{"id":"5dec507c-b439-4b6c-bbca-04fdaab38370","arxiv_id":"2508.07409","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new pipeline generates pose-controlled, view-consistent 4D character animations from one reference image and a 2D pose sequence, backed by a new 13,115-character dataset and benchmark.","lead":"CharacterShot turns a single character image and a 2D pose sequence into a 4D character animation, using a pose-conditioned video model, multi-view generation, and 4D Gaussian splatting. It also introduces a 13,115-character dataset and a benchmark, and reports strong gains over prior 4D generation methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"4D consistency is only evaluated against the generator's own multi-view videos (Table 4), so the claimed consistency may be an artifact of the closed loop rather than true 3D consistency; no independent 3D ground truth is used.","rationale":"The paper is a solid systems contribution: the dual-attention module, Character4D dataset, and neighbor-constrained 4DGS are reasonable and internally consistent; the ablations support the design choices. The 30-participant user study is real evidence for OOC generalization, though small. My main concern is the validation loop for the central claim. Table 4 explicitly uses the authors' own generated multi-view videos as ground truth, so the 4D stage is optimized and evaluated against the same signal. SSIM/LPIPS/FVD compare 2D image statistics and cannot certify spatial-view geometric consistency; a generator that hallucinates slightly different geometry in each view would still score well if the 4D representation reproduces those views. The end-to-end SOTA comparison (Table 2) may be less circular if it uses real rendered ground truth, but the paper does not clearly state the reference, and the baseline input mismatch confounds the comparison. The proposed concrete test—novel-view geometry comparison against true Character4D meshes with error bars—would settle whether the claimed consistency is real. This is an addressable weakness, not a fatal flaw, so the reader's CONDITIONAL verdict stands unchanged.","tokens_in":18411,"tokens_out":10777,"duration_ms":119033,"concrete_test":"Use the held-out Character4D test split, which has true 3D meshes, camera poses, and rendered multi-view videos. Run the full CharacterShot pipeline from a single reference image and pose sequence, then compare the optimized 4D Gaussians to the ground-truth mesh/motion rather than to the generator's pseudo-GT: render novel views not used during 4D optimization and compute LPIPS/SSIM against the corresponding ground-truth renders, and extract per-frame depth or mesh to compute Chamfer distance and F-score against the true mesh. Run with at least 3 seeds and report mean±std. If geometric error is high, or if using ground-truth multi-view videos as 4DGS input substantially outperforms using the generator's videos, the closed-loop evaluation in Table 4 is masking the actual consistency ceiling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is end-to-end 4D character animation with spatial-temporal and spatial-view consistency. The load-bearing condition is that the multi-view videos produced by the dual-attention generator (Sec. 3.3) are geometrically consistent enough to act as trustworthy supervision for the 4DGS stage (Sec. 3.4). That condition is not actually tested. Table 4, the only quantitative evaluation of the 4D optimization, states 'Ground truths are generated multi-view videos.' Because the same generated videos are both the supervision for the 4DGS and the reference for SSIM/LPIPS/FVD-F/FV4D, the metrics primarily measure how faithfully the 4D representation overfits to those particular views. They cannot detect view-inconsistent geometry, e.g., a facial feature that the generator places slightly differently in two views; 4DGS can absorb that inconsistency as blur or a float, and the score against the same pseudo-GT can still be high. The end-to-end SOTA comparison (Table 2) may be subject to a related confound: baselines are driven by a separately fine-tuned MimicMotion single-view video, so the margin could reflect input-video quality rather than the proposed 4D representation. Table 7 actually shows that replacing the generated single-view input with ground-truth video substantially improves L4GM, confirming that input quality dominates. No error bars or significance tests are reported for any table, so it is unknown whether the large-looking gaps are stable. Consequently, the 'consistent 4D' claim is currently supported only by a closed evaluation loop.