{"id":"3e100d1f-f228-4fd3-b0c8-ab11d22e30ba","arxiv_id":"2506.01288","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"WishGI bakes static global illumination into a per-object spherical-harmonics probemap, reconstructs lighting per vertex with two probe associations, and achieves memory usage about 5% of mainstream lightmaps at competitive visual quality.","lead":"WishGI is a baking pipeline that stores static scene lighting as spherical-harmonic coefficients in one compact probemap and reconstructs the light per vertex instead of per pixel, cutting fragment-shader work. On mobile GPUs it reports about 5 percent of the memory of industry lightmapping and volumetric lightmaps, with comparable or better measured error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing risk is that the per-mesh association A, optimized once in an unspecified 'standard scene' (Sec. 3.4), may not generalize to all real scenes; the paper reports no sensitivity analysis for this central pillar.","rationale":"The reader's weakest_assumption matches our own: the fixed per-mesh association A is the pivot on which the method's universality rests. We considered other candidates and set them aside. The memory claim is supported by Table 2 (Ours 0.63–0.94 KB vs Lightmap 10.68–42.72 KB and VLM 9–16 KB) and by the scene-level day-night example (0.5 MB/probemap vs 18 MB lightmaps), so it is not the weakest link. The runtime claim is supported by Tables 3–4. The use of an SH-based ground truth is a deliberate choice—the paper targets low-frequency indirect light and explicitly acknowledges the inability to represent high-frequency details, so this is a fair evaluation of the stated goal. The remaining weakness is the generalization of A: the association is optimized in a synthetic, undisclosed standard scene; if the optimum varies with lighting or occlusion structure, the method fails outside the training distribution. The paper provides no sensitivity analysis, and the three demo scenes may correlate with the standard scene. Our proposed test directly measures this sensitivity. The verdict should remain CONDITIONAL because the method is plausible and well-engineered, but the central universality claim is empirically unverified.","tokens_in":14339,"tokens_out":11994,"duration_ms":129436,"concrete_test":"Take the three Unreal demo scenes (Sun Temple, Summer HandPainted, Stylized Egypt). For each of the test meshes used in Table 2, re-run Section 3.4's association optimization using two or three different standard scenes (e.g., an empty room with a single directional light; a room with multiple area lights and boxes; an outdoor scene with an HDR sky) and with, say, 6 vs 24 mesh rotations, keeping all other parameters (100 pts/m², 120 directions, lambda=0.1, 400 iterations) fixed. Then bake the probe SH values and evaluate mRMSE on the same ground truth. If the average mRMSE across meshes changes by more than ~10% relative, or any single mesh by more than ~20%, when the standard scene is swapped, the association is sensitive to the training setup and the generalization claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 optimizes the vertex–probe association A by minimizing Eq. (10) over scenarios constructed by placing the mesh in \"a standard scene with complex lighting\" and rotating it. This A is then fixed for every instance and every scene, while Section 3.3 solves for the probe SH values via Eq. (9) given A. The quality of the final reconstruction therefore depends entirely on A, yet the paper gives no details of the standard scene, no rotation set, and no value for N_sc, and reports no sensitivity of mRMSE or visuals to these choices. If the optimal A depends on the actual occlusion layout or light distribution of a production scene, the fixed association can yield poor subspaces for the 2-probe-per-vertex reconstruction, producing light leakage or blur in exactly the scenes the method is intended for. The paper tests only three Unreal demo scenes, all of which may resemble the standard scene's characteristics. Notably, Section 5.4 concedes that for meshes that are \"highly complex\" or \"do not require reuse,\" targeted optimization based on the actual lighting environment is preferable, implicitly acknowledging that the fixed A may not be universally optimal. This is the most load-bearing assumption because the central claim of \"consistent lighting quality across all instances of the same mesh\" hinges on A being a scene-independent, geometry-only quantity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents WishGI, a static global illumination baking pipeline for low-end platforms that reconstructs per-vertex hemispherical illumination as a linear combination of a small number of scene-level spherical-harmonics probes. The probe coefficients are obtained by solving a regularized least-squares fit to ray-traced radiance over the effective hemisphere, and the vertex–probe association is optimized offline in the mesh's local space by rotating the mesh in a synthetic standard scene. The method claims roughly 5% of the memory of mainstream lightmap/volumetric lightmap approaches, lower fragment-shader texture sampling, and competitive visual quality, with experiments on three Unreal scenes and several single meshes.","tokens_in":14566,"tokens_out":5529,"duration_ms":48469,"significance":"If the claims hold, the work is a practical contribution: the