{"id":"20cb37ba-86c8-4089-a62f-78c63b7f0cde","arxiv_id":"2507.12761","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage text-guidance framework, Think-Before-Draw, uses chain-of-thought prompting to convert emotion labels into facial muscle descriptions and then progressively guides a diffusion model from coarse emotion to fine muscle control.","lead":"This paper builds a talking-head video generator that turns emotion labels into detailed facial muscle descriptions using a step-by-step reasoning language model, then uses those descriptions to control a diffusion model in two stages. It reports better image quality, expression control, and motion scores than several prior methods on the MEAD and HDTF benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 SOTA claim rests on undocumented evaluation protocols: the curated MEAD split, missing HDTF split, and absent baseline re-runs could explain the 7 dB PSNR gap.","rationale":"The reader's weakest assumption — that the Table 1 benchmark numbers were produced under comparable protocols — is exactly the load-bearing point I would attack. The paper's own text confirms the concern: Section 4.1 mentions a 'curated version of MEAD' with no curation details, and the HDTF description defines training data but not a test split. The extreme MEAD PSNR gap is a quantitative red flag that cannot be explained by the proposed CoT-FA and progressive guidance modules alone, since the ablation gains on other metrics are modest. The internal evidence for the method's value is reasonable: the ablation table shows consistent improvements from CoT-FA and the PGD strategy, and the qualitative examples support the narrative. However, the central SOTA claim is an external comparison claim, and external comparisons require controlled protocols. The paper does not supply those protocols, so the claim is conditionally supported at best. My recommendation is to keep the reader's CONDITIONAL verdict: the method may be good, but the headline claim should not be accepted until the evaluation is audited. I agree with the reader's identification of the same weakest assumption, and the concrete test of re-running baselines on the identical split directly settles whether the concern lands.","tokens_in":14262,"tokens_out":3617,"duration_ms":41736,"concrete_test":"Ask the authors to release the exact MEAD and HDTF test video/identity lists and preprocessing details (resolution, cropping, frame count, audio sampling), then run the released checkpoints of StyleTalk, EAT, EAMM, and SadTalker on that identical split. Specifically, recompute StyleTalk's MEAD PSNR using the submitted test set after the same manual curation filtering is applied; if its PSNR rises from 21.60 toward TBD's 28.64, the reported advantage is protocol-driven, not method-driven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TBD achieves state-of-the-art performance on MEAD and HDTF depends on the numbers in Table 1 being produced under comparable evaluation conditions. Section 4.1 states the model is trained on 'a curated version of the MEAD dataset' without specifying how the curation was performed, and gives only an 80/20 train/test split without saying whether the split is by identity or by video clip. For HDTF, the text says 20 subjects with 10 clips each were 'selected' for training, but no test split is defined; if the reported HDTF metrics were computed on the same clips used for training, those numbers are not valid benchmark results. No statement is made that any baseline was re-run on the exact same split with identical preprocessing, so the comparison is not controlled. The MEAD PSNR gap of 7.04 dB (TBD 28.64 vs StyleTalk 21.60) is far larger than the improvements in the ablation study and is consistent with different test sets, different video cropping/resolution, or identity leakage. Section 3.2 also states 'Each data entry undergoes manual review to filter out anomalous descriptions,' which could remove hard test samples and further inflate scores. Without the exact test identities, test videos, and baseline re-runs on that split, the SOTA claim is unverifiable and could be an artifact of protocol differences rather than the proposed method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Think-Before-Draw (TBD), a text-driven emotional talking-head generation framework with two main components: a Chain-of-Thought Facial Animation (CoT-FA) module that converts abstract emotion labels into coarse- and fine-grained text descriptions grounded in FACS/AU terminology, and a progressive guidance denoising strategy that first applies coarse text embeddings in early denoising steps and fine-grained embeddings in later steps, with an alpha-blended transition. Experiments on MEAD and HDTF report large quantitative gains over several audio-, video-, and label-driven baselines across PSNR, SSIM, FID, CPBD, and SyncConf, supplemented by qualitative comparisons, an ablation study, t-SNE visualizations, and a small user study. The central claim is state-of-the-art performance on widely used benchmarks, including MEAD and HDTF.","tokens_in":14542,"tokens_out":4784,"duration_ms":52094,"significance":"If the reported results are reproducible under controlled evaluation protocols, the framework is a meaningful step forward: it replaces discrete emotion labels with anatomically grounded, FACS-aligned text descriptions, provides an intuitive coarse-to-fine conditioning schedule for diffusion-based talking-head generation, and demonstrates fine-grained intensity control. The t-SNE style-code analysis and zero-shot portrait evaluations are useful additions. However, the paper does not release code, exact data splits, or evaluation scripts, and the quantitative evidence currently lacks the protocol transparency needed to support the