{"id":"7b3795eb-303e-45dd-b717-ecc1c48efefe","arxiv_id":"2509.03883","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive survey with a five-phase pipeline model for human motion video generation, covering over 200 papers and adding a new benchmark comparison of nine pose-guided methods.","lead":"This paper surveys more than 200 studies on generating videos of human motion from text, audio, or images, and organizes the work into five pipeline stages. It also benchmarks nine open-source pose-guided methods on a common dance-video test set, giving researchers a quantitative starting point.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark in §VI.B is unreliable: unofficial Animate Anyone and 10-video test set make Table VII rankings unsupported","rationale":"The survey's main contribution is its five-phase taxonomy and literature coverage, which is well-structured and cites a broad range of works. The benchmark is a secondary but concrete contribution; its flaws are localized and addressable. The reader's weakest_assumption correctly identified the unofficial baseline and small test set. We agree and add the internal inconsistencies (figure caption, method name, cutoff) as supporting evidence. The concern does not warrant rejection but should be fixed or the claims softened, which is consistent with the reader's CONDITIONAL verdict.","tokens_in":35641,"tokens_out":3933,"duration_ms":36769,"concrete_test":"Release the evaluation pipeline and test-video IDs, then rerun all nine methods (a) with 5 different seeds and report mean±std on the full TikTok test set, and (b) with the official Animate Anyone weights once available (or with a second independent reimplementation). If the per-method ranking changes or the top-2 differences are within one standard deviation, Table VII's conclusions should be rephrased as indicative only. Also verify the Fig. 12 caption method count and correct 'StableAnimation' to 'StableAnimator'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's quantitative contribution (Table VII, Fig. 12) compares nine pose-guided methods on the TikTok test set. This comparison is load-bearing for practitioners selecting methods, but it rests on two fragile assumptions. First, Animate Anyone is represented by an unofficial reimplementation (Moore-AnimateAnyone), not the original model; if the reimplementation deviates in architecture, training data, or post-processing, the reported L1=0.0779 cannot be attributed to the original method. Second, with only 10 test videos and no error bars or significance tests, the observed differences (e.g., SSIM 0.7558 vs 0.7436) may be within run-to-run variance. The figure caption in Fig. 12 says 'five open-source methods' while the text says nine, and Table VII lists 'StableAnimation' instead of the cited 'StableAnimator' (ref [208]), indicating the experimental presentation is not carefully checked. Additionally, StableAnimator (arXiv:2411.17697) and Animate-X (arXiv:2410.10306) are dated after the stated August 30, 2024 cutoff, so the benchmark contradicts the paper's own scope statement. These issues do not invalidate the survey's taxonomy but do undermine the benchmark's reliability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys human motion video generation, organizing the field into five phases (input, motion planning, motion video generation, refinement, output) and three driving modalities (vision, text, audio). It reviews over 200 papers, provides a table of 64 datasets, summarizes metrics, and reports a new comparative benchmark of nine pose-guided dance-video generation methods evaluated on 10 TikTok test videos with six metrics. The paper claims to be the first survey to discuss LLM-based motion planning and to provide a complete pipeline-oriented taxonomy.","tokens_in":35971,"tokens_out":3619,"duration_ms":35138,"significance":"If the survey's coverage and taxonomy hold, it would be a useful reference: the five-phase decomposition gives a unified vocabulary, the dataset table is extensive, and the LLM-planning discussion addresses an emerging topic. The paper ships an open repository of reviewed models. The independent benchmark is a potentially valuable practitioner resource, but its reliability is currently limited by the fragile experimental setup; it should not be used to rank methods until variance and faithfulness of the implementations are addressed.","major_comments":[{"comment":"The quantitative ranking relies on an unofficial reimplementation of Animate Anyone (footnote 1) and a 10-video test set. No error bars or significance tests are provided. For example, the gap between the best SSIM (MagicAnimate, 0.7558) and the second-best (UniAnimate, 0.7436) is 0.0122, which may be within run-to-run variance on 10 videos. This makes Table VII's rankings unsupported as stated. Please report variance/confidence intervals and per-video breakdowns, verify the unofficial implementation against the original (or relabel the column as the reimplementation), and either enlarge the test set or explicitly describe the benchmark as illustrative rather than comparative.","section":"Section VI.B, Table VII, Fig. 12"},{"comment":"Table VII lists 'StableAnimation [208]' while reference [208] is StableAnimator, and Fig. 12's caption says 'five open-source methods' although the text lists nine. These inconsistencies indicate the experimental presentation has not been carefully checked. Moreover, StableAnimator (arXiv:2411.17697) and Animate-X (arXiv:2410.10306) have arXiv dates after the stated August 30, 2024 cutoff, contradicting the paper's note that all statistics are current as of that date. Please correct the naming and reconcile the benchmark with the stated cutoff, or revise the scope statement.","section":"Table VII, Fig. 12, Section II scope note"},{"comment":"The paper states it does not cover video generation through NeRF or 3DGS rendering, yet Table V includes SD-NeRF [138] (a NeRF-based talking-head method) and Table VI includes TalkingGaussian [142] and GaussianTalker [143] (3D Gaussian-based). This creates an internal inconsistency in the taxonomy. Either clarify the inclusion criteria for these 3D-representation methods so they are consistent with the stated scope, or remove them from the tables.","section":"Fig. 2 caption and Tables V-VI"}],"minor_comments":[{"comment":"Typo: 'Artificial Intelligence Innovation and Incubation (Al’) Institute' should be 'AI Institute'.","section":"Author affiliations"},{"comment":"The caption says 'five open-source methods' but the text and Table VII describe nine methods; please align the caption with the actual content.","section":"Fig. 12 caption"},{"comment":"The real-time support column is labeled 'Estimated FPS based on NVIDIA RTX 4090 GPU' but the basis of the estimate is not stated. Please specify whether these are reported by the original papers or measured here.","section":"Tables III and V"},{"comment":"The legend contains garbled/overlapping text, e.g., 'text Consider (ignore) the temporal dim.(text)'. Please redraw the figure for legibility.","section":"Fig. 11"},{"comment":"The Mamba reference (Gu and Dao) lacks a full bibliographic venue/year entry; please complete it. Also change 'StableAnimation' to 'StableAnimator' in Table VII.","section":"Reference [110]"}],"recommendation":"major_revision","confidential_remarks":"The survey body is substantive and well-structured, but the benchmark section is a clear weak point and the scope inconsistencies need to be resolved. I recommend major revision; the paper could become acceptable after the quantitative claims are properly qualified and the internal contradictions are fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely useful survey. The five-phase pipeline—input, motion planning, motion video generation, refinement, output—is a workable frame for a chaotic field, and the coverage is broad: 200+ papers, 64 datasets. The emphasis on LLM-based motion planning is timely, and the summary of attention-fusion variants (Fig. 8) is a nice synthesis. Second, the survey's own benchmark is the weakest part; the quantitative rankings in Table VII should not be taken at face value.\n\nThe descriptive content is generally accurate and well-referenced. The classification into audio/text/vision-driven methods is reasonable, and the dataset table with brief descriptions is a real service to the community. I'd point a new student to this survey before the competing ones.\n\nThe soft spots are concentrated in Section VI.B. The comparison of nine pose-guided methods uses an unofficial reimplementation of Animate Anyone, so any statement like 'Animate Anyone leads in L1' is not about the original method. The test set is 10 videos, with no error bars or significance tests; differences like SSIM 0.7558 vs 0.7436 could easily be noise. There are also presentation errors: Fig. 12 says 'five open-source methods' while the text says nine, Table VII lists 'StableAnimation' for 'StableAnimator', and two of the nine methods (StableAnimator and Animate-X) are dated after the paper's stated August 30, 2024 cutoff. These are not deep conceptual flaws—they are fixable—but they undermine the benchmark's reliability.\n\nMinor: there is a typo 'Al Institute' in the biography, and the statistics date claim conflicts with those two references.\n\nBottom line: this survey deserves a serious referee and, after revision, likely acceptance. The taxonomy, coverage, and dataset list are valuable. The benchmark should either be removed, redone with proper protocols, or heavily caveated. I'd bring it to a reading group to discuss the structure and the benchmark pitfalls.","headline":"Useful survey with a valuable taxonomy; its benchmark is the weak link and should be revised or dropped.","tokens_in":36423,"tokens_out":3362,"would_cite":true,"duration_ms":29311,"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":"A five-phase pipeline organizes human motion video generation, with LLMs as motion planners.","keywords":["human motion video generation","survey","five-phase pipeline","large language models","motion planning","diffusion models","pose-guided dance generation","talking head synthesis"],"falsifier":"Rerun the nine pose-guided methods on a larger and independently annotated dance test set using official model weights wherever they exist; if the top rankings across SSIM, PSNR, LPIPS, FID, and FID-VID change materially, the survey's comparative conclusions about MagicAnimate and UniAnimate would not generalize.","tokens_in":1401,"feed_emoji":"🎬","tokens_out":1928,"duration_ms":44243,"temperature":0.7,"pith_summary":"The paper tries to give the field a shared structure by decomposing human motion video generation into five phases: input, motion planning, motion video generation, refinement, and output. It organizes more than two hundred papers into vision-, text-, and audio-driven categories and claims to be the first survey to discuss large language models as motion planners. If this framing is accepted, researchers gain a unified map for comparing methods, a taxonomy of diffusion-model architectures, and a new quantitative baseline for pose-guided dance video generation. The survey also collects 64 human-centered datasets and identifies common failure modes across current methods.","feed_headline":"Survey maps human motion video generation in five phases","feed_subtitle":"It treats LLMs as motion planners and benchmarks nine pose-guided methods on TikTok test videos.","key_machinery":"The central object is the five-phase generation pipeline (input, motion planning, motion video generation, refinement, output), used as the organizing scaffold for the whole survey. The paper's key analytic tools are a three-way modality classification (audio-, text-, and vision-driven) and a taxonomy of diffusion-based frameworks, including three input