVideoX+IPA (167.117 vs 174.121), which the authors attribute to identity-preserving generation being an out-of-distribution task; LaVieID trades a little raw realism for much higher identity fidelity."],"supporting_citations":[{"why":"Supplies the baseline diffusion transformer and latent space that LaVieID extends with the local router and temporal module.","marker":"[60]"},{"why":"Provides the main baseline, the training/evaluation dataset, and the local facial encoder architecture reused by LaVieID.","marker":"[66]"},{"why":"Off-the-shelf segmentation network that extracts the six facial components feeding the local router.","marker":"[63]"},{"why":"Supplies the FaceSim-Curricular and FaceSim-ArcFace identity-similarity metrics used to evaluate the central claim.","marker":"[8]"},{"why":"Supplies the subject and background consistency metrics used for temporal quality evaluation.","marker":"[28]"},{"why":"One of the compared state-of-the-art baselines for identity-preserving video generation.","marker":"[17]"},{"why":"One of the compared state-of-the-art baselines; LaVieID is positioned against its cross-video-pairing approach.","marker":"[70]"},{"why":"Part of the CogVideoX+IPA baseline that measures how well merely adding an image adapter preserves identity.","marker":"[61]"}],"fun_headline_variants":["Local router and temporal biases keep faces consistent","Autoregressive token chunks fix identity drift in video","Spatial local router plus autoregressive temporal bias stabilizes identity","Local facial structure prevents identity loss in generated video","Face-consistent video via local routers and temporal autoregression"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes the numeric biases it adds to the denoised video tokens keep those tokens inside the distribution the video decoder was trained to turn into images, so visual quality is not silently traded for higher face-similarity scores.","fun_headline_variants_meta":{"raw":{"variants":["Local router and temporal biases keep faces consistent","Autoregressive token chunks fix identity drift in video","Spatial local router plus autoregressive temporal bias stabilizes identity","Local facial structure prevents identity loss in generated video","Face-consistent video via local routers and temporal autoregression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2575,"prompt_tokens":758,"completion_tokens":1817,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1740}},"tokens_in":502,"tokens_out":1817,"duration_ms":14622,"temperature":1.0,"reasoning_tokens":1740,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:59:09.484472+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep $\\beta$ in Eq. (4) from 0 to 1 on the same 90-prompt benchmark while watching FaceSim-Curricular and FID together, and rerun the ablation with the route loss removed. If face-identity scores keep climbing as FID worsens, or if identity gains disappear when the segmentation supervision is dropped, the reported advantage depends on correction scale or router supervision rather than being a stable property of the architecture.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the main baseline, the training/evaluation dataset, and the local facial encoder architecture reused by LaVieID."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Off-the-shelf segmentation network that extracts the six facial components feeding the local router."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FaceSim-Curricular and FaceSim-ArcFace identity-similarity metrics used to evaluate the central claim."}],"review_version":1}