{"id":"4af94b99-bb67-4487-a483-14e66bbcc37a","arxiv_id":"2412.15609","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An interactive system combining user 2D edits, automatic pose suggestion, and re-training reduces artifacts in 3D Gaussian Splatting avatars, with moderate user-study evidence that it beats the SuperSplat editor.","lead":"3D avatar models built from a monocular video often show glitches under new poses. This paper describes a painting-based repair system with automatic pose suggestions, and reports a small user study and a crowdsourced preference test indicating it beats the SuperSplat editor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline quality claim rests on a tiny user study (n=4, no significance tests) with small metric gaps that could be noise; this is the most load-bearing threat to the central claim.","rationale":"The reader's weakest_assumption (visibility-artifact correlation) is a plausible concern about the pose-suggestion component, but it is not the most load-bearing threat to the central claim. Even if the visibility assumption failed, the 2D-editing-plus-retraining pipeline could still outperform SuperSplat; the central claim would survive. The claim's only support is the empirical comparison, and that comparison is statistically fragile: n=4 participants, 2 avatars, no significance tests, tiny metric differences, and a baseline restricted to canonical body pose and deletion-only editing (Section 7.1.3). The paper does report a crowdsourced preference (Section 7.4), but raw percentages without confidence intervals or tests are insufficient. In good faith, the system is well-motivated and clearly described; the pose-suggestion objective (Eq. 8) is a reasonable next-best-view formulation, and the limitations discussion is honest about multi-view consistency and loose clothing. However, the headline claim needs stronger empirical support. A preregistered, sufficiently powered paired study with an ablation condition would settle it. I therefore recommend keeping the CONDITIONAL verdict, with the central condition being the statistical reliability of the comparative evaluation rather than the visibility assumption per se.","tokens_in":17984,"tokens_out":8087,"duration_ms":76258,"concrete_test":"Run a preregistered within-subjects user study with at least 20 participants on the same two avatars, comparing three conditions: AvatarPerfect, AvatarPerfect with random pose suggestions, and SuperSplat. After each session, compute IoU, PSNR, and SSIM against the free-motion ground truth, and report paired bootstrap 95% confidence intervals for the mean differences. If the CI for AvatarPerfect minus SuperSplat includes zero on any primary metric, the headline claim fails; if the CI for AvatarPerfect minus random-pose AvatarPerfect includes zero, the pose-suggestion mechanism is not demonstrated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that AvatarPerfect yields higher-quality refined avatars than SuperSplat depends entirely on the empirical comparison in Section 7. With only 4 participants (all male), 2 avatars, and each condition cell averaging 2 refined results, Table 2 reports differences that are small in absolute terms (IoU 0.9094 vs 0.9115 and 0.8636 vs 0.8713; PSNR +0.4 dB; SSIM +0.008) with no significance test, confidence interval, or effect size reported anywhere in Section 7.3. The crowdsourcing preference (Table 4) is presented as raw percentages without statistical testing, and one of the 12 pairs favored SuperSplat. Given the small sample, these differences are plausibly explained by participant skill or random variation rather than by the system. Moreover, the baseline is restricted to a fixed canonical body pose and deletion-only editing (Section 7.1.3), so the comparison confounds the editing paradigm with the pose-suggestion contribution; the results do not isolate whether automatic pose suggestion is responsible for any improvement. If the observed gaps are noise, the paper's headline claim and its central novelty (pose suggestion) are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents AvatarPerfect, an interactive system for refining pre-trained 3D Gaussian Splatting (3DGS) avatars. The system renders the avatar under a suggested body/camera pose, lets the user paint over artifacts using background, inpaint, and diffusion-inpaint tools, and then updates the avatar via additional training with the original monocular video plus the edited images. A Gaussian-visibility-based optimization selects the next pose to show. The authors report a user study (4 participants) comparing AvatarPerfect with SuperSplat (deletion-only, fixed body pose) on two avatars using IoU, PSNR, and SSIM, plus a crowdsourced preference study (30 evaluators, 12 pairs). They conclude that AvatarPerfect enables users to obtain higher-quality refined 3DGS avatars than the existing 3DGS editor.","tokens_in":18256,"tokens_out":7136,"duration_ms":58430,"significance":"If the results hold, the paper offers a promising alternative to direct 3D Gaussian editing: a 2D-painting workflow with automatic pose suggestion. The system is clearly described, and the idea of using Gaussian visibility to guide the user to likely artifact locations is interesting. The manuscript also honestly discusses several limitations. However, the empirical