schemes for the main diffusion branch and five attention-fusion arrangements. For motion planning, the distinguishing mechanism is the LLM planner, which either retrieves motion descriptions from a database or projects them into a generative model's latent space.","core_discovery":"The paper claims that existing surveys cover individual subtasks but miss the whole generative process, so it proposes a five-phase pipeline spanning every step from input to deployed output. Within this pipeline, it distinguishes two motion-planning strategies: implicit feature mapping and LLM-based planning, with the latter shown in two forms—retrieval of fine-grained motion descriptions and direct projection into a motion latent space. For the generation phase, it classifies diffusion-based pose-driven methods by what feeds the main diffusion branch (pure noise, a reference image, or guided conditions) and by attention-fusion variants. It also contributes a comparative benchmark of nine o","pith_inferences":["If the five-phase framing takes hold, motion planning may be evaluated separately from video rendering, leading to benchmarks that score planner quality (e.g., retrieval rank, instruction adherence) independently of pixel-level metrics.","The LLM-planner direction suggests a future separation of semantic motion specification from appearance rendering, which would let planners and generators be upgraded independently.","The benchmark could naturally extend to audio-driven and text-driven methods under the same metrics, enabling cross-modal comparison that the current pose-guided-only setup does not provide.","The survey's deliberate exclusion of 3D skeleton motion, NeRF, and 3D Gaussian splatting leaves room for a complementary survey that bridges the 2D video and 3D representation worlds."],"forward_implications":["Any new method can be located within the five-phase pipeline and compared against its neighbors, giving the field a common coordinate system.","LLM-based motion planning becomes a distinct research branch, with retrieval-style and generative-style planners as two identifiable architectures.","The nine-method benchmark gives practitioners a quantitative starting point: MagicAnimate for structural fidelity, UniAnimate for perceptual and temporal quality, and Animate Anyone for pixel-level closeness to ground truth.","The catalog of common failure modes—hand blur, facial distortion, identity or background inconsistency, limb dislocation, and low resolution—serves as a concrete checklist for evaluating new methods.","The 64-dataset collection supplies a fast route to training data, including resolution, duration, and task-support annotations."],"supporting_citations":[{"why":"Supplies the pioneering LLM-based listener-expression planner, grounding the survey's claim that LLMs can do motion planning.","marker":"[17]"},{"why":"Extends LLM planning beyond dialogue to general avatar contexts; one of the retrieval-style planner examples.","marker":"[18]"},{"why":"Provides the generative-style LLM planner with instruction-video data, the other planner architecture the survey identifies.","marker":"[19]"},{"why":"Animate Anyone is a central pose-driven dance generation baseline and the method whose unofficial reimplementation anchors the benchmark.","marker":"[33]"},{"why":"MagicAnimate is the benchmark's top performer on SSIM and PSNR, and a key example of the pure-noise diffusion framework.","marker":"[78]"},{"why":"UniAnimate is the benchmark's top performer on LPIPS, FID, and FID-VID, and the example of a temporal Mamba-based diffusion variant.","marker":"[100]"},{"why":"Supplies the TikTok test dataset used for the nine-method comparative benchmark.","marker":"[207]"},{"why":"DisCo is one of the nine benchmarked methods and a representative early pose-driven approach using ControlNet.","marker":"[32]"},{"why":"Champ is a benchmarked method and an example of integrating a 3D parametric body model into latent diffusion.","marker":"[74]"},{"why":"MimicMotion is a benchmarked method and the example of hand-region enhancement during generation.","marker":"[36]"}],"fun_headline_variants":["Five phases of human motion video generation","Survey diagrams five-step motion video pipeline","LLM enters human motion video planning","Nine pose-guided methods benchmarked in survey","First survey to cover LLM for motion video"],"cache_read_input_tokens":38272,"weakest_assumption_plain":"The comparative benchmark assumes that an unofficial reimplementation of Animate Anyone faithfully represents the original model and that ten TikTok test videos are enough to rank the nine methods.","fun_headline_variants_meta":{"raw":{"variants":["Five phases of human motion video generation","Survey diagrams five-step motion video pipeline","LLM enters human motion video planning","Nine pose-guided methods benchmarked in survey","First survey to cover LLM for motion video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000327,"raw_usage":{"total_tokens":1653,"prompt_tokens":717,"completion_tokens":936,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":872}},"tokens_in":461,"tokens_out":936,"duration_ms":8609,"temperature":1.0,"reasoning_tokens":872,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:33:18.743062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the nine pose-guided methods on a larger and independently annotated dance test set using official model weights wherever they exist; if the top rankings across SSIM, PSNR, LPIPS, FID, and FID-VID change materially, the survey's comparative conclusions about MagicAnimate and UniAnimate would not generalize.","supporting_citations":[{"cited_title":"AgentAvatar: Disentangling Planning, Driving and Rendering for Photorealistic Avatar Agents","cited_arxiv_id":"2311.17465","evidence_quote":"Extends LLM planning beyond dialogue to general avatar contexts; one of the retrieval-style planner examples."}],"review_version":1}