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CharacterShot, a framework for 4D character animation from a single reference character image and a 2D pose sequence. The method consists of three stages: (i) fine-tuning the DiT-based image-to-video model CogVideoX with pose conditioning to obtain a 2D character animation model; (ii) extending it to multi-view video generation via a dual-attention module and camera priors, with a separately fine-tuned SV3D view generator; and (iii) optimizing a neighbor-constrained 4D Gaussian Splatting representation from the generated multi-view videos. The authors also introduce Character4D, a large dataset of 13,115 VRoidHub characters with rigged motions and multi-view renders, and CharacterBench, a benchmark built from this dataset plus out-of-distribution characters. Quantitative comparisons on CharacterBench claim consistent SOTA improvements over SV3D, SV4D, Diffusion2, STAG4D, SC4D, L4GM, and DG4D across SSIM, LPIPS, CLIP-S, FVD variants, and FV4D. The paper includes ablations of the dual-attention module, the neighbor-constrained 4DGS, and the dataset contribution, plus a 30-participant user study for out-of-Character4D characters.","tokens_in":18826,"tokens_out":2935,"duration_ms":31201,"significance":"If the claims hold, CharacterShot would be a practically valuable system: it turns a single character image and a pose sequence into a view-consistent, temporally coherent 4D representation in tens of minutes without hand modeling or motion capture. The proposed large-scale character dataset and benchmark could also be useful resources for the community. The dual-attention design is a reasonable and potentially effective way to couple spatial-temporal and spatial-view information, and the neighbor constraint for 4DGS is a sensible robustness measure. However, the current evidence for the central claim of 3D/4D consistency is weakened by a substantial evaluation-circularity issue: the 4D optimization stage is evaluated against the same generated multi-view videos that supervise it, and all quantitative comparisons lack error bars or significance tests. These are fixable through additional evaluation against independent 3D ground truth and proper statistical reporting, so the paper's core idea is not invalidated, but the strength of the empirical claims is currently disproportionate to the evidence.","major_comments":[{"comment":"The only quantitative evaluation of the 4D optimization stage uses pseudo ground truth: Table 4's caption states 'Ground truths are generated multi-view videos.' Because these same generated videos are the supervision for the 4DGS optimization and the reference for SSIM/LPIPS/FVD-F/FV4D, the metrics largely measure how well the 4D representation reproduces the generator's own outputs, not true 3D consistency. Cross-view inconsistencies in the generator could be absorbed by the 4D representation as blur or floaters while still scoring high against the same pseudo-GT. Please evaluate the 4D stage against independent 3D ground truth from the Character4D test set (e.g., held-out characters with ground-truth multi-view renders, and optionally mesh/geometry metrics such as Chamfer distance or normal consistency), or at minimum report per-sample metrics that separate view-consistency from frame","section":"Table 4, Sec. 3.4/4.3"},{"comment":"The end-to-end comparison in Table 2 is confounded by the input presented to each baseline: SV4D, Diffusion2, STAG4D, SC4D, L4GM, and DG4D are driven by single-view videos generated by a separately fine-tuned MimicMotion, while CharacterShot receives the reference image and pose sequence. Table 7 confirms that L4GM's performance is sensitive to input video quality (SSIM 0.904 with original, 0.907 with fine-tuned, 0.916 with ground-truth single-view video). This shows that a non-trivial part of the observed margin may reflect the quality of the generated single-view input rather than the proposed 4D representation. Please provide an ablation that gives the baselines access to ground-truth single-view videos (or otherwise controls for input quality) and report the comparison in that condition, alongside the current setup.","section":"Table 2 and Table 7, Sec. 4.2/B.2"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any quantitative result. While the reported margins are large, the absence of any statistical characterization makes it impossible to assess stability across test samples, especially given that the benchmark is constructed by the authors from their own dataset. Please report standard deviations or confidence intervals over CharacterBench samples, and where feasible run a paired significance test (e.g., Wilcoxon signed-rank) for the primary comparisons. This is particularly important for the 4D optimization comparison in Table 4, where the