{"id":"eba86ae3-74fa-4349-8ffd-3f533fd8a6f1","arxiv_id":"2607.27982","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"2D skeletal sketches and rigidity maps injected via gated adapters into frozen UniRig and Puppeteer backbones recover target rigs better than geometry-only baselines and support iterative editing.","lead":"ViP-Rig lets artists control 3D character skeletons and skin weights by drawing 2D bone sketches and color-coded rigidity maps on a mesh rendering. It freezes strong pretrained riggers and steers them with compact visual tokens, aiming to replace tedious 3D hierarchy editing.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Quantitative gains rest on GT-derived prompts that baselines never receive, so target-recovery tables do not isolate controllable rigging from privileged conditioning.","rationale":"The reader correctly isolates the evaluation proxy as the weakest link. The architecture (frozen backbones + zero-init gated / symmetric adapters + Dense-to-Compact encoding) is coherent, ablations (Tables 3–4) support the design choices, and zero-shot MR numbers are a positive signal. None of that removes the circularity: quantitative tables measure reconstruction of the annotation that generated the prompt, while baselines are denied that signal. Qualitative editing demos are encouraging but insufficient to carry the claim. A controlled free-form / edited-prompt quantitative protocol (or a small user study) is the natural condition for upgrading from CONDITIONAL; until then the verdict stays CONDITIONAL at moderate confidence. No deeper mathematical inconsistency is present.","tokens_in":11525,"tokens_out":572,"duration_ms":13141,"concrete_test":"Hold out a subset of Art-XL2.0 test meshes. Generate two prompt sets: (1) the paper’s GT-projection prompts and (2) independent free-hand or heavily perturbed sketches (or edited unconditional UniRig/Puppeteer projections) that intentionally deviate from GT topology/rigidity. Re-run ViP-Rig and report CD-J2J/B2B and skinning ℓ1 against both the prompt-implied target and the original GT. If the advantage over geometry-only baselines shrinks by >50% or localization fails under set (2), the load-bearing quantitative claim does not transfer to real user control.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim equates superior CD/precision/ℓ1 under “prompt-guided evaluation” (Tables 1–2) with successful controllable rigging. Yet Sec. 4.1 constructs every benchmark prompt by projecting the exact GT skeleton (or entropy of GT skinning weights) onto a VLM-selected view, then scores recovery of that same annotation. Baselines (RigNet, UniRig, Puppeteer, …) receive only geometry and therefore cannot exploit the same channel. Consequently the reported margins mix (a) genuine adapter control with (b) the simple fact that the model is shown a rendering of the answer. Free-form / edited prompts appear only in qualitative Fig. 4; no quantitative protocol measures fidelity, localization, or stability when the prompt is sparse, multi-view-ambiguous, or intentionally different from any GT. If the quantitative edge collapses once prompts are no longer GT projections, the strongest empirical support for the claim disappears.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"ViP-Rig proposes a two-stage framework for controllable character rigging conditioned on user 2D visual prompts: a skeletal sketch for joint layout/branching and a color-coded rigidity map for regional deformation. Compact tokens from a frozen DINOv2 encoder plus stage-specific Perceiver Resamplers are injected into frozen pretrained backbones—layer-wise zero-initialized gated adapters into UniRig’s autoregressive skeleton generator, and symmetric pre-/post-interaction gated adapters into Puppeteer’s point–joint skinning streams—while preserving geometric priors. Training synthesizes prompts by projecting GT skeletons and by discretizing normalized Shannon entropy of GT skinning rows. On Articulation-XL2.0 and zero-shot ModelsResource, under prompt-guided evaluation, the method reports lower skeleton Chamfer distances (Table 1) and better skinning precision/recall/ℓ1 (Table 2) than geometry-only baselines, with ablations (Tables 3–4) supporting full layer-wise fusion and dual-path injection. Qualitative Figure 4 illustrates prompt-first rigging and result-guided editing.","tokens_in":11921,"tokens_out":1409,"duration_ms":37918,"significance":"If the controllability claims hold under realistic user prompts, the work addresses a genuine production gap: automatic riggers produce a single plausible rig, whereas artists need task-specific skeletons and regional skinning without heavy 3D hierarchy editing or per-joint weight painting. Conditioning frozen SOTA backbones via gated adapters is a clean, reusable design, and supporting both prompt-first and iterative edit loops is practically valuable for AI-generated assets. The zero-shot ModelsResource transfer and the ablation evidence for persistent and symmetric injection are concrete technical strengths. Significance is currently tempered by how much of the quantitative gain is attributable to privileged GT-derived prompts rather than free-form control.","major_comments":[{"comment":"§4.1 Evaluation protocol together with §3.1–3.2 Training: benchmark I_skel is built by projecting the ground-truth joints/connections onto a VLM-selected view, and I_skin by discretizing normalized Shannon entropy of the GT skinning rows; Tables 1–2 then score recovery of those same annotations. Geometry-only baselines (RigNet, MagicArticulate, UniRig, Puppeteer) never receive this channel. The reported margins therefore mix genuine adapter control with privileged conditioning on a 2D rendering of the target. Free-form, sparse, or intentionally non-GT prompts appear only