{"id":"5b2fcd5f-1861-4fa8-8fc8-3b9d6e0f4b7d","arxiv_id":"2504.17062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An extended intrinsic image representation that adds a transparency channel and an explicit compositing pipeline for synthesizing and editing transparent materials such as glass.","lead":"The paper adds a transparency channel to standard intrinsic-image material maps, letting glass and windows be edited and recomposed in screen space. It proposes a deterministic compositing recipe based on established reflection and transmission models, instead of learned generation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transmission term reuses the reflection Fresnel factor, making Eq. 21's glass ~25x too dark at normal incidence; the central transparent-material result is unsupported as written.","rationale":"The reader's weakest assumption was the thin-surface model (zero thickness, parallel faces, negligible refraction offset). That is a legitimate scope limitation and is explicitly acknowledged in the paper. On re-reading, a more fundamental issue appears earlier in the derivation: the transmission lobe in Eq. 8 reuses the Fresnel reflectance F and Eq. 21 then scales the transmitted layer by the same integrated reflectance factor (AF0+B) as the reflection layer. For typical window glass this makes the transmitted background roughly 25x too dark at normal incidence and reverses the angular behavior. This is not an external disagreement about consensus or a missing secondary effect; it is an inconsistency between Eq. 21 and the elementary behavior of a thin dielectric slab. It matters for the central claim: the LPIPS comparison in Table 2 is for composed images that include glass regions, so if the transmission layer is unphysically dark, the reported 'outperforms diffusion' numbers do not validate the model. I therefore recommend REJECT as written, with the path to revision being a corrected Fresnel factor for transmission and a re-run of the comparison. This is a partial agreement with the reader: the thin-surface concern is real, but the Fresnel-factor error is more load-bearing and testable.","tokens_in":12035,"tokens_out":9979,"duration_ms":100841,"concrete_test":"Set up the minimal case: planar thin glass, F0=0.04, A=1, R≈0, Abg=1, Amr=0, T=1, and evaluate Eq. 21 at normal incidence. The stated model gives Itran≈0.04. Compute the reference with a two-interface thin-slab formula or one path-traced frame: Itran≈(1-F)^2≈0.92. Also evaluate at large view angle (θ→80°): Eq. 21 raises the factor toward 1 while the physical transmittance drops toward 0. If these discrepancies reproduce, the transmission term must be reformulated (e.g., using 1-F with the correct index-of-refraction prefactor) before any claim about transparent material synthesis can be evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing problem is not the acknowledged thin-surface approximation but the Fresnel factor used for transmission. Eq. 8 defines the BTDF with F(ht,ωo), the same Fresnel reflectance as the reflection lobe, and Eq. 21 then scales the transmitted background by (AF0+B), the same precomputed reflectance integral used for Ispec. For a smooth dielectric with F0=0.04, that factor is about 0.04 at normal incidence; physically a thin glass slab transmits about (1-F)^2 ≈ 0.92. The error is not merely a missing energy-conservation term: it is wrong physics at the qualitative level. At grazing incidence the physical transmittance (1-F)^2 goes to zero, while Eq. 21's factor grows toward 1, so the model predicts bright background at grazing and dark background at normal incidence—the opposite of glass. This is independent of the thin-surface assumption: even for an ideal thin parallel slab, the transmission layer in Eq. 21 is dark by roughly 24x at normal view. It also contradicts the paper's own interpretation that A=(1,1,1) represents 'pure glass with no energy loss' (Sec. 5.2), which would require the transmitted background to pass almost unchanged. Because transparency is the central extension and the LPIPS comparison in Fig. 7 depends on it, the core claim as written is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ePBR, an extension of intrinsic image representations to support transparent materials. The method augments standard intrinsic channels (albedo, roughness, metallic, normal, depth, irradiance, mirror reflectance, background radiance) with a transparency map T stored in the blue channel of the PBR material image. The rendering model combines a diffuse Lambertian term, a microfacet specular reflectance term, and a specular transmittance term under a thin-surface assumption, then derives screen-space compositing equations: a split-sum style reflection layer, a transmission layer