{"id":"eb10588d-bd6f-48e5-bbcd-7088e8f404f9","arxiv_id":"2608.09143","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UniMoFlow is a single flow-matching model for both text-to-motion generation and instruction-driven editing, trained with a 55,641-triplet synthetic dataset and a source-anchored sampling mode.","lead":"UniMoFlow is a system that edits 3D human motion from text instructions, such as changing arm movements or adding a hop midway through a walk, while keeping the rest of the motion intact. It also introduces a large synthetic dataset of 55,641 edit examples and an inference method that makes editing strength adjustable at test time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dataset construction and all editing metrics share the same SnapMoGen evaluator, and UniMoFlow trains in the same latent space, so the headline edit gains may reflect distribution matching rather than instruction following; an independent evaluator/human validation is needed.","rationale":"I agree with the reader's weakest assumption and find it load-bearing. The evaluator circularity is structural: the same SnapMoGen text-motion embedding is used for (i) filtering synthesized targets, (ii) defining every editing metric, including TR@k, Match, Struct, Region, PosRatio, and FID, and (iii) implicitly shaping UniMoFlow's training distribution through the shared latent space and flow objective. A model that reproduces the SnapMoGen base-DiT/FlowEdit output distribution is therefore expected to score well without necessarily following instructions better. The comparison is still fair across baselines in the narrow sense that all were trained on the same triplets and scored with the same metrics, but the benchmark's validity as a measure of instruction-following is not established. The paper's own conclusion flags the dependence on automatic filtering and future human preference signals, which corroborates this concern. The generation claims are more independently grounded: Table 4 uses the SnapMoGen test split and compares against published methods, and the architecture/ablations are reported carefully. I do not see an internal contradiction that would justify rejection; rather, the central editing claim needs an external validity check before it can be accepted as stated. The reader's CONDITIONAL verdict is therefore appropriate, and my stress-test does not change it.","tokens_in":20939,"tokens_out":4305,"duration_ms":44818,"concrete_test":"Recompute the Table 2/3 comparisons using an independently trained evaluator (e.g., TMR trained on HumanML3D) instead of the SnapMoGen evaluator, on the same 5,835 Omni-MoEdit test pairs. If UniMoFlow's TR@1/GT R@1 advantage over MotionLab drops below the reported margin (0.6347 vs 0.4839; 0.6762 vs 0.5197) or reverses, the central edit claim is an artifact of the self-referential evaluator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The editing claim rests on a closed loop: Omni-MoEdit targets are synthesized by a SnapMoGen-trained base DiT with FlowEdit (§3, Fig. 2), filtered by SnapMoGen evaluator thresholds (Match≥0.6, R@1≥0.7, Δmatch≥0.1, Struct≥0.4), and then scored in that same evaluator's embedding space via Eqs. 9–29. UniMoFlow is trained on these filtered targets in the same Wan-VAE latent space and with the same flow-matching paradigm as the synthesis backbone. Consequently, the large gains in Tables 2–3 may measure proximity to the SnapMoGen base-DiT/target distribution, which UniMoFlow is architecturally predisposed to reproduce, rather than genuine instruction-following; prior editors trained on the same triplets but in different motion spaces are doubly disadvantaged. The SAFE strength w=1.0 (Table 7) is another unresolved choice, since no validation split is named. This does not invalidate the generation results (Table 4), which use an independent SnapMoGen test split, but the central 'substantially outperforms prior editors' claim is not yet supported. The conclusion itself concedes the dependence on automatic filtering and defers human preference signals to future work, so the concern is acknowledged in-text rather than hypothetical.