{"id":"60a2611f-1ce5-4a32-9760-6c5b32a31103","arxiv_id":"2411.19459","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Fleximo generates a video of a person from one reference image and a text description of the motion, by chaining a text-to-3D-motion model, a skeleton rescaling step, and a trained skeleton adapter that fills in hand and face details.","lead":"This paper introduces Fleximo, a system that creates videos of a specific person performing motions described in plain text, using only one photo of that person. It combines a text-to-motion model with a motion-to-video generator, plus new tricks to fix scale and add hand details.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's PSNR/SSIM/LPIPS/DreamSim are computed against the static reference image, so they penalize genuine motion and cannot support the claim that Fleximo outperforms competitors; MotionScore is unvalidated and its margin is tiny.","rationale":"The reader identifies the anchor-point rescale as the weakest assumption, which is a real and honestly reported limitation. However, the most load-bearing issue for the paper's headline claim is the evaluation methodology: the quantitative comparison in Table 1 uses reference-image similarity metrics that inherently reward less motion, so they cannot measure the quality of a motion-generation task. The only motion-specific metric, MotionScore, is not validated and shows a very small, unquantified margin. The user study is too small to carry the weight of the comparison. Even if the pipeline works exactly as described, the evidence does not establish that Fleximo outperforms existing methods. This is an unacknowledged flaw in the paper's argument, distinct from the acknowledged skeleton-scaling limitation. I therefore recommend keeping the CONDITIONAL verdict: the system contribution is plausible and the demos are encouraging, but the comparative claim needs a valid evaluation (motion-aware metrics, validated MotionScore, larger user study, and artifact release) before it can be accepted. The reader's CONDITIONAL verdict is appropriate; my specific reason differs from the reader's stated weakest_assumption, hence 'disagree' on that narrow question.","tokens_in":13472,"tokens_out":8905,"duration_ms":80502,"concrete_test":"Compute, for each of the 400 MotionBench videos, the mean absolute DWPose keypoint displacement between the first and last frame as a motion-magnitude measure, and include a static baseline (the reference image repeated). Then correlate PSNR/SSIM (computed against the reference image) with motion magnitude across I2VGen-XL, VideoCrafter, DynamiCrafter, and Fleximo. If higher reference-image PSNR correlates with lower motion magnitude, or if the static baseline matches or exceeds Fleximo's PSNR/SSIM, the identity/video-quality metrics in Table 1 are biased toward static output and cannot support the outperformance claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim 'outperforms existing text-conditioned image-to-video generation methods' rests on Table 1 and the user study in Section 5.2 and 5.4. Table 1 reports PSNR, SSIM, LPIPS, and DreamSim by comparing generated frames to the single reference image. For a motion-generation task this is not a valid ground truth: a video that freezes on the reference image would score perfectly, while any real motion reduces pixel-level similarity. Higher reference-image similarity therefore conflates 'static' with 'high quality', so Fleximo's better numbers cannot be read as evidence of better motion videos. The only motion-following metric, MotionScore (Section 4), is an unvalidated LLM-plus-CLIP similarity; its advantage over DynamiCrafter is 0.6990 vs 0.6868 (Δ = 0.0122), with no error bars, no significance test, and no correlation with human judgment. The user study uses only 8 participants and no statistical analysis. The acknowledged failure case (Section 5.6) shows scale jitter and identity loss on large movements, but the metrics in Table 1 would not detect such a failure because they reward staying close to the reference image. If the evaluation is invalid, the comparative claim is unsupported even though the qualitative demos and the skeleton-adapter design are plausible.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new task, text-to-human motion video generation, in which a single reference image and a natural-language description are used to synthesize a video of the reference identity performing the described motion. To avoid training on large text-video datasets, Fleximo first generates 3D human motion with a pretrained text-to-motion model (T2M-GPT), projects the resulting meshes to 2D skeletons, and aligns those skeletons to the reference image using an affine transformation followed by a neck-anchored per-bone rescaling (Section 3.2). A skeleton adapter trained on HumanVid fills in missing hand and face keypoints (Section 3.3), and the resulting skeleton video is fed to a MimicMotion-style motion-to-video model together with the reference image; an optional refinement pass and LLM-based decomposition into motion segments support longer or multi-part motions. The paper also introduces MotionBench, a benchmark of 400 videos over 20 identities and 20 motions, and MotionScore, an LLM-plus-CLIP metric for motion-text alignment. Quantitative comparisons against I2VGen-XL, VideoCrafter, and DynamiCrafter, a small user study, ablations, and failure cases are reported.","tokens_in":13692,"tokens_out":4511,"duration_ms":38810,"significance":"The proposed task is well motivated and the modular pipeline is a sensible way to sidestep