{"id":"222d93a1-0b70-46b6-9b92-5fffcd0451cc","arxiv_id":"2508.00398","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"OSF uses flow-depth edge detection to provide occlusion-robust edge guidance for a single-stage stylization network, improving quality and speed in drawing-based 3D animation.","lead":"This paper presents a method to keep hand-drawn animation styles intact when body parts overlap in 3D motion. It combines depth and optical-flow cues to guide a faster single-stage stylization network, improving quality under occlusions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FDED's occlusion-robustness is conditional: flow-based edge recovery presupposes an unoccluded previous-frame edge map and reliable RAFT flow on stylized frames, and the paper does not test persistent-occlusion sequences.","rationale":"The reader's weakest-assumption analysis points to the same dependency identified here: FDED's flow branch requires an occlusion-free previous edge map and reliable optical flow. The paper is honest about this in footnote 6 and provides a plausible way to detect occlusion-free frames, but it does not evaluate the failure mode where no such frame exists at the start of a motion or where occlusion persists throughout. The reported ablations, efficiency numbers, and qualitative comparisons are credible and support the method's value under the intended condition, but they do not establish unconditional occlusion-robustness. Because the concern is a missing evaluation of the method's stated precondition rather than an observed contradiction, the appropriate verdict is CONDITIONAL: accept if the persistent-occlusion stress test confirms the mechanism, or if the claim is explicitly restricted to sequences with an unoccluded bootstrap frame.","tokens_in":14648,"tokens_out":4858,"duration_ms":53492,"concrete_test":"Construct a stress subset of 20 Mixamo motions whose first rendered frame already has at least 30% visible-area occlusion and whose occlusion never drops below 10% for the first 30 frames. Run OSF on this subset with d_{i-1} initialized from the first, already-occluded frame, and compare contour CLIP/SSIM against the same metrics on sequences that begin unoccluded. If USNet+FDED's gain over USNet is substantially smaller on the persistent-occlusion subset, the occlusion-robustness claim must be scoped to motions with a pre-occlusion bootstrap frame.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires FDED to recover contours that depth-based edge detection misses in occluded frames. The flow branch (Sec. 3.2, Eq. 5) warps source points from the previous frame's edge map d_{i-1} by RAFT optical flow. Footnote 6 concedes that the method assumes d_{i-1} is well-constructed without occlusion. This is load-bearing, not a technicality: if a motion starts in an occluded pose, or maintains occlusion across all frames (e.g., crossed arms held through a clip), then d_{i-1} already omits the contour the flow branch is supposed to restore, and f_i inherits the same gap. The paper reports an 'occlusion set' of 20 motions but never states how many of those have an occlusion-free starting frame, and Figure 13's robustness curve begins at 0% occlusion, so the bootstrap regime is not tested. A second unresolved dependency is flow quality: RAFT is trained on photorealistic video, while the inputs here are flat, stylized, texture-sparse drawings. If optical flow is inaccurate at contour points, the estimated p*_i are wrong and the filtering/interpolation in Eq. 5 cannot fix them. The paper reports no flow endpoint error or failure analysis on stylized projections. These are scoping limitations, not internal contradictions, but they mean the headline 'occlusion-robust' is only established for motions starting from an unoccluded reference with adequately trackable appearance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Occlusion-robust Stylization Framework (OSF) for drawing-based 3D animation. OSF introduces Flow-Depth Edge Detection (FDED), which combines a depth-based edge detector with a flow-based branch that warps edge points from the previous frame using RAFT optical flow to recover contours in occluded regions. The stylization network is re-designed as a single-stage unified network (USNet) with an edge-guided contrastive ranking loss, replacing the two-stage texture-then-contour pipeline of DrawingSpinUp. Experiments on 120 Amateur Drawings characters with 20 non-occluded and 20 occluded motions report consistent improvements in CLIP, SSIM, FID, and LPIPS over baselines, and show a 2.4x speedup and roughly 2.3x memory reduction relative to the two-stage baseline. The main weaknesses are that the flow-based edge recovery assumes a well-constructed, occlusion-free previous-frame edge map, and that the paper does not validate optical flow quality on stylized inputs or report variance for its quantitative claims.","tokens_in":15011,"tokens_out":5910,"duration_ms":51261,"significance":"If the occlusion robustness claim were fully established, the paper would be a useful contribution to the small but growing area of drawing-based 3D animation: it identifies a concrete failure mode (occluded contours), proposes a simple and interpretable fix (propagate previous-frame edges via optical flow), and demonstrates that the fix can be plugged