state-of-the-art claim. The contribution is potentially valuable, but the verification burden is not yet met.","major_comments":[{"comment":"The central state-of-the-art claim is not verifiable because the evaluation protocol is underspecified. The MEAD split is described only as \"a curated version\" with an 80/20 ratio; there is no statement of whether the split is by identity or by video clip, no description of the curation, and no confirmation that baselines were re-run on the exact same split with identical preprocessing. For HDTF, 20 subjects with 10 clips each are \"selected\" for training, but no test split is defined; if the reported HDTF metrics were computed on clips also used for training, those numbers are not valid benchmark results. The 7.04 dB MEAD PSNR gap over StyleTalk (28.64 vs 21.60) is far larger than the improvements shown in the ablation study and is consistent with protocol differences rather than a genuine method advantage. The authors must specify the exact test identities and videos, describe the curation, and re-run all baselines under identical conditions.","section":"§4.1 (Datasets) and Table 1"},{"comment":"No error bars, standard deviations, or statistical significance tests are reported for any quantitative metric, so it is impossible to assess whether the observed gaps are meaningful. Moreover, the internal numbers are inconsistent: the full model's FID in Table 2 is 16.725, while Table 1 reports TBD's MEAD FID as 17.28 for what appears to be the same configuration. This discrepancy suggests either different evaluation sets or a reporting error. The authors should report mean and standard deviation over multiple runs or seeds, and clarify the relationship between the evaluation sets used in Tables 1 and 2.","section":"§4.2, Table 1 and §4.4, Table 2"},{"comment":"The statement \"Each data entry undergoes manual review to filter out anomalous descriptions\" introduces a selection step that can bias the evaluation. The paper does not specify whether this filtering occurred before the train/test split or whether the test set was frozen before manual review. If anomalous or hard test examples were removed from the test set, the reported scores are inflated. The authors should clarify the temporal ordering of filtering and splitting, and ideally evaluate on a fixed test set that is not subject to manual filtering.","section":"§3.2 (Step-IV)"},{"comment":"The quantitative comparison omits the closest text-driven fine-grained baselines, including FG-EmoTalk [14], ExpCLIP [15], and Expressive Talking Avatars [16], all of which are cited in the related work. Since the paper's contribution is specifically fine-grained text-driven control, comparing only against EAT as a text-driven baseline does not support the claim of state-of-the-art performance over text-driven methods. These baselines should be included in the quantitative comparison, or a clear justification for their exclusion should be provided.","section":"§2.1 and §4.1 (Comparison setting)"}],"minor_comments":[{"comment":"There are typos such as \"vitrual\" instead of \"virtual\" and grammatical issues such as \"a effective solution\"; these should be corrected.","section":"§1"},{"comment":"The stage-split notation is ambiguous: s1 = t * 0.4 uses t both as the total number of sampling steps and as the current timestep, and \"P si = t\" should be a summation. Algorithm 2 would also benefit from clearer naming and a precise definition of the stage boundaries.","section":"§3.4"},{"comment":"The baseline \"ETA\" should be \"EAT\" to match the method name used elsewhere in the paper.","section":"Table 3"},{"comment":"The subfigure captions appear to be swapped relative to the text: the text describes (a) as identity features from one speaker and (b) as expression features, while the caption labels (a) as emotion-related clustering and (b) as identity-related dispersion.","section":"Figure 11"},{"comment":"Hallo and Sonic are discussed as qualitative comparisons only, yet they are not listed in the quantitative comparison table; the paper should state explicitly that these methods are evaluated qualitatively and not on the same benchmark.","section":"§4.3"},{"comment":"The user study is based on only 10 evaluators and 10 samples per condition; the authors should report per-item variance, confidence intervals, or a statistical test to support the MOS comparisons.","section":"§4.6"},{"comment":"References [25] and [40] both cite the Hallo paper; one duplicate should be removed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to become acceptable after a major revision that pins down the evaluation protocol. I would ask the editor to require exact test splits, a description of the MEAD curation, and baseline re-runs on the same split before the quantitative claims can be checked. The omission of the closest text-driven baselines should also be addressed; the current comparison set does not support the abstract's state-of-the-art claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering contribution. The core idea—using a four-step Chain-of-Thought prompt to decompose emotion labels into FACS-based muscle descriptions, then feeding coarse and fine text into different denoising stages—is a sensible recombination of known pieces, and the specific combination for talking heads looks new. The ablation supports both components: descriptions help, CoT helps more, and the progressive guidance adds a further gain. The intensity-control demo is a nice touch. I give credit for the clear presentation and for grounding the prompts in anatomy rather than just label-to-text mapping.