{"id":"1c24d2c0-a6c5-4867-8548-fdacf89e0ef0","arxiv_id":"2508.09959","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding an L1 sparsity penalty to the motion dictionary of the LIA portrait animator produces disentangled, human-interpretable motion vectors that support controllable image and video editing and scale to roughly one billion parameters.","lead":"LIA-X is an autoencoder that animates a still portrait using a driver video, and adds a sparsity penalty to its motion dictionary so individual facial movements (smile, yaw, blink) can be adjusted separately. The paper shows the model scales to about one billion parameters and reports better or comparable scores than existing portrait animators on several benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'edit-warp-render' advantage rests on an emergent, untrained linear-edit property (Eq. 11) that is neither quantitatively validated nor fully specified; without it, the central interpretability claim is unsupported.","rationale":"The paper's novel selling point is not the autoencoder backbone (largely LIA) but the Sparse Motion Dictionary and its use for 'edit-warp-render.' The Sparse Motion Dictionary is presented as the mechanism producing interpretable vectors; interpretability is then operationalized by Eq. 11. I read Section 5.2 as the only evidence for this operationalization. It is anecdotal and cherry-picked: 'almost all activated vectors correspond to human-understandable meanings' with a_i chosen by hand. No quantitative metric establishes that the latent code manipulation is causally linked to the intended semantic change, is monotonic in a_i, or transfers to unseen identities. This is a load-bearing gap because Eq. 10's edit step is precisely what differentiates LIA-X from LIA; if editing does not generalize, the method reduces to a scaled-up LIA with an L1 penalty, and the claimed advantage on large-pose cross-reenactment (where editing is supposed to align source and driving) is unsupported. I also note the quantitative comparisons are confounded: LIA-X trains on a much larger, partly internal dataset, so Table 1/2 improvements may reflect scale rather than sparsity/interpretability. The reader's verdict of CONDITIONAL is appropriate; the authors should release code/data or run the proposed edit-validation study. No internal contradiction in the equations was evident, so I do not escalate to REJECT, but the central contribution is not yet demonstrated.","tokens_in":11265,"tokens_out":12381,"duration_ms":129950,"concrete_test":"Conduct a controlled evaluation on 100 held-out AAHQ/VoxCelebHQ portraits: for each dictionary vector that the paper labels as controlling a semantic (e.g., 'smile', 'mouth open', 'yaw'), render edited images x_{a} = G_r(T(G_f(z_{s->s}+a·d_i), x_s)) for a in {-0.5,-0.4,...,0.5} and measure the corresponding AU/pose (e.g., OpenFace) as a function of a. Accept the editing claim only if the measurement changes monotonically in the expected direction for at least 80% of sources and the edit direction is consistent. Additionally, on a subset verify that z_{s->E(s)} equals E(x_{E(s)})+w_{r->E(s)}—i.e., re-encode the rendered edited image—otherwise Eq. 10 is not a valid animation of the edited portrait.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the Sparse Motion Dictionary supports interpretable, controllable editing depends on Eq. 11, where adding a_i d_i to the source's self-reconstruction code z_{s->s} is assumed to edit pose/expression. This property is never trained or measured. The network is trained only with reconstruction losses plus L1 on coefficients (Eq. 9), so nothing enforces that the latent direction d_i corresponds to a semantically consistent image change, nor that the effect transfers across identities. Section 5.2 shows only qualitative, post-hoc examples with manually chosen a_i values; there is no attribute-level accuracy, monotonicity, or cross-identity consistency metric. Moreover, Eq. 10's edit-warp-render formula silently substitutes z_{s->E(s)} (a code from the original source) for the self-reconstruction code of the edited source; it is not specified how x_{E(s)} is rendered or whether this shortcut equals animating a truly edited portrait. If the linear-add assumption fails or is identity-specific, the key advantage over LIA is unsupported, and the reported cross-reenactment gains in Table 2 (which use ID similarity and image quality, not semantic correctness) cannot be attributed to editing. The sparse penalty itself might still improve reconstruction via regularization, but the interpretable-editing story would be unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LIA-X, a self-supervised autoencoder for portrait animation built on LIA's linear latent