into existing stylization networks. The unified single-stage network with contrastive learning is a clean contribution, and the computational analysis is a credible practical advantage. The paper also ships a reproducible pipeline built on public components (Wonder3D, RAFT, Mixamo), which is a strength. However, the central claim of 'occlusion-robust' is currently scoped more narrowly than the title and abstract suggest: it is supported only for motions that begin from an occlusion-free reference frame and for which RAFT provides accurate correspondences on flat stylized images. The quantitative evaluation also lacks error bars and a specified human-evaluation protocol, which weakens the support for the perceptual claims. With additional experiments that directly probe these dependencies, the contribution would be solid.","major_comments":[{"comment":"The flow-based edge detection assumes that the previous frame's edge map d_{i-1} is well-constructed and occlusion-free. This is load-bearing: if a motion starts in an occluded pose, or maintains occlusion across frames (e.g., crossed arms held through the entire clip), then d_{i-1} already omits the contour the flow branch is supposed to restore, and f_i inherits the same gap. The paper reports an 'occlusion set' of 20 motions but never states how many of those motions have an occlusion-free starting frame, and Figure 13's robustness curve begins at 0% occlusion, so the bootstrap regime is not tested. The authors should either (a) enforce the assumption by detecting and using an occlusion-free reference frame, as suggested in footnote 6, or (b) explicitly scope the claim to motions that begin from an unoccluded reference and provide experiments on persistent-occlusion sequences to characterize the failure mode.","section":"§3.2, Eq. (5), and footnote 6"},{"comment":"The optical flow estimator RAFT is trained on photorealistic video, while the inputs here are flat, stylized, texture-sparse drawings. Since the flow branch relies on accurate point correspondences at contour points, the paper should validate flow quality on stylized projections. Concretely, because the 3D animation pipeline provides known correspondences from the rigged mesh, the authors can measure flow endpoint error on the stylized 2D projections, or compare against an oracle flow in an ablation. Without such a check, the possibility that flow errors at contour points produce incorrect p*_i that the filtering/interpolation in Eq. (5) cannot repair remains an open correctness risk for the central claim.","section":"§3.2 and §4.1"},{"comment":"All metrics are reported as averages over 10 runs with different seeds, but no standard deviations or confidence intervals are given. Given the often small margins between the best and second-best methods (e.g., Table 1: texture CLIP 0.974 vs 0.971; contour CLIP 0.982 vs 0.978), the reader cannot assess whether the improvements are statistically meaningful. Please report variance and, where appropriate, significance tests or per-seed ranges.","section":"§4.2 and Tables 1-2"},{"comment":"The human evaluation is described only as 'Human evaluation of preferences is performed,' with no protocol details: number of participants, number of comparisons per participant, whether the evaluation was pairwise or single-stimulus, whether the participants were blind to method, and what instructions they received. The Human column in Table 1 is reported as a single score per method, so the perceptual claim is not verifiable. Please provide a full protocol description and appropriate statistical analysis.","section":"§4.2 and Table 1 (Human column)"}],"minor_comments":[{"comment":"The memory reduction is reported as 2.1x, but Table 3 gives 11.62 GB / 5.05 GB approximately 2.30x; please reconcile the numbers and use a consistent figure throughout.","section":"Abstract, Introduction, and Table 3"},{"comment":"The notation f_i = h·g(p*_i, d_i) is terse; the roles of the filter g and the interpolator h (dilation) could be described explicitly for readers who do not rely solely on Figure 5.","section":"§3.2, Eq. (5)"},{"comment":"The figure labels the y-axes as 'CLIP' and '1-LPIPS' while the text refers to 'consistency' and 'fidelity'; please make the terminology consistent with the metric definitions in §4.2.","section":"§4.4, Figure 13"},{"comment":"The statement that the stylization network is fine-tuned on a single drawing should clarify that this is the same assumption used in DrawingSpinUp [68], to avoid framing it as a newly introduced limitation.","section":"§2 and §4.1"},{"comment":"The contour extraction method is referenced only as 'the method in [68]'; please briefly describe or cite the contour estimator, since contour quality is a central evaluation component.