pseudo-GT setup already narrows the gap between methods.","section":"Tables 1-8, Sec. 4"}],"minor_comments":[{"comment":"The neighbor loss uses the symbol u_t_i, but u is not defined before its use. Please define it (presumably the position of Gaussian i at time t) and clarify the dimension of L_t_i.","section":"Eq. (4)-(6), Sec. 3.4"},{"comment":"There is a typo: 'Objverse' should be 'Objaverse' in Section 2.2 context; the reference [8] is Objaverse. Also, 'followed [84]' in Sec. 3.4 should be 'following [84]'.","section":"Sec. 3.5"},{"comment":"The caption says 'G-Mesh' but the text and references use 'DG-Mesh'. Please make the naming consistent.","section":"Fig. 8"},{"comment":"Minor typo: 'sequnce' should be 'sequence' in the abstract. Also, the claim 'first DiT-based 4D character animation framework' should be checked against Human4DiT (cited as [57]), which is also DiT-based and handles 4D human generation; the distinction should be stated clearly (e.g., 'from a single reference image and 2D pose sequence') to avoid overclaiming.","section":"Abstract & Sec. 1"},{"comment":"The user study uses 30 volunteers, which is small. Please report the exact question format, the standard deviation across participants, and whether the differences are statistically significant. Also, a breakdown by character type (anime vs. real vs. AI-generated) would strengthen the out-of-distribution generalization claim.","section":"Appendix B.3, Table 8"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clearly stated and well-motivated pipeline, and the proposed dataset and benchmark could be useful. The main risk is that the empirical evidence for the central 3D/4D consistency claim is currently circular in the 4D stage and confounded in the end-to-end comparison. These are load-bearing but fixable with additional experiments and statistical reporting. I would not recommend rejection, as the method itself appears sound and the issues are within the scope of a revision. However, the authors should be pushed to provide independent 3D ground-truth evaluation for the 4D stage and to control for input-quality confounds before the SOTA claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what to know: this is a competent and useful systems paper. The task combination—single character image plus 2D pose sequence to 4D character—is genuinely new relative to SV4D, Human4DiT and friends. The dual-attention module is a sensible design choice for multi-view video generation, and the neighbor-constrained 4DGS loss is a practical fix for the misalignment artifacts that plague optimization from generated views. The Character4D dataset, 13k VRoid characters with rigged motions and 21 viewpoints, is a real resource, and the benchmark built from it is reasonable for relative comparisons.\n\nThe paper does its component-level homework: ablations for camera prior, dual-attention vs. view attention, and the neighbor loss are all present and show consistent improvements. The view generator comparison in Table 5 against Zero123XL, InstantMesh, Hi3D, and SV3D is a legitimate evaluation with ground-truth rendered views, and the margins there are credible.\n\nThe soft spot is exactly where the stress test lands. Table 4, the only quantitative evidence for the neighbor-constrained 4DGS and for 4D consistency, states that ground truths are the authors' own generated multi-view videos. That is a closed loop: the 4D stage is trained on those videos and then scored against the same videos, so the metrics mostly measure overfitting to the generator's idiosyncrasies, not true geometric consistency. A facial feature that appears in slightly different positions across views can be absorbed by 4DGS as blur or floaters and still score well against the pseudo-GT. This doesn't invalidate the neighbor constraint—the visual results in Fig. 8 show fewer artifacts than the baselines when all are optimized on the same pseudo-GT—but it cannot support the claim of 'spatial-view consistency' in an absolute sense. Table 7 is honest in showing that better input video helps L4GM, which actually supports the worry that input quality drives a big part of the margin in Table 2.\n\nThree more things, smaller but real. No error bars or significance tests anywhere, so the large-looking margins in Tables 1 and 2 are of unknown stability. The limitation section mentions 'confidence-aware pose guidance' as though it were part of the method, but the method section never describes it—that looks like an unfinished edit. And code, models, and datasets are promised but not released, which matters because the benchmark is self-built and the pseudo-GT evaluation can be checked only with the artifacts.