local-space inverse probe association is a novel idea that could reduce probe redundancy, and the reported fragment-sampling reduction (3.08 vs. 12.2 samples in Table 3) is a plausible path to low-end support. The optimization is differentiable and the core formulas are standard least squares, which lends credibility to the method. However, the current evidence leaves the generalization of the fixed association insufficiently quantified, which is central to the claim of consistent quality.","major_comments":[{"comment":"The vertex–probe association A is optimized in a single 'standard scene' with unspecified complexity, lighting, and rotation set, and no sensitivity analysis is reported. Because A is then fixed for all scenes and instances, the paper's central claim of consistent lighting quality depends on this generalization. Specify the standard scene, the number of scenarios N_sc, and the rotation set, and report mRMSE or visual results across a range of lighting environments (e.g., directional, point, area lights, varying occlusion) and scene structures.","section":"Section 3.4 / Eq. (10)"},{"comment":"The evaluation metric mRMSE is essentially the same cosine-weighted hemispherical difference that the baking loss minimizes (Eq. 4). The method is therefore being evaluated on its own objective, which biases the quantitative comparison in Table 2. Please add an independent metric, such as per-pixel PSNR/SSIM of rendered images against a path-traced reference, or demonstrate that the weighting choice does not change the ranking of methods.","section":"Section 5.1 / Eq. (13) vs. Section 3.3 / Eq. (4)"},{"comment":"The numerical ground truth is the third-order SH projection of ray-traced radiance, and the method itself reconstructs SH2/SH3 coefficients. This shared representation can favor the method. Please justify the choice of SH order for the ground truth, and either report the error against the original ray-traced radiance or use a representation-agnostic metric to confirm the quantitative results.","section":"Section 5.1 / Ground truth"},{"comment":"The headline 'approximately 5% of the memory' is not substantiated. Table 2 reports per-object probe memory only; the scene-level Time-of-Day example in Section 5.4 shows 4MB for eight probemaps versus 18MB for a single lightmap, which is about 22% of the memory for one time point. Table 3 reports bandwidth in MB/frame, not storage. Provide a scene-level storage comparison that supports the 5% figure.","section":"Abstract / Section 5.2 / Section 5.4"},{"comment":"The derivation of the normal equation (9) is deferred to a supplementary that is not available to the reader. Since this is the core of the baking pipeline and the notation (e.g., the products involving w, T(Y), B, W) is non-trivial, please include the full derivation in the main text or in an accessible supplementary document.","section":"Eq. (9)"},{"comment":"The limitation paragraph concedes that for meshes that are 'highly complex' or 'do not require reuse', targeted optimization based on the actual lighting environment is preferable. This directly qualifies the claim that the fixed association A works consistently for all instances. Please characterize these failure cases quantitatively (e.g., which of the tested meshes required targeted optimization) and incorporate this into the central claim.","section":"Section 5.4"}],"minor_comments":[{"comment":"K-Medoids is attributed to Lloyd [1982], which is the reference for k-means; please use a k-medoids reference or clarify the clustering variant.","section":"Section 3.4"},{"comment":"The phrase 'we use twelve 10-bits' should be 'twelve 10-bit values' and similarly for 'fifteen 8-bits'.","section":"Section 3.2"},{"comment":"The column header 'Mesh | Mem.' appears to combine the mesh name and memory; reformat for readability.","section":"Table 2"},{"comment":"The terms 'pixel' for the two 16-byte blocks in Figure 4 are confusing; use 'texel' or 'data block'.","section":"Section 3.2 / Figure 4"},{"comment":"There is an unresolved placeholder '[Knodt et al. 2023; ?]'; fill in the missing reference.","section":"Section 2"},{"comment":"The sentence 'Each vertex needs to be assigned the top n probes with the highest weights as geometric priors to avoid visual discontinuities in Figure 7(a) when the mesh triangles are dense' is unclear; specify what 'top n' is and how it relates to the two-probe-per-vertex setup.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong systems contribution, but the missing sensitivity analysis for the association generalization and the overlap between loss and metric are significant enough that I cannot recommend acceptance in the current form. I agree with the reader's conditional verdict; the authors should be asked to provide the missing experiments and derivations. The work aligns well with ACM TOG if these concerns are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a competent applied graphics paper from the Tencent/USTC group. The core idea—baking a per-mesh vertex–probe association in local space once, then reusing it for every instance—is genuinely new relative to the scene-level probe distribution work they cite (Cupisz, Iwanicki & Sloan, Volumetric Lightmaps). The engineering is careful: a single probemap, vertex-based reconstruction in the base pass, LOD-aware sampling, and mobile FPS measurements that look credible. The SH fitting with cosine weighting toward effective directions is a sensible reformulation, and the least-squares closure in Eqn. (9) is standard. The ablations (geometric prior, gradient regularization) are short but informative. They are also honest about limitations: no high-frequency shadows, potential LOD popping, simple meshes, dynamic objects. That honesty counts in their favor.