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniMoFlow, a unified latent flow-matching model that handles both text-to-motion generation and instruction-driven motion editing in a single self-attention context, together with a synthetic dataset Omni-MoEdit and an inference-time refinement scheme SAFE. The dataset is built by using Qwen3-8B to generate edit instructions and target descriptions from SnapMoGen captions, synthesizing target motions with a base DiT and FlowEdit, and filtering candidates with a frozen SnapMoGen evaluator. UniMoFlow is trained jointly on generation data and Omni-MoEdit triplets, with source-motion tokens placed in the same attention context as text and noisy target tokens. SAFE initializes from the source latent and uses the difference between instruction-conditioned and null-text velocity fields to steer the edit. Experiments report substantial improvements over supervised and training-free editing baselines on target-text alignment, source preservation, edit localization, and cycle consistency, while maintaining competitive text-to-motion generation quality on the SnapMoGen test split.","tokens_in":21199,"tokens_out":4825,"duration_ms":45250,"significance":"If the central claim is supported, the paper makes a useful contribution: it offers a principled way to scale editing supervision from synthetic triplets, a compact architecture that unifies generation and editing, and a controllable inference-time editing mechanism. The manuscript is also strong on experimental hygiene in several respects: the baselines are reproduced from released code with a common data protocol, the ablations isolate the contributions of mixed training, latent flow matching, token-level tags, and attention masks, and the new metrics are defined explicitly in the supplementary material. The generation results in Table 4 are evaluated on an independent SnapMoGen test split and are not affected by the editing-evaluator concern. However, the main editing comparison is currently evaluated with the same frozen SnapMoGen evaluator that was used to filter the synthetic training data, and UniMoFlow is trained in the same latent space and flow-matching paradigm as the synthesis backbone. This creates a real risk that the headline gains in Tables 2 and 3 measure proximity to the synthesis pipeline rather than instruction-following.","major_comments":[{"comment":"The headline editing comparison is a closed loop. Candidate triplets are retained only when the frozen SnapMoGen evaluator reports Match ≥ 0.6, R@1 ≥ 0.7, Δmatch ≥ 0.1, and Struct ≥ 0.4; the same evaluator's embedding space supplies every editing metric in Tables 2 and 3 through Eqs. (9)–(29); and UniMoFlow is trained on the retained targets in the same Wan-VAE latent space and flow-matching paradigm used by the synthesis backbone. The large SAFE/Native gains over MotionLab, TMED, and OmniME may therefore reflect proximity to the SnapMoGen base-DiT target distribution rather than genuine instruction-following. The baselines are additionally disadvantaged because they operate in different motion representations (e.g., 28-joint windows for MotionReFit and raw 296-D features for MotionLab), so they cannot benefit from the same latent-space prior. The generation results in Table 4 are not affected by this concern, but the central claim that UniMoFlow 'substantially outperforms prior editors' is not yet supported. I request at least one genuinely independent validation: human preference ratings on a sampled test set, an off-the-shelf text–motion retriever not used in filtering, or a zero-shot probe on MotionFix-style human-annotated pairs. The conclusion's own concession that the system 'still depends on automatic filtering' makes this a required validation rather than an optional extra.","section":"Dataset synthesis and evaluation protocol (Fig. 2, Eqs. 9–29)"},{"comment":"The main SAFE results use w = 1.0, which is selected from the sweep in Table 7 on Omni-MoEdit with no named validation split. If this sweep was performed on the test split that is also used for Tables 2 and 3, the reported SAFE numbers are the best of five hyperparameter settings and the comparison is optimistic. Please either perform the w sweep on the held-out validation split and freeze w before test evaluation, or report Tables 2 and 3 for a range of w values (e.g., 0.5, 1.0, 2.0) so the reader can assess the sensitivity of the main claims. A similar issue applies to the generation results: the CFG scale 4.5 in Table 4 is selected on the SnapMoGen test split according to Figure 7. The generation comparison should use a validation-based selection, or the authors should justify why test-set selection does not bias the comparison against the baselines.","section":"SAFE drift strength (Table 7) and generation CFG (Fig. 7)"},{"comment":"The paper states that the residual v_c − v_u 'distills the velocity component attributed to the edit instruction while remaining anchored to the source motion.' This is an important claim for SAFE, but it is not formally established or empirically isolated. Table 7 shows that larger w degrades FID and retrieval, which is consistent with drift away from the source manifold, but there is no analysis of how the residual behaves across edit types, motion lengths, or source-motion complexity. I would like to see a quantitative decomposition, for example measuring source fidelity and