the lack of large text-motion-video paired data: reusing T2M-GPT and a pretrained motion-to-video model is practical, and the skeleton adapter addresses a real gap between 18-point text-to-motion skeletons and the full-body keypoints expected by motion-to-video models. If the comparative claims were properly supported, Fleximo would be a useful baseline and the MotionBench/MotionScore package would be a community resource. However, the current quantitative evidence does not establish the central claim of outperforming existing methods: the main image-similarity metrics are computed against the static reference image, which rewards stillness, MotionScore is an author-proposed metric with no validation and a very small margin, and the user study is too small and lacks statistical analysis. The paper's own failure case for large positional shifts further limits the scope of the claimed flexibility. These issues are fixable with additional experiments, which is why I am not recommending rejection.","major_comments":[{"comment":"The PSNR, SSIM, LPIPS, and DreamSim numbers in Table 1 are computed by comparing each generated frame to the single reference image, as stated in Section 5.2. For a motion-generation task this is not a valid notion of ground truth: a video that simply freezes on the reference image would score perfectly, and any real motion necessarily reduces pixel-level similarity. The large reported gap (PSNR 16.647 for Fleximo vs. 9.607 for DynamiCrafter) is therefore consistent with Fleximo staying closer to the reference and cannot support the claim that Fleximo generates better motion videos. These metrics should be replaced or supplemented with motion-aware evaluations, for example comparing extracted pose sequences against the text-to-motion skeletons or against ground-truth motion from reference videos, and per-video error bars or confidence intervals should be reported.","section":"Section 5.2, Table 1"},{"comment":"MotionScore is an author-proposed metric whose validity is not demonstrated. The paper reports that MotionLLM is prompted to describe the motion in the generated video and that CLIP cosine similarity between that description and the input text is computed, but it does not show that MotionScore correlates with human judgments, does not report per-motion or per-video distributions, and does not provide any statistical test. The reported advantage over DynamiCrafter is 0.6990 versus 0.6868, a difference of 0.0122 that is likely within the noise of an LLM-plus-CLIP pipeline. The authors should validate MotionScore against human ratings, show that it detects failures such as the basketball case in Figure 9, and report confidence intervals and significance tests.","section":"Section 4, MotionScore"},{"comment":"The user study does not support the strength of the claims drawn from it. There are only 8 participants, and with 800 videos and each participant scoring 100 videos, each video appears to receive a single rating; no inter-rater agreement, significance test, or confidence interval is reported. The 2.3-point gap in motion alignment between Fleximo and DynamiCrafter therefore cannot be separated from rating noise. The study should be expanded or, at minimum, the authors should report per-item statistics, a repeated-measures analysis, and inter-rater agreement.","section":"Section 5.4, Table 2 and Supplementary Section 9"},{"comment":"The acknowledged failure case is load-bearing for the central claim. The paper states that for motions with large positional shifts, such as playing basketball, the skeleton scaling becomes inconsistent, causing scale jitter and poor identity preservation. This is precisely the kind of dynamic motion that the task is meant to support, and several motions in MotionBench (running, jumping, boxing, playing golf) are likely to trigger the same behavior. The metrics in Table 1 would not detect such failures because they reward proximity to the reference image. The authors should quantify how often the pipeline degrades on each motion type, for example by reporting per-motion MotionScore or per-motion human ratings, and should qualify the claims of 'flexible' and 'any desired length' accordingly.","section":"Section 5.6, Figure 9"}],"minor_comments":[{"comment":"The notation yd and yr is described as 'y-coordinates of detected bodies and the reference body,' but it is not clear how many keypoints are used in the polyfit or whether both arms and legs are included; please specify the exact set of keypoints used for the affine fit.","section":"Section 3.2, Eq. (3)"},{"comment":"The paper uses both 'skeleton adapter' and 'skeleton adaptor'; please unify the terminology.","section":"Section 3.3 vs. Figure 2"},{"comment":"The ablation study is presented as Figure 6 in the main text but the text refers to it as 'Fig. 14'; please correct the cross-reference.","section":"Section 5.5"},{"comment":"References [7] and [8] are the same paper (Guo et al., 'Generating diverse and natural 3d human motions from text') and should be merged.","section":"References"},{"comment":"The sentence 'The results in Tab. 2 are averaged on 200 videos' is imprecise; it should read 'averaged over the 200 videos generated by each method,' and the fact that each video receives a single rating should be stated explicitly.","section":"Supplementary Section 9"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid systems contribution with a plausible pipeline and useful qualitative demonstrations, but the comparative claim currently rests on evaluation practices that reward static outputs and on an unvalidated