navigation. The main novelty is a 'Sparse Motion Dictionary': an L1 penalty on the motion coefficients in the training loss (Eq. 9) is claimed to produce semantically disentangled, interpretable motion vectors. At inference, users can 'edit' a source portrait by adding a scaled dictionary vector to its self-reconstruction code (Eq. 11), then animate with an 'edit-warp-render' formula (Eq. 10). The paper reports state-of-the-art quantitative results on self-reenactment (Table 1) and cross-reenactment (Table 2), a scalability study up to 0.9B parameters (Tables 3-4), and qualitative demonstrations of editing and 3D-aware manipulation.","tokens_in":11681,"tokens_out":5363,"duration_ms":63883,"significance":"If the interpretability and controllability claims hold, LIA-X would be a meaningful step toward fast, controllable portrait animation without explicit 3D representations. The sparse-dictionary idea is simple and potentially useful, and the large-scale training study is a concrete engineering contribution. However, the central evidence is observational: the editing operation in Eq. (11) is never trained for or quantitatively measured, no controlled ablations isolate the effect of the sparsity penalty, and the quantitative comparisons are confounded by substantially larger model and dataset scale. The significance is therefore currently conditional rather than demonstrated.","major_comments":[{"comment":"The central editing operation is neither trained for nor quantitatively validated. The training loss (Eq. 9) supervises reconstruction of x_{s->d} and adds L1 on coefficients; nothing enforces that z_{s->s}+a_i d_i changes only the named semantic attribute, nor that the edit transfers across identities. The paper states 'Surprisingly, we found...' and shows selected examples. This is post-hoc observational evidence. Please add quantitative tests: attribute classifiers (pose angles, mouth/eye openness) on edited outputs, monotonicity of attribute magnitude with a_i, cross-identity consistency, and negative controls (random directions, a dense dictionary, or random sparse directions). Without such tests, the 'interpretable Sparse Motion Dictionary' claim is unsupported.","section":"Section 5.2, Eq. (11)"},{"comment":"The edit-warp-render formula is under-specified. z_{s->E(s)} is defined via Eq. (11), but no edited image x_{E(s)} is ever rendered or used as an actual source. Eq. (10) composes the original source's self-reconstruction code with the original driving motion difference. Please specify: (i) how x_{E(s)} is obtained, (ii) how a_i is chosen automatically or by the user to match the driving frame, and (iii) why substituting z_{s->E(s)} for z_{s->s} is equivalent to animating a genuinely edited portrait. If this is an inference-time heuristic, that should be stated and validated separately from the reconstruction training.","section":"Section 4.3, Eq. (10)"},{"comment":"The reported comparisons are confounded by training scale and data. LIA-X is trained on 0.5M sequences / 94M frames / 55k identities with up to 0.9B parameters, whereas the baselines use smaller models and datasets. The wins do not isolate the sparse dictionary or the edit-warp-render mechanism. Provide ablations with matched data and model size: LIA-X with and without the sparsity penalty, with a dense dictionary, and with editing disabled. Also report confidence intervals or repeated-run variability. Table 2's Identity Similarity and no-reference Image Quality metrics do not measure semantic correctness of motion transfer; include pose/expression alignment metrics or a user study to attribute the gain to editing.","section":"Tables 1 and 2"},{"comment":"Sparsity is demonstrated on only two videos and with no quantitative metric (e.g., fraction of active coefficients, entropy, dictionary utilization). More importantly, coefficient sparsity is neither necessary nor sufficient for semantic disentanglement; a random sparse dictionary can also activate few vectors. Please report dictionary statistics over the validation set and show that each vector corresponds to the same semantic attribute across many identities and source images, rather than only the curated examples in Figs. 4-6. The labels attached to motion vectors (yaw, pitch, smile, etc.) appear to be assigned post hoc to vectors selected for illustration.","section":"Section 5.1, Fig. 3"}],"minor_comments":[{"comment":"Typo: 'enrouraging' should be 'encouraging'.","section":"Figure 2 caption"},{"comment":"The perceptual loss weight is written λ without a subscript, while Eq. (9) uses λ1 and λ2. Align the notation.","section":"Eq. (5)"},{"comment":"The perturbation range a_i ∈ [-0.5, 0.5] is given without context: no typical coefficient magnitudes or normalization of dictionary vectors are reported, so 'small perturbation' is undefined.","section":"Section 5.2"},{"comment":"The improvements from Middle (0.3B) to Large (0.9B) are within 0.01 LPIPS and 0.08 PSNR on VoxCelebHQ; the statement that these results 'clearly demonstrate' the effectiveness of scaling is overstated without error bars or statistical testing.","section":"Tables 3 and 4"},{"comment":"The text says 'To ensure a fair comparison, we train LIA-X at two different resolutions,' but LIA-X is also trained on substantially more data and with much larger capacity than the baselines. A fair comparison is not established by matching resolution alone.