\n\nFor a reader: anyone working on 4D generation or character animation gets value from the dataset, the benchmark, and the dual-attention idea. It deserves a serious referee, and a revision that adds independent 3D geometric evaluation, error bars, and the artifacts would make the SOTA claim credible. Recommend sending to peer review.","headline":"Solid systems paper with a genuinely new task setting, but the headline consistency numbers are measured against the model's own outputs, so the SOTA claim is not yet earned.","tokens_in":19329,"tokens_out":2339,"would_cite":true,"duration_ms":22767,"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":"CharacterShot claims that a single reference character image and a 2D pose sequence, run through a pose-conditioned diffusion transformer and a neighbor-constrained 4D Gaussian-splat reconstruction, can yield a controllable, view-consistent","keywords":["4D character animation","pose-guided animation","diffusion transformer","multi-view video generation","4D Gaussian splatting","single-image animation","temporal consistency","character dataset"],"falsifier":"Run CharacterShot on a character from Character4D whose true 3D mesh and motion are known, then render the optimized 4D Gaussians at a held-out camera viewpoint and time that were not used in the multi-view video set, and measure image distance against a rendering of the true mesh. If the held-out novel-view error is close to the reported training-view numbers, the consistency is real; if it degrades sharply, the consistency was an artifact of optimizing against self-generated videos.","tokens_in":18358,"feed_emoji":"🎬","tokens_out":9875,"duration_ms":87027,"temperature":0.7,"pith_summary":"CharacterShot sets out to prove that a moving, viewable-from-any-angle 3D character—a 4D character animation—can be produced from a single reference character image and a 2D pose sequence, without modeling, rigging, or motion capture. The paper's approach is to make a diffusion-transformer image-to-video generator pose-controllable, extend it to generate multiple synchronized viewpoints with a dual-attention module and camera conditioning, and then reconstruct a temporally stable 4D Gaussian-splat representation from those videos. The payoff, if the claims hold, is that character animation drops from a weeks-long CGI pipeline to roughly an hour of GPU computation that an individual creator can run. To support this, the authors introduce a 13,115-character training set and a benchmark, CharacterBench, and report consistent gains over prior methods on multi-view and 4D quality metrics.","feed_headline":"One character image + 2D poses yield a consistent 4D character","feed_subtitle":"One image and a pose sequence become a view-consistent 4D character in about an hour.","key_machinery":"The dual-attention module: it processes video tokens in two arrangements at once—frames grouped within each view and views grouped within each frame—using parallel full-attention blocks, so spatial-temporal and spatial-view coherence are learned jointly rather than through separate attention mechanisms. The other load-bearing piece is neighbor-constrained 4D Gaussian splatting (4DGS): a coarse-to-fine optimization of deformable 3D Gaussians over time, where a neighbor loss compares each point's displacement from its local group center between consecutive frames and is activated by a binary gate only when a point moves beyond a threshold. That gate stops noisy outlier points from dragging loc","core_discovery":"CharacterShot's central claim is that a dynamic 3D character can be generated directly from one reference image and a 2D pose sequence, bypassing the usual CGI chain of modeling, rigging, and motion capture. The paper builds this around a diffusion-transformer video model that is made pose-controllable in a pretraining stage, then extended to generate several synchronized viewpoints through a dual-attention module and a camera prior. The multi-view videos are lifted into a 4D representation by a coarse-to-fine Gaussian-splatting optimization with a neighbor constraint designed to suppress view-to-view outliers. With these pieces, the authors report that CharacterShot preserves the character'","pith_inferences":["The paper does not state this, but if the multi-view generator produces systematic artifacts, the 4D benchmark's pseudo-ground-truth comparison will hide them; an evaluation on independent 3D scans or held-out rendered views is the natural next check.","The paper does not state this, but the pose-conditioned 2D animator is a standalone product: it can animate any character image with any pose sequence even if the 4D reconstruction stage is never run.","The paper does not state this, but the