self-metric with a negligible margin. I would ask the editor to require (i) motion-grounded quantitative evaluation, (ii) validation of MotionScore against human judgments, and (iii) a statistically adequate user study before considering acceptance. The MotionBench resource is potentially valuable, but the paper should make clear that the benchmark's reference images include generated images and that MotionScore is an author-proposed metric without external validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhang, Zhou, et al. ask whether a single photo plus natural language can drive a human motion video, and their answer is a working pipeline: T2M-GPT generates 3D motion, gets projected and rescaled to the reference image, a skeleton adapter hallucinates hand keypoints, and MimicMotion renders the video. That task formulation is genuinely new relative to the cited work, which needs pose videos or skeleton sequences, and the skeleton adapter is a real component. The paper is also honest: it reports the basketball failure case in Section 5.6, and it does not pretend that large positional shifts work smoothly.\n\nWhat I would push back on is the evaluation. Table 1 computes PSNR, SSIM, LPIPS, and DreamSim against the static reference image. For a motion-generation task that is not a valid ground truth: a frozen frame would score perfectly, and any real motion reduces the score. Those numbers therefore cannot support the claim that Fleximo outperforms DynamiCrafter or the others. The motion-following metric, MotionScore, is a reasonable idea but unvalidated, and the reported margin (0.699 vs 0.687) has no error bars or significance test. The user study has 8 participants. That is a load-bearing weakness because the paper's central claim is exactly the comparative superiority.\n\nThe anchor-rescale step is also the soft spot the authors themselves flag: for motions with big positional shifts, per-frame affine fits and bone rescaling become inconsistent, and the identity suffers. Since the benchmark includes jumping, boxing, and playing golf, those failures are not edge cases; they are part of the target task. None of this makes the paper worthless, but it makes the \"outperforms\" claim unsupported as written.\n\nWho should read this? People building text-driven human animation systems, especially those trying to reuse T2M and M2V models. The pipeline is a sensible composition and the skeleton adapter is worth borrowing. As a referee, I would send it out but require a reworked evaluation: motion-aware metrics with reference to the actual motion, validation of MotionScore against human judgments, error bars, and ideally a released implementation. The paper deserves the referee time, but it needs meaty revisions before it supports its own conclusions.","headline":"A plausible new-task pipeline whose comparative claim is undermined by a static-reference evaluation and a tiny, unvalidated motion metric.","tokens_in":14314,"tokens_out":2209,"would_cite":false,"duration_ms":19845,"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":"Fleximo generates human motion videos from one photo and text","keywords":["text-to-human motion video generation","text-to-motion","motion-to-video generation","skeleton adapter","anchor point rescale","LLM planning","MotionBench","MotionScore"],"falsifier":"Run Fleximo on a motion with large positional shift (e.g., \"A person is playing basketball\" with side-facing movement) and measure frame-to-frame variation of the per-bone scaling factors and face identity similarity against the reference image. If the scaling variance is high and identity similarity drops sharply relative to smaller motions, the anchor-point rescale assumption is violated and the central claim of consistent high-quality generation fails.","tokens_in":13217,"feed_emoji":"🎬","tokens_out":4811,"duration_ms":37381,"temperature":0.7,"pith_summary":"This paper introduces a new task, text-to-human motion video generation, where the input is a single reference image plus natural language and the output is a video of that person performing the described motion. The authors claim that by combining pretrained text-to-3D motion models with a motion-to-video generator, they can produce high-quality motion videos without needing millions of text-video training pairs. The key moves are projecting text-generated 3D skeletons to 2D, rescaling them to match the reference person via anchor-point alignment, and using a skeleton adapter to fill in missing hand and face details. The paper also contributes MotionBench, a 400-video benchmark, and MotionScore, a metric for motion-text alignment, and reports that Fleximo outperforms existing text-conditioned image-to-video methods.","feed_headline":"Fleximo generates human motion videos from one photo and text","feed_subtitle":"No pose video needed: a pretrained text-to-motion model and a skeleton adapter make text the only control.","key_machinery":"The load-bearing mechanism is the anchor-point based rescale combined with the skeleton adapter. The rescale method chooses the neck as a fixed anchor and then rescales each connected body part (arms, legs, hands) individually to match the reference image's limb lengths, after a global affine alignment; this is what makes the text-generated skeleton match the identity in the reference image. The skeleton adapter is a latent video diffusion model that takes a handless skeleton video and a reference skeleton image and generates a complete skeleton video with realistic hand keypoints, bridging the format gap between T2M-GPT's 18 body points and the full-body