evidence is preliminary: the main comparison has a very small sample, no statistical inference, and the baseline is restricted in a way that does not isolate the pose-suggestion contribution. The pose-suggestion objective also contains a technical ambiguity. With these issues addressed, the work could be a useful contribution to interactive 3D avatar refinement.","major_comments":[{"comment":"The headline claim of higher quality is based on means of two refined results per avatar per condition, with no variance, confidence intervals, or significance tests. The reported differences are small (e.g., IoU 0.9094 vs 0.9115 and 0.8636 vs 0.8713; PSNR +0.4 dB; SSIM +0.008). At n=4 participants and 2 avatars, these gaps are plausibly due to individual differences. The conclusion in Section 9 that the system 'significantly enhances' visual quality is not supported by the data. Per-participant results and an appropriate statistical analysis (or a clear pilot framing) are needed.","section":"Section 7.3, Table 2"},{"comment":"The comparison confounds the editing paradigm with the pose-suggestion contribution. SuperSplat is restricted to Gaussian deletion and a fixed canonical body pose; AvatarPerfect additionally offers inpainting/diffusion editing and additional training. The observed improvements could stem from those components rather than from the proposed automatic pose suggestion. An ablation (e.g., AvatarPerfect with random or user-chosen poses, or with manual camera-only pose selection) is required to substantiate the central novelty.","section":"Sections 7.1.3 and 7.3"},{"comment":"The objective function contains an undefined quantity \\bar{V} and, as written, does not directly reward displaying barely visible Gaussians. The term clip(V[i]-\\bar{V}) is zero for low-visibility Gaussians, so the sum only penalizes candidate views that show already-visible Gaussians, and it is indifferent to whether low-visibility Gaussians appear. This does not match the stated goal of 'display as many Gaussians as possible, which are less visible'. Please correct the formulation, define \\bar{V}, and provide a derivation or intuition for how the objective selects views with many low-visibility Gaussians.","section":"Section 5.2.2, Eq. (8)"},{"comment":"The central assumption that 'barely visible Gaussians are more likely to generate visual artifacts' is never tested against artifact location data. Without such validation, it is unclear whether the pose-suggestion optimization points users to views where artifacts actually occur. A simple test would be to correlate per-Gaussian visibility in the input video with artifact presence in the initial avatar, e.g., by comparing visibility of deleted/edited Gaussians in user trials with that of all Gaussians.","section":"Sections 5.2.1 and 5.2.2"},{"comment":"The crowdsourced preference percentages are reported without any significance testing, and the 'Same quality' responses are not incorporated. For example, in the Avatar 1 (P1 & P3) fencing row, SuperSplat was preferred by 71.4% of evaluators; the paper explains this in the text but does not test whether the overall 11/12 preference for Ours is statistically reliable. Binomial tests or confidence intervals for each pair, as well as an analysis aggregating the 12 questions, are needed.","section":"Section 7.4, Table 4"}],"minor_comments":[{"comment":"The keyword 'Gaussian splitting' should be 'Gaussian Splatting'.","section":"Section 1, Keywords"},{"comment":"The phrase 'four male' is awkward; consider 'four participants (all male, aged 21–26)'.","section":"Section 7.1.1"},{"comment":"The word 'Gaussian' is misspelled as 'Gaussain' in several places (e.g., Figure 1 caption); please correct these typos.","section":"Figure 1 and elsewhere"},{"comment":"The 'Same quality' row is never discussed in the text; a sentence interpreting these responses would help.","section":"Table 4"},{"comment":"The use of the pose decoder D and the latent code dimension (z ∈ R32) is only briefly mentioned; a few more details in the main text or a clear reference to the supplement would improve reproducibility.","section":"Section 5.2.2"},{"comment":"The caption claims 'less than 10 minutes per avatar', while Table 3 reports total times of about 635–658 seconds for the same five-iteration protocol; clarify which process the caption refers to.","section":"Figure 9 caption"}],"recommendation":"major_revision","confidential_remarks":"This is within the scope of an HCI systems venue, but the evaluation standards are below what is typically required to support the strong concluding claim. The lack of statistical testing and the confounded comparison are the main obstacles. I would encourage the editors to treat this as a major-revision candidate rather than a reject, as the system concept is novel and the crowdsourcing data, if properly analyzed, could provide some support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AvatarPerfect is a sensible, well-described system that combines 2D image editing with automatic body and camera pose suggestion for cleaning up artifacts in 3DGS avatars. The pose suggestion, driven by Gaussian visibility, is a reasonable adaptation of next-best-view estimation and is the genuinely new piece here. The paper is clearly written and unusually honest about its limitations.