same pipeline could be driven by real-time skeleton tracking to make an animatable avatar, provided the generation latency drops from tens of minutes to near-interactive speeds.","The paper does not state this, but the out-of-distribution generalization claim is primarily perceptual rather than geometric, because the benchmark mixes in-house rendered characters with internet and model-generated images rated by 30 human volunteers."],"forward_implications":["Pose-controlled 4D animation becomes accessible to non-specialists: the reported pipeline takes 20–40 minutes for multi-view video generation plus about 30 minutes for 4D optimization on one H800 GPU, and needs only a character image and a skeleton sequence as input.","Any existing 2D pose data—from dance videos, motion libraries, or manual keyframing—can be retargeted onto arbitrary characters, since the motion signal is just a 2D pose sequence.","Because the dual-attention module learns spatial-temporal and spatial-view consistency jointly, the same animated character can be rendered from arbitrary viewpoints without per-view training.","The neighbor constraint specifically targets artifacts like disappearing hands and outlier 3D points, which earlier 4D-Gaussian methods exhibit when optimized on generated multi-view videos."],"supporting_citations":[{"why":"Supplies the pretrained DiT-based image-to-video model that is fine-tuned for pose-controlled 2D character animation.","marker":"[87]"},{"why":"Serves as the view generator and camera-prior formulation that is fine-tuned on Character4D for novel-view characters.","marker":"[70]"},{"why":"Defines the multi-view video generation paradigm and FV4D evaluation protocol CharacterShot extends and is compared against.","marker":"[81]"},{"why":"Provides the coarse-to-fine progressive 4D Gaussian fitting and total-variation loss that the neighbor-constrained 4DGS is built on.","marker":"[84]"},{"why":"Establishes the pose-guided 2D character animation baseline that is fine-tuned to generate single-view inputs for fair comparison.","marker":"[96]"},{"why":"Gives the data-curation pipeline used to build Character4D from rigged characters and retargeted motions.","marker":"[75]"},{"why":"Supplies the 13,115 character models from which Character4D is constructed.","marker":"[71]"},{"why":"Provides the whole-body pose detector used to extract pose sequences for training and inference.","marker":"[88]"}],"fun_headline_variants":["Single image and pose sequence yield a consistent 4D character","From one reference image to 4D character animation","One photo plus 2D poses becomes a 4D character","CharacterShot: 4D characters from one image and poses","A solo image and pose list create a 4D character"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Everything in the final 4D result is only as trustworthy as the multi-view videos produced earlier, and the paper's own benchmark treats those generated videos as the ground truth for the 4D stage, so errors introduced by the generator can become permanent without being measured.","fun_headline_variants_meta":{"raw":{"variants":["Single image and pose sequence yield a consistent 4D character","From one reference image to 4D character animation","One photo plus 2D poses becomes a 4D character","CharacterShot: 4D characters from one image and poses","A solo image and pose list create a 4D character"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":1844,"prompt_tokens":747,"completion_tokens":1097,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1022}},"tokens_in":491,"tokens_out":1097,"duration_ms":10379,"temperature":1.0,"reasoning_tokens":1022,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:07:45.967918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CharacterShot on a character from Character4D whose true 3D mesh and motion are known, then render the optimized 4D Gaussians at a held-out camera viewpoint and time that were not used in the multi-view video set, and measure image distance against a rendering of the true mesh. If the held-out novel-view error is close to the reported training-view numbers, the consistency is real; if it degrades sharply, the consistency was an artifact of optimizing against self-generated videos.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the multi-view video generation paradigm and FV4D evaluation protocol CharacterShot extends and is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the pose-guided 2D character animation baseline that is fine-tuned to generate single-view inputs for fair comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the whole-body pose detector used to extract pose sequences for training and inference."}],"review_version":1}