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eq. (11) lands: the interpretability and editing claims are the load-bearing contribution, and they currently rest on qualitative, post-hoc examples. The paper may be salvageable with a quantitative evaluation of the editing property, controlled ablations for the sparsity penalty, and matched-scale comparisons; without those, the central claim is not established. I would encourage the editor to request such an experiment rather than reject outright, since the idea is simple and the large-scale training results are potentially useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nLIA-X takes LIA and adds an L1 sparsity penalty on the motion coefficients. That is a small but real change, and it does visibly sparsify the active dictionary per frame (Fig. 3). The new inference-time trick is \"edit-warp-render\": using a single learned motion vector to edit the source portrait before animation. They also scale the autoencoder to ~1B parameters and show some payoff.\n\nWhat's genuinely good: the sparse dictionary does seem to reduce the number of active coefficients, the editing demos for yaw/pitch/roll and mouth/eyes are visually convincing, and the scaling tables are honest enough to note that gains flatten from 0.3B to 0.9B. The limitations section is candid about fixed resolution and conv architecture limits.\n\nThe soft spots are substantial. First, the central interpretability claim rests entirely on qualitative examples. Eq. 11 treats z_{s->E(s)} = z_{s->s} + a_i d_i as if adding a scaled dictionary vector reliably edits pose or expression, but nothing in the training objective enforces that. There is no attribute-level accuracy, no monotonicity test, no cross-identity consistency check, no quantitative disentanglement metric, and no user study. The stress-test note is right: if that linear-add property fails on arbitrary sources, the whole edit-warp-render advantage is unsupported. The paper is also vague about exactly how the edited source is rendered in Eq. 10, which compounds the concern. Second, the comparisons are confounded: LIA-X trains on ~94M frames including an undocumented internal dataset, while the baselines appear to use their standard smaller training sets. Tables 1 and 2 are not apples-to-apples. Third, the internal dataset is a reproducibility blocker.\n\nThat said, the paper does not overclaim everywhere. It calls the interpretability \"surprising\" and frames the editing as discovered rather than trained, and it explicitly admits the 0.9B gains are small. The core architecture is coherent and the sparsity effect is plausible. The paper deserves a serious referee, but it needs major revision: release code and data details, retrain or at least report baselines under matched data and compute, and add quantitative evaluations of the editing operation. If the linear-edit property holds up, this is a useful practical tool; right now the evidence is mostly a demo.\n\nMy recommendation: send it to peer review, but expect heavy revision on the evaluation side. A careful referee could make this much better.","headline":"LIA-X is a simple, plausible sparsity tweak on LIA; the interpretability story is supported mostly by cherry-picked visuals and the SOTA comparisons are confounded by much larger training data.","tokens_in":12107,"tokens_out":2998,"would_cite":false,"duration_ms":31746,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LIA-X makes portrait animation controllable by learning a sparse motion dictionary whose vectors are human-interpretable, enabling an 'edit-warp-render' pipeline that improves reenactment quality.","keywords":["portrait animation","latent image animator","sparse motion dictionary","interpretable representations","image editing","video editing","3D-aware manipulation","scalability"],"falsifier":"A quantitative editing fidelity test would settle the claim: apply Eq. (11) with a 'smile' vector to a large, diverse set of neutral-expression source portraits, then measure smile intensity with a pretrained facial attribute classifier. If the edit succeeds for only a