\n\nWhat the paper does well: the system design is thoughtful. Editing in 2D sidesteps the difficulty of direct 3D Gaussian manipulation, the three tools map onto the two artifact types, and the additional training with a masked loss is a clean way to incorporate user edits. The user study is well-intentioned, with counterbalancing and a practice session. The crowdsourced preference study, though small, shows a strong directional signal in favor of AvatarPerfect in 11 of 12 animation pairs.\n\nThe soft spots are real, and the stress-test note lands. The evidence for the headline claim is thin: four participants, all male, two avatars, per-cell n of two refined results. The quantitative gaps in Table 2 are tiny (IoU +0.002 and +0.008; PSNR +0.4 dB; SSIM +0.008) and no significance tests, confidence intervals, or effect sizes are reported. These differences could easily be noise. The baseline SuperSplat is restricted to deletion with a fixed body pose, so the comparison tests the whole pipeline rather than isolating the pose-suggestion contribution. And the core assumption—that artifacts correlate with low Gaussian visibility—is plausible but never directly validated against artifact location data.\n\nThese are addressable issues, not fatal ones. The system itself is credible, and the pose-suggestion formulation makes sense even if the empirical proof is not yet solid. I would not say the central claim is disproven; I would say it is not yet established.\n\nThis paper is for an HCI/graphics audience interested in interactive editing of neural representations. It contributes a working system and a novel pose-suggestion approach. It deserves serious peer review, but a referee should push for more participants, proper statistical reporting, and an ablation that separates pose suggestion from the 2D-editing paradigm. With those additions it could be a solid venue paper; without them, the quantitative claim should be softened.","headline":"A well-built interactive system for refining 3DGS avatars, but the headline quality claim rests on a weak user study; it deserves peer review with requests for stronger evidence.","tokens_in":18782,"tokens_out":1883,"would_cite":true,"duration_ms":16795,"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":"2D painting of suggested views beats direct 3D Gaussian deletion for cleaning up avatars, the paper claims.","keywords":["3D Gaussian Splatting","avatar refinement","user-assisted editing","pose suggestion","next best view","image inpainting","monocular video","human avatar"],"falsifier":"A decisive check: render the initial avatar under many poses, flag pixels where the output diverges from a ground-truth free-motion frame or where raters identify floating splats and color anomalies, then compare those locations with the accumulated per-Gaussian visibility from the training video; if artifact regions are not enriched in low-visibility Gaussians, the pose-suggestion objective is not the source of the reported gain and random pose selection should perform equally.","tokens_in":17813,"feed_emoji":"🖌️","tokens_out":7699,"duration_ms":63496,"temperature":0.7,"pith_summary":"This paper is trying to establish that the practical fix for artifacts in 3D Gaussian Splatting avatars is a mixed-initiative loop: the system decides where to look by suggesting a body and camera pose, the user simply paints away the bad splats and colors in the rendered 2D image, and the avatar is retrained on the corrected frame together with the original video. The payoff, if true, is that a non-expert can clean up an avatar produced from a monocular video without learning 3D manipulation, and the hard part—finding the hidden problem Gaussians—is automated. In a within-subjects study with four participants, the 2D-painting workflow beat a direct 3D-deletion editor on IoU, PSNR, and SSIM for both test avatars, and independent raters preferred the refined avatars in 11 of 12 animated pairs. The larger consequence is that monocular pipelines become usable: a short interactive pass can recover visual quality that fully automatic methods leave on the table under novel poses.","feed_headline":"2D painting beats direct Gaussian deletion for avatar cleanup","feed_subtitle":"Suggested poses expose hidden Gaussians, and edited 2D images retrain the avatar, beating direct 3D deletion.","key_machinery":"The central object is the per-Gaussian visibility weight $v^p_i=T^p_i\\alpha^p_i$ from the 3DGS $\\alpha$-blending equation, summed over pixels to give $v^I_i$ for a rendered image. Accumulating these visibility vectors over the input video and the edited frames yields the history $V$; the pose-suggestion step minimizes $\\sum_i \\mathrm{clip}(V[i]-\\bar V)\\,\\sigma(G,D(z),\\tau)[i]$ over body-pose latent $z$ and camera $\\tau$, which selects views where the least-seen Gaussians occupy the most pixels. Everything else—the painting tools, the masked training loss, the oversampling of edited frames—exists to make that chosen view informative and to transfer the 2D fix back into 3D without forgetting the original video.","core_discovery":"On the paper's own terms, the discovery is that 'barely visible' is a computable, optimizing target: each Gaussian's contribution $v^p_i=T^p_i\\alpha^p_i$ to every pixel accumulates into a visibility history, and the next view is chosen to maximize the