\n\nThe soft spots are real, and they cluster in the evaluation. Table 1's SOTA claim depends on protocol comparability that is never established. MEAD is \"curated\" with no description of how; the 80/20 split is not characterized by identity or clip. For HDTF, the paper says 20 subjects with 10 clips each were selected for training, but no test split is defined. If the reported HDTF numbers include training clips, they are not benchmark numbers. Baselines are not re-run on the same split, and the 7 dB PSNR gap over StyleTalk is large enough that different cropping, resolution, or test identities could explain it. The statement in Section 3.2 that each data entry undergoes manual review to filter out anomalous descriptions is another selection step that could inflate scores. Also missing from the comparison are FG-EmoTalk and ExpCLIP, both cited but not tested, despite being the closest text-driven baselines.\n\nThe ablation is internally consistent, but there are no error bars or significance tests, so the reported gains could be noise. No code or data is released. That said, I do not see circularity: the central comparisons are against external baselines, and the risk is protocol ambiguity rather than a derivation that assumes the conclusion. The method itself is credible and the idea is worth pursuing.\n\nI would send this to peer review rather than desk reject, because the core idea is testable and the ablation is positive, but I would make review contingent on the authors providing exact test splits, a description of the curation, a controlled re-run of baselines, and ideally code. If those cannot be produced, the SOTA claim should be withdrawn. The paper is useful for researchers in emotion-controllable talking heads, mainly as an idea paper that still needs a trustworthy evaluation.","headline":"Useful recombination of CoT and progressive conditioning, but the SOTA claim is unverifiable until the evaluation protocol is documented and baselines are re-run.","tokens_in":15065,"tokens_out":1631,"would_cite":false,"duration_ms":20473,"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":"Think-Before-Draw claims that splitting an emotion label into coarse and muscle-level text prompts, applied in that order during diffusion denoising, yields fine-grained control over emotional talking-head video and state-of-the-art…","keywords":["emotional talking-head generation","chain-of-thought reasoning","facial action units","text-driven facial animation","progressive guidance denoising","diffusion model","facial expression control","zero-shot generation"],"falsifier":"Re-run MakeItTalk, SadTalker, PC-AVS, EAMM, StyleTalk, and EAT on the exact same curated MEAD split and preprocessing pipeline used for TBD, and recompute the Table 1 metrics with the same evaluation code; if the PSNR gap over StyleTalk collapses or reverses, the state-of-the-art claim for the stated protocol does not hold.","tokens_in":14079,"feed_emoji":"🎭","tokens_out":8317,"duration_ms":81041,"temperature":0.7,"pith_summary":"The authors set out to show that discrete emotion labels are too coarse to drive natural talking-head animation, and that decomposing those labels into muscle-level text before generation yields finer, more controllable expressions. Their framework, Think-Before-Draw, first runs a chain-of-thought module that expands an emotion label into a holistic prompt (identity plus overall emotion) and a fine-grained prompt naming the facial action units and muscles involved. A diffusion-based video generator then applies the holistic prompt during early denoising steps and the fine-grained prompt during later steps, mimicking an artist's global-to-local workflow. The authors report that this scheme outperforms prior audio-, video-, and label-driven methods on the MEAD and HDTF benchmarks, and that it generalizes to previously unseen portrait images. If correct, the framework gives text-driven talking-head systems a practical route from abstract emotion semantics to anatomically grounded facial motion.","feed_headline":"Chain-of-thought turns emotion labels into muscle-level video prompts","feed_subtitle":"A two-stage denoiser uses global emotion first, facial-muscle details second, with the best reported results on MEAD and HDTF.","key_machinery":"The load-bearing machinery is the combination of a Chain-of-Thought Facial Animation module (CoT-FA) with a progressive guidance denoising schedule. CoT-FA is a four-tier prompt pipeline—facial information extraction, facial action unit analysis, facial muscle analysis, and prompt design—that converts an emotion label and a reference image into two text embeddings: a coarse global condition and a fine muscle-level condition. The progressive strategy feeds the coarse embedding into the denoising U-Net's text-attention layers during the first denoising stage and switches to the fine embedding in the second stage, blending the two with an alpha weight that depends on the timestep's position within its stage rather than using a fixed value. This global-to-local conditioning scheme is the mechanism that carries the argument, because it is what lets a single diffusion backbone produce expressions that are both emotionally coherent and locally detailed.","core_discovery":"The central claim is that the semantic gap between an abstract emotion label and facial motion can be closed by decomposing the label before generation. The paper's CoT-FA module performs this decomposition in four stages: it extracts identity attributes from the reference image, identifies the Facial Action Units associated with the emotion under the Facial Action Coding System, analyzes the underlying facial muscles and their synergies, and finally designs two text prompts—a coarse global description and a fine muscle-level description. The progressive guidance denoising strategy then schedules these conditions so that the global prompt steers the early denoising steps and the local prompt steers the