edit success separately for localized edits versus structural edits, to support the claim that SAFE is complementary to native editing rather than simply a weaker version of it.","section":"SAFE mechanism (Eqs. 7–8)"}],"minor_comments":[{"comment":"The sentence 'For zero-shot or training-free editing, we mainly use such samplers to synthesize candidate pairs with Qwen3-8B' is confusing about whether training-free editors are used only for data synthesis or also as evaluation baselines; please clarify.","section":"Introduction and Related Work"},{"comment":"Equation (5) has unbalanced delimiters in the expectation terms; please fix the typesetting.","section":"Equation (5)"},{"comment":"The paragraph on source-anchored null-text fields ends with 'Section .' and a blank cross-reference; please fill in the correct section number.","section":"SAFE section"},{"comment":"The abbreviations 'UniMoFlow (S)' and 'UniMoFlow (SAFE)' are used inconsistently across Tables 2 and 3; please unify them with the main text.","section":"Tables 2 and 3"},{"comment":"The statement that 'only the first 148 motion channels are used' for the Struct metric requires a citation to the evaluator implementation or a specification of which channels are included; otherwise the metric cannot be reproduced independently.","section":"Evaluation metric definitions (Eq. 16)"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is interesting and the experiments are extensive, but the closed-loop evaluation is the main risk. I would recommend requiring an independent evaluation or human study before acceptance, and also requiring the authors to clarify the validation-split discipline for the SAFE strength and CFG hyperparameters. The paper is likely salvageable with a major revision rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real empirical effort, not a toy: Omni-MoEdit is 55k+ triplets with source and target captions and reverse instructions, and UniMoFlow is a genuinely unified architecture with in-context source conditioning and a continuous SAFE control knob. Second, the headline editing results should be read with caution: the targets are synthesized by a SnapMoGen-trained DiT, filtered by the SnapMoGen evaluator, and scored in that same evaluator's embedding space, so the large gains over prior editors may partly reflect distribution matching rather than instruction following.\n\nWhat's good: the ablations are extensive and fair—unified training helps, token-level tags and latent flow matching each earn their keep, and the baselines are reproduced on the same triplets rather than warmed over with their original datasets. The generation results on the SnapMoGen test split are more independent and show the unified model is at least competitive with strong continuous-space generators. The new metrics (PosRatio, Region, Cycle-Con) are clearly defined and show the authors thought about the failure modes of earlier evaluation.\n\nSoft spots, in order. One: the closed loop. This is structural, not a detail. Until the edits are scored by an evaluator not derived from the synthesis backbone, or by human raters, the 'substantially outperforms' claim is conditional. The paper's own conclusion concedes the dependence on automatic filtering, which is honest but doesn't fix the bias. Two: SAFE's w=1.0 is chosen from a sweep in Table 7 with no stated validation split; that is hyperparameter tuning on the test set, albeit for one scalar. Three: the code, data, and checkpoints are promised but not shipped; for a dataset-driven claim, that's a real gap.\n\nNet: this deserves a serious referee. The right revision path is to (i) add an independent or human-validated edit evaluation, (ii) report w selection on validation, and (iii) release the artifacts. I'd be disappointed if the editing claims are taken at face value before that, but the core architecture and dataset are substantial contributions. I'd send it out.","headline":"A serious, well-ablated unification of generation and editing for 3D human motion, with a large synthetic dataset—but the headline editing gains rest on a self-referential benchmark and need independent validation.","tokens_in":21777,"tokens_out":2640,"would_cite":true,"duration_ms":24558,"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":"The paper claims that grounding 3D motion editing in generation via a unified flow-matching model with full source context improves target alignment, edit effectiveness, and cycle consistency over prior editors while keeping generation…","keywords":["3D human motion editing","text-to-motion generation","instruction-driven editing","flow matching","latent transformer","synthetic dataset","source-anchored sampling","motion evaluation metrics"],"falsifier":"Take a random sample of 500 