screen-space presence of the least-seen Gaussians while a pose decoder keeps the body natural. The user then corrects that rendered frame with three 2D tools (background erase, solid-color inpaint, and diffusion inpaint), and the avatar is updated by additional training on the original video plus the masked edited frame, with edited frames oversampled so the fix sticks. The paper reports that after five such iterations the refined avatars beat the direct-deletion editor on IoU, PSNR, and SSIM for both avatars used, and that crowdsourced viewers chose them as higher quality in 11 of 12 motion/avatar pairs.","pith_inferences":["Going beyond the paper, the same visibility loop should generalize to any artifact type that leaves a visible trace in a rendered image, as long as a per-element visibility measure exists; the paper notes meshes, NeRFs, and SDFs would each need a new such measure.","A natural extension is to automate the painter: an artifact detector operating on suggested views could close most of the same gaps, leaving the human to judge only ambiguous edits.","Because the visibility history weights edited frames by only $w=0.01$ but training oversamples them at probability $0.3$, there is an unexplored explore/exploit knob: raising or lowering $w$ controls how fast the system stops revisiting improved regions.","The pose suggestion could double as a diagnostic: applying it to a freshly trained avatar predicts where novel-pose artifacts are most likely before any user edit begins."],"forward_implications":["Cleanup no longer requires selecting individual Gaussians in 3D; the user only paints on the suggested rendered image and presses 'update avatar'.","Because edited frames are added with a binary mask while original frames are still sampled, corrections are local and the avatar keeps its appearance elsewhere on the body.","The same visibility objective proposes nontrivial body poses, not just new camera angles, exposing occluded areas where the two classic artifacts form.","The reported quality gains imply direct 3D deletion is at a structural disadvantage: deleting whole Gaussians can expose previously hidden color errors, while masked retraining repairs them in place."],"supporting_citations":[{"why":"Baseline direct-3D-deletion editor that the central comparison is defined against.","marker":"[35]"},{"why":"Generates the initial 3DGS avatars from monocular video and supplies the motions used in the crowdsourced evaluation.","marker":"[21]"},{"why":"Provides the differential Gaussian rasterization and alpha-blending equations from which per-Gaussian visibility is defined.","marker":"[17]"},{"why":"Supplies the body-pose representation used for pose suggestion and skinning-based deformation.","marker":"[28]"},{"why":"Provides training and ground-truth videos used for quantitative evaluation of the refined avatars.","marker":"[27]"},{"why":"Provides the pose decoder that maps the optimized latent code to natural body poses, keeping suggestions plausible.","marker":"[34]"},{"why":"Segments actor and background masks used to compute IoU and SSIM/PSNR metrics.","marker":"[18]"},{"why":"Powers the diffusion-inpaint tool that fills complex-texture regions during user editing.","marker":"[36]"}],"fun_headline_variants":["2D paint outshines direct Gaussian editing for avatars","Smart pose suggestions boost avatar touch-ups","Edit avatars in 2D, beat 3D deletion","User-driven 2D fixes improve 3D avatars","Pose-guided 2D retouching perfects 3D avatars"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pose suggestion assumes a correlation between how rarely a Gaussian was visible in the input video (and prior edited frames) and how likely it is to produce a visible artifact in a new pose.","fun_headline_variants_meta":{"raw":{"variants":["2D paint outshines direct Gaussian editing for avatars","Smart pose suggestions boost avatar touch-ups","Edit avatars in 2D, beat 3D deletion","User-driven 2D fixes improve 3D avatars","Pose-guided 2D retouching perfects 3D avatars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000282,"raw_usage":{"total_tokens":1651,"prompt_tokens":909,"completion_tokens":742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":656}},"tokens_in":525,"tokens_out":742,"duration_ms":6177,"temperature":1.0,"reasoning_tokens":656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:15:18.002883+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check: render the initial avatar under many poses, flag pixels where the output diverges from a ground-truth free-motion frame or where raters identify floating splats and color anomalies, then compare those locations with the accumulated per-Gaussian visibility from the training video; if artifact regions are not enriched in low-visibility Gaussians, the pose-suggestion objective is not the source of the reported gain and random pose selection should perform equally.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline direct-3D-deletion editor that the central comparison is defined against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides training and ground-truth videos used for quantitative evaluation of the refined avatars."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Powers the diffusion-inpaint tool that fills complex-texture regions during user editing."}],"review_version":1}