keypoints required by motion-to-video models.","core_discovery":"The central claim is that a reference image plus natural language is sufficient to generate realistic human motion videos, and that this can be achieved by bridging two existing model families: text-to-motion and motion-to-video. Concretely, Fleximo projects skeletons from a pretrained text-to-3D motion model into 2D, rescales them frame-by-frame to the reference image using an affine transform followed by per-bone anchoring at the neck, enriches the sparse skeleton with a skeleton adapter that synthesizes hand and face keypoints, and refines the final video by re-extracting DWPose skeletons from a first-pass anchor video. With LLM-based planning, long text is decomposed into discrete motion segments that can be concatenated into videos of any desired length. The paper reports that this pipeline outperforms text-conditioned image-to-video baselines on identity fidelity, video quality, and motion following.","pith_inferences":["If scale alignment is the main failure mode, a depth-aware or learned skeleton-rescale step could extend Fleximo to motions with large translation and out-of-plane rotation, such as running toward the camera.","The LLM planning template is not motion-specific and could be applied to other long-horizon generation tasks where a text instruction must be segmented into executable sub-steps.","MotionScore could be repurposed as a general motion-following evaluation for any generated human video, though its reliability depends on the captioning LLM's ability to describe motion accurately.","The skeleton adapter's training scheme—animate a full skeleton from a handless one—is a general way to inject fine-grained missing details and could be applied to facial expressions, props, or object interactions in other animation pipelines."],"forward_implications":["Users can animate any reference photo with natural language, without sourcing pose videos, making motion control as easy as typing a sentence.","Because the pipeline reuses large pretrained models and only needs a modest training set for the adapter, text-to-human motion video becomes feasible without millions of text-video pairs.","Long and compound instructions can be generated by the LLM planner, enabling multi-motion videos that concatenate seamlessly.","The MotionScore metric offers a way to judge motion following that avoids the bias of binary or forced-choice questioning.","The refinement loop improves identity consistency by reusing DWPose skeletons from a first-generated anchor video, suggesting a general self-improvement strategy for skeleton-guided animation."],"supporting_citations":[{"why":"Supplies the text-to-3D motion generator that turns input text into 3D mesh vertices, the upstream of the whole pipeline.","marker":"[39]"},{"why":"Provides the motion-to-video backbone and full-body skeleton format that Fleximo fine-tunes and that defines what the skeleton adapter must produce.","marker":"[43]"},{"why":"Detects reference-image keypoints for the anchor-point rescale and extracts skeleton videos for the refinement stage.","marker":"[38]"},{"why":"Provides the human motion video training set used to train the skeleton adapter and fine-tune the video model.","marker":"[34]"},{"why":"Supplies the LLM that decomposes long motion text into discrete segments for arbitrary-length generation.","marker":"[29]"},{"why":"A text-conditioned image-to-video baseline that Fleximo compares against on MotionBench.","marker":"[35]"},{"why":"Another image-to-video baseline that Fleximo compares against and outperforms on identity and quality metrics.","marker":"[42]"}],"fun_headline_variants":["One photo and text generate human motion videos","No pose video needed: text controls motion video","Fleximo turns a single photo and prompt into video","Text as the sole guide for human motion video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that a text-generated skeleton can be aligned to the reference person by an affine transform followed by per-bone rescaling, and that this alignment stays consistent for all frames of a motion; when a motion involves large positional shifts, the scaling becomes inconsistent and identity is lost.","fun_headline_variants_meta":{"raw":{"variants":["One photo and text generate human motion videos","No pose video needed: text controls motion video","Fleximo turns a single photo and prompt into video","Text as the sole guide for human motion video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1676,"prompt_tokens":1031,"completion_tokens":645,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":586}},"tokens_in":647,"tokens_out":645,"duration_ms":5976,"temperature":1.0,"reasoning_tokens":586,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:09:52.816484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Fleximo on a motion with large positional shift (e.g., \"A person is playing basketball\" with side-facing movement) and measure frame-to-frame variation of the per-bone scaling factors and face identity similarity against the reference image. If the scaling variance is high and identity similarity drops sharply relative to smaller motions, the anchor-point rescale assumption is violated and the central claim of consistent high-quality generation fails.","supporting_citations":[{"cited_title":"Generating human motion from textual descrip- tions with discrete representations","cited_arxiv_id":null,"evidence_quote":"Supplies the text-to-3D motion generator that turns input text into 3D mesh vertices, the upstream of the whole pipeline."}],"review_version":1}