Omni-MoEdit test triplets stratified by the five edit types and have independent human raters judge whether the synthesized target motion both realizes the edit instruction and preserves all unspecified source content; if the evaluator's acceptance thresholds (Matching Score ≥ 0.6, R@1 ≥ 0.7, DeltaMatch ≥ 0.1, Structure ≥ 0.4) pass pairs that human raters consistently reject, or if agreement differs by edit type, the dataset and all editing metrics built on that evaluator inherit the bias.","tokens_in":20678,"feed_emoji":"🏃","tokens_out":9457,"duration_ms":78454,"temperature":0.7,"pith_summary":"The paper tries to establish that instruction-driven editing of 3D human motion can be made a native capability of a text-to-motion generator instead of a separate trained editor. To do this it builds Omni-MoEdit, a 55,641-pair synthetic dataset in which a language model turns motion captions into edit instructions, a text-to-motion DiT synthesizes target motions through FlowEdit, and a frozen evaluator filters for target-text alignment and source preservation. Its model UniMoFlow is a latent flow-matching transformer—a generative model that learns a velocity field from noise to data—that places edit text, noisy target tokens, and the full source-motion token sequence in one self-attention context, with mode-specific masks switching between generation and editing. A source-anchored sampling mode, SAFE, edits by integrating the difference between instruction-conditioned and null-text velocity fields while starting from the source latent. If the paper is right, one compact model can generate and edit human motion, and large synthetic triplets can replace scarce manual edit annotations.","feed_headline":"One flow model generates and edits 3D human motion","feed_subtitle":"Training on 55K synthetic edit pairs plus source-anchored sampling improves target alignment and cycle consistency over prior editors.","key_machinery":"The central object is a unified latent flow-matching transformer with in-context conditioning. Source and target motions are compressed by a causal VAE—a temporal autoencoder that turns every four motion frames into one continuous latent token—and a frozen T5 encoder supplies text tokens; at editing time the context sequence is noisy target, edit text, separator, and full source, with time-conditioned additive modality tags and a mode-specific global self-attention mask letting the same backbone act as generator or editor. The model is trained with rectified-flow interpolation in a length-dependent shifted time, with separate losses on edit triplets and generation samples, and text dropout that retains the source condition, which produces the null-text velocity field SAFE exploits at inference. SAFE starts from the source latent and takes steps along gated differences between instruction-conditioned and null-text velocities, with a strength parameter that trades source fidelity against edit magnitude without retraining. The accompanying Omni-MoEdit pipeline is the other load-bearing mechanism: a language model generates commands and target captions, a base DiT synthesizes targets via FlowEdit, and thresholds on Matching Score, R@1, DeltaMatch, and Structure filter the triplets.","core_discovery":"The central claim is that grounding editing in generation works across data, architecture, and inference: a model trained jointly on text-to-motion generation and Omni-MoEdit edit triplets learns to localize and apply body-part, amplitude, temporal, action, and style changes while preserving unedited structure. UniMoFlow (SAFE) raises target-text retrieval at rank 1 from 0.4839 to 0.6347 and improves edited-vs-ground-truth retrieval from 0.5197 to 0.6762 relative to the strongest prior editor in its comparison, while cutting FID from 26.06 to 12.45; UniMoFlow also reports the best FID (15.331) and Matching Score (0.716) among continuous-space generators on the SnapMoGen test split. The authors interpret these results as showing that full token-level source access plus a shared generative prior outperforms compressed source conditioning and training-free adaptation, and that native editing and source-anchored sampling cover complementary regimes, with SAFE suited to localized refinements and the native editor more reliable for large structural reconstructions.","pith_inferences":["As an extension beyond the paper, the editing-quality ceiling is likely set by the caption diversity and the evaluator's embedding, so swapping the source corpus or the evaluator would change which edit types succeed.","As an extension beyond the paper, any flow-matching generator trained with source-conditioned null-text dropout could be turned into an editor with the SAFE recipe, which suggests a direct transfer test to image or audio