small fraction of identities or produces inconsistent magnitudes, the linear editing assumption fails. Alternatively, compare animation quality (PSNR, LPIPS, identity similarity) with and without the pre-editing step across many source-driving pairs with large pose/expression differences; if the improvement is not consistent and","tokens_in":1758,"feed_emoji":"🎭","tokens_out":6307,"duration_ms":92694,"temperature":0.7,"pith_summary":"LIA-X is an autoencoder-based portrait animator that models facial motion as linear navigation through a learned sparse motion dictionary. The paper argues that adding an L1 sparsity penalty to the motion coefficients forces each reconstruction to use few motion vectors, and that the resulting vectors become interpretable, corresponding to semantic factors like head yaw, pitch, roll, mouth, eyes, and eyebrows. This interpretability supports a novel edit-warp-render strategy: before transferring motion from a driving video, the user can edit the source portrait using the learned vectors to align its pose and expression with the driving frame. The paper claims this reduces the large pose/expression gap that degrades prior warp-render methods, and that LIA-X outperforms state-of-the-art approaches on self-reenactment and cross-reenactment benchmarks. It also demonstrates that the architecture scales to about one billion parameters with improved performance.","feed_headline":"Sparse motion dictionary makes portrait animation editable","feed_subtitle":"LIA-X learns interpretable motion vectors, so users can adjust pose and expression before animating — and reenactment improves.","key_machinery":"The Sparse Motion Dictionary $D_m = \\{d_1, \\dots, d_M\\}$ is a set of orthogonal motion vectors in latent space, with an L1 sparsity penalty $S(A_{r\\to d})$ on the coefficients used to reconstruct each image. The sparsity constraint is the central mechanism: it forces the network to use a minimal set of motion vectors for each reconstruction, which empirically disentangles facial dynamics into interpretable factors. The linear editing operation $z_{s\\to E(s)} = z_{s\\to s} + a_i d_i$ then leverages these vectors for direct, user-controllable manipulation of the source portrait before animation.","core_discovery":"LIA-X discovers that applying an L1 sparsity constraint to the motion coefficient vector $A_{r\\to d}$ in a latent autoencoder produces a Sparse Motion Dictionary whose individual vectors are interpretable. During self-supervised training, this penalty encourages the network to reconstruct each frame by activating only a few motion vectors; at inference, these vectors can be manipulated directly. Editing the source portrait is formulated as $z_{s\\to E(s)} = z_{s\\to s} + a_i d_i$, adding a scaled learned motion vector to the source's own reconstruction code. This edit is then followed by the standard warp-render animation step, yielding the 'edit-warp-render' pipeline. The paper claims this al","pith_inferences":["The linear editing property of Eq. (11) may generalize beyond the qualitative examples: if the motion dictionary is truly interpretable, then composing vectors (e.g., 'smile' + 'yaw') should produce combined edits, and editing should work consistently across arbitrary source identities. This is testable with attribute classifiers on a large set of portraits.","The sparsity-induced interpretability might extend to other video-to-video tasks such as full-body animation or speech-driven avatars, where a sparse motion dictionary could reveal interpretable control axes for body pose or lip sync.","The diminishing returns from 0.3B to 0.9B parameters suggest the model is dataset-limited rather than architecture-limited; training on larger, more diverse datasets may unlock further gains, an explicit hypothesis the paper leaves for future work."],"forward_implications":["Portrait animation becomes a two-stage process: first edit the source to match the driving frame's pose and expression, then warp-render; this narrows the initial gap that degrades prior methods.","Users can perform fine-grained image editing (e.g., open/close eyes, smile, frown) by adjusting individual motion vectors, and video editing by applying the same linear edit to a video's first frame and then animating.","The model exhibits 3D-aware manipulation of yaw, pitch, and roll without any explicit 3D representation, since these transforms emerge as interpretable motion vectors.","Scaling the autoencoder from 0.05B to 0.9B parameters monotonically improves reconstruction metrics on VoxCelebHQ and TalkingHead-1KH, though gains diminish at the largest size.","Autoencoder-based latent navigation provides faster inference than diffusion-based animators, making LIA-X a practical complement for controllable