\n\nThe soft spots, in order of importance. First, the fixed association A is the load-bearing piece, but the training setup is under-specified: the paper says they \"construct a standard scene with complex lighting\" and rotate the mesh, but gives no description of that scene, no N_sc, no rotation set, and no sensitivity analysis. If the optimal A depends on the lighting or occlusion distribution of the target scenes, the generalization claim is unquantified. They test three Unreal demo scenes, all plausibly similar to their standard scene. Their own Section 5.4 concedes that for \"highly complex\" meshes, targeted per-scene optimization is preferable—so the universal-association claim already has known exceptions. Second, the quantitative metric partially aligns with the training loss: both Eqn. (4) and Eqn. (13) weight directions by cos(n, d). That biases the mRMSE comparison toward the method. The visual comparisons mitigate this, but the metric alone is not neutral. Third, the \"approximately 5% of memory\" claim is baseline-dependent—against a 64px lightmap the saving is about 23%, not 5%. They should pin down the reference. Fourth, Eqn. (9)'s derivation lives only in a supplementary I cannot inspect; I didn't spot an error in the main text, but that step is unverified.\n\nNone of these are load-bearing flaws. The method is plausible, the paper is honest, and the engineering value is real. I'd send this to a referee who knows probe-based GI. It deserves serious engagement, especially on the probe-association idea. My recommendation: engage with it, and push for a sensitivity analysis of A and a precise memory comparison.","headline":"A well-engineered mobile GI baking paper with a genuinely new per-mesh probe-association idea, but the generalization of that fixed association is under-analyzed and the metric slightly favors the method.","tokens_in":15180,"tokens_out":2987,"would_cite":true,"duration_ms":31566,"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":"Per-vertex spherical-harmonic reconstruction with a mesh-local probe association delivers static global illumination at roughly 5 percent of the memory of industry lightmap methods, while cutting fragment-shader sampling.","keywords":["static global illumination","spherical harmonics","inverse probe distribution","light baking","low-end platforms","lightmap","fragment shader sampling","level of detail"],"falsifier":"Bake one mesh with the fixed association under adversarial real scenes—strong directional sources, large occluders, and light distributions absent from the rotation-based training—then compare its mRMSE and visible light leakage against the same mesh with probes optimized in that actual scene; a large quality gap would falsify the generalization claim.","tokens_in":14017,"feed_emoji":"💡","tokens_out":8431,"duration_ms":83804,"temperature":0.7,"pith_summary":"This paper proposes a static global illumination baking pipeline aimed at low-end real-time platforms. It claims that reconstructing light per vertex from a small set of spherical-harmonic probes, with each mesh assigned an offline-optimized probe association in its own local space, produces lighting quality competitive with mainstream industry lightmap methods while using only about 5 percent of their memory. The method also reduces fragment-shader texture sampling and removes the need for UV mapping. If the claim holds, mobile and other constrained devices could run convincing baked global illumination in forward rendering without an extra light pass.","feed_headline":"Global illumination baked in 5% of the memory, same light","feed_subtitle":"Per-mesh spherical-harmonic probes cut fragment-shader sampling, skip UV maps, and keep quality across instances.","key_machinery":"The load-bearing object is the per-mesh association $A$: for each vertex, a short list of probe indices and normalized weights that defines the vertex's spherical harmonics as a linear combination of probe spherical harmonics. It is initialized by K-medoids clustering on the mesh surface with a visibility-aware distance (Euclidean for mutually visible points, pathfinding distance for occluded ones), then refined by minimizing the lighting loss over rotated views of the mesh in a standard scene. This association removes scene-space probe placement and lets all instances of one mesh share a single lightweight embedding; the spherical-harmonic coefficients themselves are then solved from the normal equations of a linear least-squares problem, Eq. (9), with the regularization controlled by $\\lambda$.","core_discovery":"On its own terms, the paper's central discovery is that pixel-level probe interpolation is unnecessary for static objects: a vertex-based illumination model, where each vertex's spherical-harmonic vector is a weighted combination of probes and fragments interpolate those vectors barycentrically, can carry the full reconstruction. The probes are fit by minimizing a hemisphere-weighted lighting error $\\mathbf{E}_{\\mathrm{light}}$ plus a gradient-regularization