later steps, with a timestep-dependent blending weight that smooths the transition between stages. On the MEAD benchmark the method reports PSNR 28.64, SSIM 0.88, and FID 17.28, and on HDTF PSNR 23.28, SSIM 0.74, and FID 16.17, which the authors present as state-of-the-art results; the lone exception is lip-sync confidence on HDTF, attributed to the uniform neutral text labels used because that dataset has no emotion annotations.","pith_inferences":["Inference: If the MEAD numbers in Table 1 were produced under the same protocol for all methods, the unusually large PSNR gap over StyleTalk (28.64 vs. 21.60) would be strong evidence for the method; if baselines were not re-run on the same curated split, that gap needs verification before being read comparatively.","Inference: The progressive guidance idea is not specific to faces: any conditional diffusion task with a natural coarse-to-fine prompt ordering, such as scene-level then object-level text-to-video, could adopt the same two-stage scheduling.","Inference: Because CoT-FA prompts pass through manual review, the pipeline's ceiling depends on how reliably the language model produces anatomically correct muscle descriptions; an automated check against known action-unit-muscle mappings would be a cheap way to test that dependency.","Inference: The three annotated intensity levels suggest a path to continuous intensity control: if the language model can interpolate between mild, moderate, and intense descriptions, users could dial expression strength continuously rather than choosing from discrete levels."],"forward_implications":["Text-driven talking-head systems can replace flat emotion labels with multi-level prompts, giving users explicit control over expression intensity as well as emotion category.","The two-stage guidance schedule can be layered onto any diffusion-based talking-head generator that already accepts text conditioning, since it only changes how conditions are ordered during denoising.","Because each generated expression is traceable to the action units and muscles named in the prompt, the framework makes emotional animation more explainable and easier to debug.","For a fixed 40-step denoising budget, two-stage partitioning is reported to beat three- and four-stage partitioning, since too many switches destabilize generation.","On HDTF, where emotion annotations are absent, the method still improves image quality and FID over baselines using neutral labels, with the only reported weakness being lip-sync confidence."],"supporting_citations":[{"why":"Supplies the Facial Action Coding System and action-unit semantics that ground the fine-grained muscle descriptions.","marker":"[17]"},{"why":"Supplies the chain-of-thought prompting technique that CoT-FA uses to decompose emotion labels.","marker":"[18]"},{"why":"The multimodal language model that carries out the four-stage chain-of-thought decomposition from image and text.","marker":"[13]"},{"why":"Supplies the MEAD dataset with emotion and intensity annotations used for training and evaluation.","marker":"[45]"},{"why":"Supplies the HDTF dataset used as the second benchmark.","marker":"[46]"},{"why":"Serves as a label/text-driven emotional talking-head baseline that TBD compares against.","marker":"[10]"},{"why":"Serves as a video-driven style-transfer baseline that TBD must beat on emotion control.","marker":"[7]"},{"why":"Supplies the pretrained motion-module weights used to initialize the temporal layers.","marker":"[42]"}],"fun_headline_variants":["Emotion labels become muscle-level prompts via chain-of-thought","Think-before-draw: CoT converts emotions to facial muscle specs","Two-stage denoising: global emotion first, micro-muscles second","CoT decomposition turns emotion labels into fine-grained control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that all methods in Table 1 were evaluated under the same protocol, since the paper trains on an unspecified curated version of MEAD and does not state that the baselines were re-run on that same split with identical preprocessing.","fun_headline_variants_meta":{"raw":{"variants":["Emotion labels become muscle-level prompts via chain-of-thought","Think-before-draw: CoT converts emotions to facial muscle specs","Two-stage denoising: global emotion first, micro-muscles second","CoT decomposition turns emotion labels into fine-grained control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4122,"prompt_tokens":1034,"completion_tokens":3088,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":3014}},"tokens_in":650,"tokens_out":3088,"duration_ms":25755,"temperature":1.0,"reasoning_tokens":3014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:38:40.333290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run MakeItTalk, SadTalker, PC-AVS, EAMM, StyleTalk, and EAT on the exact same curated MEAD split and preprocessing pipeline used for TBD, and recompute the Table 1 metrics with the same evaluation code; if the PSNR gap over StyleTalk collapses or reverses, the state-of-the-art claim for the stated protocol does not hold.","supporting_citations":[{"cited_title":"Ekman, W","cited_arxiv_id":null,"evidence_quote":"Supplies the Facial Action Coding System and action-unit semantics that ground the fine-grained muscle descriptions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MEAD dataset with emotion and intensity annotations used for training and evaluation."},{"cited_title":"Zhang, L","cited_arxiv_id":null,"evidence_quote":"Supplies the HDTF dataset used as the second benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a label/text-driven emotional talking-head baseline that TBD compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a video-driven style-transfer baseline that TBD must beat on emotion control."}],"review_version":1}