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a graphics/vision venue. The main risk is overclaiming 'occlusion-robust' given the unvalidated bootstrap assumption and flow-quality dependency. I would recommend the editor seek a reviewer with optical flow experience, as the method's central mechanism depends on RAFT on stylized inputs. The heavy self-citation pattern in the references (e.g., [33,36,58,59,61,63]) is not problematic per se but could be streamlined for readability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a straightforward, well-executed improvement to drawing-based 3D animation. The new thing is FDED: it combines depth-based edge detection with a flow-based branch that warps edges from the previous frame to fill in contours lost to occlusion, and it feeds those edges into a single-stage stylization network trained with an edge-guided contrastive loss. That combination is new for this task, and the ablations actually support it—adding the flow branch to either their USNet or the previous DSU baseline gives consistent gains on both texture and contour metrics, and the qualitative examples show the difference where occlusions occur. The speedup from two-stage to single-stage is real (2.4x faster, 2.1x less memory) and not a gimmick.\n\nThe soft spots are real but not disqualifying. First, the flow branch assumes the previous frame's edge map d_{i-1} is clean and occlusion-free; footnote 6 says this, but the paper never tests what happens when the motion starts in an occluded pose or keeps the overlap for a long stretch. The robustness curve in Figure 13 starts at 0% occlusion, so the bootstrap regime is untested. Second, RAFT is used off-the-shelf on flat, texture-sparse stylized frames; there is no flow quality analysis on those inputs. If the flow is inaccurate at contour points, the whole correction fails. These are scoping limitations, not contradictions, but they mean \"occlusion-robust\" is established only for sequences that begin from a clean reference and have trackable appearance. Third, the paper reports 10-run averages without error bars, and the human evaluation is a single preference number with no protocol details. That's an easy fix for the authors.\n\nI think the reader's report is fair. The paper is honest about its assumption, and it doesn't oversell past 40% occlusion. For a paper in this niche, it deserves a serious referee. With the missing statistics and a test of the bootstrap condition addressed, it would be a clean accept.\n\nRecommendation: send to peer review, and ask for error bars, human evaluation details, and an experiment or discussion of the persistent-occlusion case.","headline":"Solid niche contribution with a real advance in occlusion handling, but the 'occlusion-robust' claim is conditional on clean starting frames and reliable flow on stylized images.","tokens_in":15485,"tokens_out":2721,"would_cite":true,"duration_ms":25725,"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":"Occlusion-robust edge guidance stabilizes contours in drawing-based 3D animation.","keywords":["drawing-based 3D animation","occlusion robustness","optical flow","edge detection","stylization","image-to-3D diffusion","contrastive learning","single-stage stylization"],"falsifier":"Render a synthetic motion where a limb passes completely behind a torso for several frames, with a known ground-truth contour; measure whether FDED's recovered edge matches the ground truth at the occluded boundary, and whether the output contours stop flickering across those frames.","tokens_in":14492,"feed_emoji":"🎨","tokens_out":5972,"duration_ms":54241,"temperature":0.7,"pith_summary":"The paper claims that contour flicker and stroke blur in drawing-based 3D animation come from a pose gap: the stylization network is trained on one clean, occlusion-free drawing, but at inference it must stylize poses where body parts overlap. To fix this, the authors propose OSF, which feeds the stylization network an occlusion-robust edge map. The edge map is built by FDED: depth edges cover visible contours, and optical flow recovers edges hidden by occlusion from the previous frame. A single-stage U-Net with edge-guided contrastive learning then stylizes the 2D projection, running 2.4x faster and using 2.1x less memory than the two-stage baseline. If correct, the framework would keep hand-drawn contours and stroke patterns stable in dynamic, occluded motions.","feed_headline":"Flow-based edges stop contour flicker in 3D drawing animation","feed_subtitle":"The new pipeline recovers occluded contours with optical flow and runs 2.4x faster than the two-stage baseline.","key_machinery":"The central object is FDED, a two-part edge detector that fuses a depth-based edge map $d_i$ with a flow-based edge map $f_i$. Depth-based detection applies Gaussian adaptive thresholding to the 3D mesh's depth map and produces the unoccluded edge $e_u$. Flow-based detection takes points lying on the previous clean edge $d_{i-1}$, shifts them by the RAFT flow vector $v_{i-1}$ to estimated positions $p^*_i = p_{i-1} + v_{i-1}$, keeps only points inside the current depth edge $d_i$, and dilates the survivors to form the occluded edge $e_o$. The union $e = e_o \\cup e_u$ is concatenated with the 2D projection $Z_i$ as guidance for a single U-Net trained with reconstruction loss plus an edge-guided contrastive ranking loss.","core_discovery":"The paper claims that the quality drop in drawing-based 3D animation under occlusion is caused by a stylization pose gap: stylization networks are fine-tuned on one occlusion-free drawing pose but must stylize arbitrary target poses at inference, many with overlapping body parts. To close the gap, OSF augments the stylization network's input with an occlusion-robust edge map produced by FDED. FDED keeps the depth-based edge for unoccluded regions and recovers the missing occluded