generation."],"supporting_citations":[{"why":"Latent Image Animator: the original latent navigation model that LIA-X extends with the sparse dictionary and edit-warp-render strategy.","marker":"[45]"},{"why":"LIA journal version and the VoxCelebHQ dataset used for training and self-reenactment evaluation.","marker":"[48]"},{"why":"Sparse dictionary coding, the conceptual inspiration for the Sparse Motion Dictionary and its L1 penalty.","marker":"[24]"},{"why":"TalkingHead-1KH dataset used for training and self-reenactment evaluation.","marker":"[40]"},{"why":"HDTF dataset, from which driving videos are selected for the cross-reenactment evaluation.","marker":"[57]"},{"why":"FOMM, a first-order motion model baseline that LIA-X compares against in self-reenactment.","marker":"[28]"},{"why":"LivePortrait, a state-of-the-art keypoint-based animator used as a comparison baseline.","marker":"[11]"},{"why":"X-Portrait, a diffusion-based animator used as a comparison baseline, highlighting speed and quality trade-offs.","marker":"[49]"}],"fun_headline_variants":["Sparse motion vectors make portrait animation editable","LIA-X: interpretable motion codes for precise facial control","Edit-warp-render: controllable portrait reenactment","Sparse motion dictionary enables fine-grained expression editing","Face animation with interpretable sparse motion codes"],"cache_read_input_tokens":13952,"weakest_assumption_plain":"The edit-warp-render advantage rests on the assumption that simply adding a scaled learned motion vector to the source's own reconstruction code (Eq. 11) reliably edits the portrait's expression or pose; this linear editing property is never trained for or quantitatively evaluated, and the examples shown were selected post hoc from vectors that happened to work.","fun_headline_variants_meta":{"raw":{"variants":["Sparse motion vectors make portrait animation editable","LIA-X: interpretable motion codes for precise facial control","Edit-warp-render: controllable portrait reenactment","Sparse motion dictionary enables fine-grained expression editing","Face animation with interpretable sparse motion codes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1265,"prompt_tokens":745,"completion_tokens":520,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":445}},"tokens_in":489,"tokens_out":520,"duration_ms":6774,"temperature":1.0,"reasoning_tokens":445,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:40:33.760892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A quantitative editing fidelity test would settle the claim: apply Eq. (11) with a 'smile' vector to a large, diverse set of neutral-expression source portraits, then measure smile intensity with a pretrained facial attribute classifier. If the edit succeeds for only a small fraction of identities or produces inconsistent magnitudes, the linear editing assumption fails. Alternatively, compare animation quality (PSNR, LPIPS, identity similarity) with and without the pre-editing step across many source-driving pairs with large pose/expression differences; if the improvement is not consistent and","supporting_citations":[{"cited_title":"Latent image animator: Learning to animate im- ages via latent space navigation","cited_arxiv_id":null,"evidence_quote":"Latent Image Animator: the original latent navigation model that LIA-X extends with the sparse dictionary and edit-warp-render strategy."},{"cited_title":"Lia: Latent image animator","cited_arxiv_id":null,"evidence_quote":"LIA journal version and the VoxCelebHQ dataset used for training and self-reenactment evaluation."},{"cited_title":"Emergence of simple- cell receptive field properties by learning a sparse code for natural images","cited_arxiv_id":null,"evidence_quote":"Sparse dictionary coding, the conceptual inspiration for the Sparse Motion Dictionary and its L1 penalty."},{"cited_title":"One-shot free-view neural talking-head synthesis for video conferenc- ing","cited_arxiv_id":null,"evidence_quote":"TalkingHead-1KH dataset used for training and self-reenactment evaluation."},{"cited_title":"Flow-guided one-shot talking face generation with a high- resolution audio-visual dataset","cited_arxiv_id":null,"evidence_quote":"HDTF dataset, from which driving videos are selected for the cross-reenactment evaluation."},{"cited_title":"First order motion model for image animation","cited_arxiv_id":null,"evidence_quote":"FOMM, a first-order motion model baseline that LIA-X compares against in self-reenactment."},{"cited_title":"X-portrait: Expressive portrait anima- tion with hierarchical motion attention","cited_arxiv_id":null,"evidence_quote":"X-Portrait, a diffusion-based animator used as a comparison baseline, highlighting speed and quality trade-offs."}],"review_version":1}