qualitatively in Figure 4. The central claim of controllable rigging needs a quantitative protocol that measures fidelity/localization when prompts are user-like (edited, incomplete, multi-view-ambiguous, or deliberately different from GT), and/or baselines given comparable structural side information. Without that, Tabl","section":"§4.1, Tables 1–2"},{"comment":"§3.2 Training, Eq. (12): I_skin is synthesized by thresholding normalized Shannon entropy of each point’s GT weight distribution into rigid/transitional/flexible. This is a convenient proxy, not an artist-authored rigidity intent, and it is derived from the same W that the model is trained to predict. The paper should justify the three-level discretization (thresholds, sensitivity) and show that control still holds when rigidity maps are painted independently of GT entropy (e.g., user edits in Figure 4 quantified, or held-out artist maps). Otherwise skinning gains in Table 2 may partly reflect reconstructing the entropy field that generated the prompt rather than general regional control.","section":"§3.2, Eq. (12), Table 2"},{"comment":"Tables 1–2 report point estimates only (no standard deviations, multiple seeds, or significance tests) on a 2K Art-XL2.0 split and 270 MR characters. Given modest absolute margins in several cells (e.g., skinning ℓ1 0.335→0.309 vs Puppeteer on Art-XL2.0; precision 87.6%→88.3%), variability matters for the claim of consistent superiority. Please add error bars or multi-run statistics, at least for the main comparisons.","section":"Tables 1–2"}],"minor_comments":[{"comment":"Figure 2 and Eqs. (3)–(4): gate parameterization (per-layer α_ℓ, β_ℓ; four independent skinning adapter copies) is clear, but the number of Perceiver tokens K and feature dimension d are never stated in the main text; please report them for reproducibility.","section":"§3.1–3.2, Figure 2"},{"comment":"§4.1 says detailed sampling/augmentation and prompt perturbation settings are in the supplementary material; for a methods-heavy systems paper, a short main-text summary of perturbation magnitudes (landmark jitter, missing branches, color bleed) would help readers judge robustness claims without the supplement.","section":"§4.1"},{"comment":"Related work cites Animator-Centric Skeleton Generation (Sun et al. 2026) for global bone-density control; a brief quantitative or qualitative comparison against that form of weak structural conditioning would clarify the incremental value of dense 2D sketches.","section":"§2"},{"comment":"Single-view prompts are inherent to the interface; failure modes under heavy self-occlusion or wrong chosen view are not analyzed. A short limitation paragraph or failure-case figure would set expectations.","section":"§5 Conclusion"},{"comment":"Minor prose issues: missing spaces in compounds (“Existingautomatic”, “prompt-firstrigging”, “resultingskinningweights”) appear in the abstract and introduction; clean for camera-ready.","section":"Abstract, §1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation circularity raised by the skeptic is real and load-bearing for the strongest claim, but the architecture and qualitative editing story are credible and likely fixable with additional experiments rather than a redesign. I would not reject on novelty grounds: 2D visual prompting for both skeleton topology and regional skinning on frozen rigging backbones is a reasonable systems contribution for cs.CV/graphics. Fit is appropriate for a vision/graphics venue that accepts controllable 3D generation work. No concerns about misconduct or citation manipulation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real move here is practical: freeze UniRig and Puppeteer, compress a single-view 2D skeletal sketch and a three-class rigidity map with DINOv2 + Perceiver Resampler, and inject with zero-init gated adapters (layer-wise for the AR skeleton decoder; symmetric pre/post JPCA for skinning). That is a clean control interface artists can actually draw, and the paper shows both prompt-first blueprints and edit-the-projection loops, including on image-to-3D meshes.\n\nWhat it does well: ablations are coherent (full layer-wise fusion beats prefix/CA-k; dual-path skinning beats explicit scale or single stream). Zero-shot ModelsResource holds the ranking. Skinning is fairly isolated with GT skeletons. Citations sit on the right recent stack (UniRig, Puppeteer, MagicArticulate, RigNet). No math theater; the equations are standard gated cross-attn + cosine softmax.\n\nSoft spot, in proportion: the quantitative claim is \"prompt-guided target recovery\" where every benchmark I_skel is a projection of the GT joints/bones and I_skin is entropy-binned from the GT weight rows, then scored against those same GTs, while baselines get geometry only. That is not circular nonsense—the model still has to lift 2D to 3D hierarchy and weights—but the margins mix privileged conditioning with adapter skill. Free-form / sparse / intentionally non-GT prompts live only in Fig. 4. No user study, no error bars, no code in-text. That is the main interpretive caveat, not a collapse of the method.