editing.","As an extension beyond the paper, the reverse instructions in Omni-MoEdit could serve as a cycle-consistency training loss, not just an evaluation metric."],"forward_implications":["Large-scale instruction-driven motion editing can be supervised without manually curated triplets: the synthesis-and-filter recipe yields 55,641 pairs covering body-part, amplitude, temporal, action, and style edits, about 3.5 times larger than prior adjustment sets.","Unified generation-editing training is mutually beneficial: generation batches regularize residual artifacts in synthetic edits, and edit batches strengthen language-motion grounding, so a single checkpoint serves both tasks.","Source-anchored inference gives a continuous edit-strength dial: adjusting the drift weight controls how far the output moves from the source toward the instruction, without retraining, which suits interactive animation workflows.","Semantics-aware metrics matter: reporting positive improvement ratio, region alignment, and cycle consistency alongside reconstruction metrics separates realizing the requested change from preserving the source, so under-editing no longer masquerades as success.","Edits that require drastic structural reconstruction remain the hard regime, with the paper stating that the native editor is more reliable there, while SAFE is preferred for fine-grained adjustments."],"supporting_citations":[{"why":"Supplies SnapMoGen source captions, base text-to-motion motions, and the frozen evaluator used to filter Omni-MoEdit and compute all editing and generation metrics.","marker":"Guo et al. 2025a"},{"why":"The Qwen3-8B model that generates edit commands, target captions, and reverse instructions from source captions.","marker":"Yang et al. 2025a"},{"why":"FlowEdit, the inversion-free editing sampler used to synthesize target motions from the base DiT conditioned on the target description.","marker":"Kulikov et al. 2025"},{"why":"MotionFix, which defines the instruction-driven triplet editing task and supplies the TMED baseline and architecture starting point.","marker":"Athanasiou et al. 2024"},{"why":"InstructPix2Pix, whose synthetic-pair amortization principle motivates training the final editor in one conditional forward pass.","marker":"Brooks, Holynski, and Efros 2023"},{"why":"DiT, the diffusion-transformer backbone adapted for the base text-to-motion synthesis model and UniMoFlow.","marker":"Peebles and Xie 2023"},{"why":"Wan, whose causal video VAE is adapted into the 1D causal VAE providing the 32-D continuous latent space used by both generation and editing.","marker":"Wan Team et al. 2025"}],"fun_headline_variants":["Flow model unites 3D motion generation and editing","Grounding editing in generation boosts 3D motion alignment","SAFE: source-anchored refinement for 3D motion editing","Training on synthetic edit pairs improves 3D motion control","UniMoFlow: generate and edit 3D motion with one model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline treats the frozen evaluator's similarity scores as ground truth for both 'the target motion realizes the instruction' and 'the source is preserved', so if that embedding space is biased or blind to some edit categories, the dataset and every reported editing comparison inherit the bias.","fun_headline_variants_meta":{"raw":{"variants":["Flow model unites 3D motion generation and editing","Grounding editing in generation boosts 3D motion alignment","SAFE: source-anchored refinement for 3D motion editing","Training on synthetic edit pairs improves 3D motion control","UniMoFlow: generate and edit 3D motion with one model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000438,"raw_usage":{"total_tokens":2249,"prompt_tokens":992,"completion_tokens":1257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1169}},"tokens_in":608,"tokens_out":1257,"duration_ms":10871,"temperature":1.0,"reasoning_tokens":1169,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:40:10.115228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of 500 Omni-MoEdit test triplets stratified by the five edit types and have independent human raters judge whether the synthesized target motion both realizes the edit instruction and preserves all unspecified source content; if the evaluator's acceptance thresholds (Matching Score ≥ 0.6, R@1 ≥ 0.7, DeltaMatch ≥ 0.1, Structure ≥ 0.4) pass pairs that human raters consistently reject, or if agreement differs by edit type, the dataset and all editing metrics built on that evaluator inherit the bias.","supporting_citations":[{"cited_title":"and Varol, G","cited_arxiv_id":null,"evidence_quote":"MotionFix, which defines the instruction-driven triplet editing task and supplies the TMED baseline and architecture starting point."}],"review_version":1}