edge by warping the previous frame's edge points along RAFT optical flow, filtering them to the current depth edge, and interpolating with dilation. With this guidance and an edge-guided contrastive loss, the single-stage USNet preserves contours and strokes under occlusion, and the framework runs 2.4x faster with 2.1x less memory than the two-stage baseline.","pith_inferences":["Because FDED only relies on consecutive rendered frames and a depth map, it could be lifted into other image-animation systems that render depth, not only drawing stylization.","A natural cold-start test is to begin a motion at an occluded pose without any clean previous edge map; the current pipeline assumes such a frame exists.","The edge-guided contrastive loss may transfer to other tasks where ground-truth labels are sparse along one channel, such as line-art colorization or sketch cleanup.","The reported 40% occlusion ceiling points to a concrete follow-up: combine FDED with patch-wise encoding to maintain quality in heavily occluded poses."],"forward_implications":["Animations of overlapping body parts keep stable contours because the flow-recovered edge supplies contour information that depth alone misses.","FDED is a drop-in condition: adding it to the two-stage DrawingSpinUp baseline also improves its occlusion-set consistency and fidelity.","The single-stage unified stylization network removes the intermediate ground-truth label and the extra network pass, cutting inference time by 2.4x and memory by 2.1x.","Edge-guided contrastive learning makes the stylization network converge faster and to a lower reconstruction loss, with sharper textures and contours.","Occlusion robustness is bounded: when the occlusion rate rises above about 40%, output quality still degrades, which the paper attributes to multiple occluded areas obscuring contours and strokes."],"supporting_citations":[{"why":"Supplies the optical flow estimator RAFT that tracks edge points between consecutive frames.","marker":"[52]"},{"why":"Generates the 3D structure from a single drawing that is rigged and projected to produce the 2D frames being stylized.","marker":"[30]"},{"why":"Provides the two-stage stylization baseline that OSF improves upon and compares against.","marker":"[68]"},{"why":"Provides the Amateur Drawings test characters and the single-pose training setup used in evaluation.","marker":"[46]"},{"why":"Supplies the U-Net encoder-decoder architecture used as the unified stylization network backbone.","marker":"[43]"},{"why":"Defines the Gaussian adaptive thresholding used by the depth-based edge detector.","marker":"[50]"},{"why":"Serves as the alternative 3D edge detector compared in the ablation study.","marker":"[27]"},{"why":"Serves as the non-stylization image-to-3D baseline in experimental comparisons.","marker":"[51]"}],"fun_headline_variants":["Optical flow closes the stylization pose gap in 3D animation","Single-stage stylization with flow-carried edges runs 2.4x faster","Warped edges keep drawing strokes intact under occlusion","Flow-guided edges fix flicker and blur in occluded animation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole recovery mechanism depends on the previous frame's edge map being clean and occlusion-free, and on the optical flow correctly tracking those edge points into the current frame.","fun_headline_variants_meta":{"raw":{"variants":["Optical flow closes the stylization pose gap in 3D animation","Single-stage stylization with flow-carried edges runs 2.4x faster","Warped edges keep drawing strokes intact under occlusion","Flow-guided edges fix flicker and blur in occluded animation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1562,"prompt_tokens":1000,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":616,"tokens_out":562,"duration_ms":5669,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:09:41.009493+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic motion where a limb passes completely behind a torso for several frames, with a known ground-truth contour; measure whether FDED's recovered edge matches the ground truth at the occluded boundary, and whether the output contours stop flickering across those frames.","supporting_citations":[{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the optical flow estimator RAFT that tracks edge points between consecutive frames."},{"cited_title":"Wonder3d: Sin- gle image to 3d using cross-domain diffusion","cited_arxiv_id":null,"evidence_quote":"Generates the 3D structure from a single drawing that is rigged and projected to produce the 2D frames being stylized."},{"cited_title":"DrawingSpinUp: 3D Animation from Single Character Drawings","cited_arxiv_id":"2409.08615","evidence_quote":"Provides the two-stage stylization baseline that OSF improves upon and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Amateur Drawings test characters and the single-pose training setup used in evaluation."},{"cited_title":"Computer vision: algorithms and applica- tions","cited_arxiv_id":null,"evidence_quote":"Defines the Gaussian adaptive thresholding used by the depth-based edge detector."},{"cited_title":"Neural contours: Learning to draw lines from 3d shapes","cited_arxiv_id":null,"evidence_quote":"Serves as the alternative 3D edge detector compared in the ablation study."}],"review_version":1}