\n\nWho it is for: people building controllable character pipelines and anyone extending frozen 3D generative backbones with 2D artist input. Worth a serious referee. I would engage, cite for the interface pattern, and push for a non-GT prompt protocol and release.","headline":"Solid systems paper: 2D skeletal/rigidity prompts into frozen UniRig/Puppeteer via gated adapters; tables win under GT-derived prompts, free-form control is qualitative only.","tokens_in":12594,"tokens_out":482,"would_cite":true,"duration_ms":8376,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"2D skeletal and rigidity sketches can steer frozen rigging models to the particular skeleton and skinning an artist wants, not just a plausible one.","keywords":["character rigging","skeleton generation","skinning weights","visual prompting","gated adapters","controllable 3D animation","autoregressive skeleton models"],"falsifier":"Give artists freehand skeletal and rigidity sketches on held-out meshes (not projections of the ground-truth annotation), measure whether the output joints, hierarchy, and animated deformations match the intended edits better than geometry-only baselines and remain stable under sparse or multi-view-ambiguous drawings.","tokens_in":12355,"feed_emoji":"🦴","tokens_out":928,"duration_ms":18160,"temperature":0.7,"pith_summary":"The same 3D character mesh can need very different bones and deformation behavior depending on the animation task, so artists routinely edit an automatic rig by hand. Most learning systems only produce one geometry-driven rig and give little direct control over joint layout or regional stiffness. ViP-Rig lets a user draw or edit a 2D bone sketch and a color-coded rigidity map on a mesh rendering, then injects compact features from those images into frozen pretrained skeleton and skinning networks through gated adapters. Under evaluation that reconstructs target rigs from such prompts, it recovers joints, branches, and skinning weights more accurately than geometry-only baselines on a large articulation dataset and in zero-shot transfer, while qualitative examples show localized prompt-first rigging and iterative edit loops.","feed_headline":"2D bone sketches steer frozen models to the rig you want","feed_subtitle":"Gated adapters turn skeletal and rigidity drawings into controllable joints and skinning on a given mesh.","key_machinery":"Dense-to-Compact Visual Prompt Encoding plus task-specific gated injection: a frozen vision encoder and Perceiver Resampler compress the 2D prompt into fixed-length tokens; layer-wise gated cross-attention adapters steer autoregressive skeleton decoding, and symmetric pre/post adapters condition both point and joint streams before final skinning matching.","core_discovery":"ViP-Rig establishes that compact visual tokens from user 2D skeletal sketches and rigidity maps, injected via zero-initialized gated adapters into frozen pretrained skeleton generators and dual-stream skinning matchers, can explicitly control joint placement, branching, and regional deformation while preserving geometric priors—and that this recovers target skeletons and skinning weights better than geometry-conditioned baselines under prompt-guided evaluation.","pith_inferences":["If 2D visual prompts are enough to disambiguate underdetermined rigs, similar compact sketch conditioning may transfer to other underdetermined 3D authoring steps such as part segmentation, blendshape design, or cloth pinning.","The gap between annotation-projected benchmark prompts and true freehand sketches is the main open measurement problem; a public freehand-prompt test set would stress-test whether adapter control is truly user-facing.","Symmetric conditioning of both sides of a matching function is a general pattern that may apply whenever a user wants to bias pairwise assignments (e.g., correspondence or binding) without rewriting the matcher."],"forward_implications":["Artists can specify different task-specific rigs for one mesh by drawing 2D bone layouts and rigidity regions instead of editing a full 3D hierarchy or painting dense weights.","Frozen state-of-the-art skeleton and skinning backbones can be reused for controllable rigging without full retraining, via small gated adapters.","Prompt-first rigging extends to unrigged and image-to-3D assets when a single aligned rendering is available.","Result-guided loops—edit the projected skeleton or derived rigidity map, regenerate, repeat—become a practical workflow for elaborating or simplifying rigs.","Regional rigidity cues can separate adjacent parts that geometry alone would skin similarly (e.g., rigid armor next to a flexible shoulder)."],"fun_headline_variants":["Draw 2D bones to steer frozen generators toward target joints","Visual skeletal tokens control branching in frozen rig models","Gated adapters map bone sketches to joints and skinning weights","Rigidity maps and bone doodles modulate frozen skinning backbones","Compact 2D prompts recover target skeletons better than geometry alone"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That success at recovering a rig from prompts built by projecting the same ground-truth skeleton and entropy-derived rigidity labels is a fair stand-in for real freehand artist control on ambiguous meshes.","fun_headline_variants_meta":{"raw":{"variants":["Draw 2D bones to steer frozen generators toward target joints","Visual skeletal tokens control branching in frozen rig models","Gated adapters map bone sketches to joints and skinning weights","Rigidity maps and bone doodles modulate frozen skinning backbones","Compact 2D prompts recover target skeletons better than geometry alone"]},"model":"grok-4.5","effort":"low","cost_usd":0.004194,"raw_usage":{"total_tokens":1296,"prompt_tokens":841,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":41944000,"prompt_tokens_details":{"text_tokens":841,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":388,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":841,"tokens_out":67,"duration_ms":6887,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T21:26:04.356877+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Give artists freehand skeletal and rigidity sketches on held-out meshes (not projections of the ground-truth annotation), measure whether the output joints, hierarchy, and animated deformations match the intended edits better than geometry-only baselines and remain stable under sparse or multi-view-ambiguous drawings.","supporting_citations":[],"review_version":1}