term, which reduces baking to a linear least-squares problem in the probe coefficients. The other half is the inverse probe distribution: K-medoids clustering with a visibility-aware distance initializes the vertex–probe association $A$, and gradient descent in a synthetic standard scene refines it, so the association embeds into the mesh and every instance reuses it. The paper reports lower multi-directional RMSE than Directional Lightmap and Volumetric Lightmap on most test meshes, with fragment texture samples per texel dropping to 3.08 from 12.2 (VLM) and 5.14 (Lightmap).","pith_inferences":["The paper does not quantify how the fixed association $A$ degrades when a scene's lighting layout differs strongly from the synthetic standard scene; a sweep over lighting configurations, with per-scene optimized probes as the reference, would reveal the true generalization margin.","A natural testable extension is dynamic objects: interpolating spherical harmonics from nearby probes, which the paper sketches as viable, could be validated against path-traced ground truth for moving characters.","The paper itself notes in Section 5.4 that spherical harmonics and the vertex-probe structure cannot represent sharp shadows and that LOD transitions may pop; these are acknowledged boundaries that bound the 'competitive lighting' claim to low-frequency, static lighting.","The probe-count ceiling of 256 raises a packing question the paper leaves open: how many distinct meshes can share one scene probemap before texture resolution or index precision becomes the limiting factor."],"forward_implications":["Static scene lighting can be baked and rendered at roughly 5 percent of the memory of Directional Lightmap or Volumetric Lightmap configurations at comparable or better measured mRMSE.","Fragment-shader texture sampling drops to about 3.08 samples per texel, versus 12.2 for Volumetric Lightmaps and 5.14 for Directional Lightmaps, which directly relieves the bottleneck on tile-based deferred GPUs.","Artists can skip UV mapping for static geometry, since illumination is stored at vertices and in a single scene-wide probemap, eliminating UV-gap waste and draw-call overhead.","Level-of-detail support is built in by switching between second-order and third-order spherical harmonics, taking one or two texture samples depending on distance.","Time-of-day updates reduce to interpolating among a small set of probemaps, with a day-night cycle demonstrated at 4 MB total for eight time points."],"supporting_citations":[{"why":"Defines Directional Lightmap, the industry baseline whose memory footprint, texture sampling, and mRMSE the paper compares against.","marker":"[Sloan and Silvennoinen 2018]"},{"why":"Defines Volumetric Lightmaps, the second main baseline, with nine 3D textures and per-pixel trilinear interpolation.","marker":"[Games 2024c]"},{"why":"Introduces spherical-harmonics precomputed radiance transfer and the SH representation that the baking and reconstruction pipeline builds on.","marker":"[Sloan et al. 2002]"},{"why":"Supplies the edge-based least-squares vertex baking gradient used in the regularization term of the baking loss.","marker":"[Kavan et al. 2011]"},{"why":"Provides the inverse probe-distribution inspiration, combining light-field information with learned probe placement to reduce light leakage.","marker":"[Michal Iwanicki and Shirley 2024]"},{"why":"K-medoids clustering is used to initialize probe positions from geometric priors during mesh import.","marker":"[Lloyd 1982]"},{"why":"A* pathfinding, combined with Dijkstra search, defines the visibility-aware distance between occluded sample points.","marker":"[Hart et al. 1968]"}],"fun_headline_variants":["Global illumination baking with 5% memory footprint","SH baking cuts GI memory to 5% without quality loss","Per-mesh SH probes deliver GI at 5% memory cost","Lightweight GI baking: 5% memory, no extra passes","Static GI baked with SH, using 5% of usual memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one per-mesh vertex–probe association, tuned once in a synthetic standard scene and reused by every instance, keeps its quality in arbitrary real scenes; if the optimal association depends on the actual arrangement of lights and occluders, the fixed weights will leak light or blur shading in layouts unlike the training set.","fun_headline_variants_meta":{"raw":{"variants":["Global illumination baking with 5% memory footprint","SH baking cuts GI memory to 5% without quality loss","Per-mesh SH probes deliver GI at 5% memory cost","Lightweight GI baking: 5% memory, no extra passes","Static GI baked with SH, using 5% of usual memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2559,"prompt_tokens":929,"completion_tokens":1630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":1544}},"tokens_in":545,"tokens_out":1630,"duration_ms":12262,"temperature":1.0,"reasoning_tokens":1544,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:44:58.417022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Bake one mesh with the fixed association under adversarial real scenes—strong directional sources, large occluders, and light distributions absent from the rotation-based training—then compare its mRMSE and visible light leakage against the same mesh with probes optimized in that actual scene; a large quality gap would falsify the generalization claim.","supporting_citations":[],"review_version":1}