obtained by applying the GGX blur kernel twice to the background and multiplying by albedo, and a final composition I = (1-T)(1-M)Idiff + Ispec + TItran. The authors evaluate the approach on five InteriorVerse scenes and report lower LPIPS than the diffusion-based RGB↔X baseline, arguing that their deterministic composition provides better control and fidelity for high-specular and transparent regions.","tokens_in":12288,"tokens_out":2212,"duration_ms":23104,"significance":"If the model worked as claimed, it would offer a practical lightweight alternative to Monte Carlo rendering for editing and compositing transparent objects in screen-space intrinsic representations, with the transparency map occupying no extra memory. The paper clearly states its thin-surface simplification and identifies future extensions. However, the central claim depends on the physical correctness of the transmission term, and that term currently uses a reflectance Fresnel factor instead of a transmittance factor, producing qualitatively wrong brightness behavior for glass. The empirical validation is also limited to five scenes with manually adjusted channels, no error bars, and no code release, so the headline 'outperforms diffusion-based method' is not yet established. The paper's contribution is conceptually useful but requires a corrected transmission model and stronger validation.","major_comments":[{"comment":"The transmission term uses the same Fresnel reflectance factor F(ht, ωo) as the reflection lobe, and Eq. (21) scales the transmitted background by (AF0+B), the precomputed reflectance integral. For a smooth dielectric with F0=0.04 at normal incidence this yields a transmittance of about 0.04, whereas a thin glass slab transmits approximately (1-F)^2 ≈ 0.92. At grazing angles the physical transmittance (1-F)^2 tends to zero, while Eq. (21)'s factor grows toward 1, predicting the opposite qualitative behavior of glass. This contradicts the paper's own interpretation that A=(1,1,1) represents 'pure glass with no energy loss' in Sec. 5.2. Since transparency is the core extension and the LPIPS comparison in Fig. 7 depends on the resulting transmission layer, the central claim is not supported as written. The transmission model should use a proper transmittance Fresnel factor, e.g. (1-F)^2 for a thin slab, and the precomputed lookup should be recomputed accordingly.","section":"Sec. 3.3, Eq. (8) and Sec. 4.3, Eq. (21)"},{"comment":"The claim that the method 'outperforms diffusion-based method for the entire image' rests on five scenes with no error bars or statistical testing, and the comparison setup is partially hand-tuned: T is set as the inverse of the ground-truth mask, M and T are manually modified to fit the renderings, Abg is set to 1, and E is estimated from RGB↔X. This does not establish a general advantage, especially for the transparent material aspect that is the paper's central novelty. The authors should either provide a broader quantitative comparison with fixed, non-manual channel generation, or restrict the claim to a qualitative demonstration.","section":"Sec. 5.3 and Table 2"},{"comment":"The reflection and transmission layers both use the same precomputed (AF0+B) factor, but the transmission layer in Eq. (21) is additionally multiplied by albedo A to approximate absorption. This conflates the reflectance-based Fresnel weight with a volume absorption effect and can darken or color-shift the transmitted background in a way that is not physically motivated. At minimum, the paper should state this as an ad hoc approximation with a quantitative demonstration of its effect, rather than presenting it as part of the physically derived compositing model.","section":"Sec. 4.2, Eq. (19) and Eq. (21)"}],"minor_comments":[{"comment":"The half-vector definition ht = −(ωo + ηωi)/||ωo + ηωi|| uses η but the refractive index is not defined in the text; please clarify the notation and its relationship to F0.","section":"Sec. 3.3, Eq. (8)"},{"comment":"The roughness evaluation compares 'Ours' with path-traced ground truth, but no numerical error is reported and the figure is small; adding quantitative roughness-sweep errors would strengthen the claim that the double-convolution approximation is accurate.","section":"Fig. 4"},{"comment":"The albedo row says that with A=(0,0,0) 'light cannot scatter or pass through the surface', but Eq. (21) multiplies the transmitted background by A, so a pure black albedo would make the background completely invisible; the physical interpretation of A as both diffuse reflectance and transmission absorption should be discussed more carefully.","section":"Sec. 5.2, Fig. 6(d)"},{"comment":"The sentence 'T is the inverse of the ground-truth mask since we found that most of the nonmasked areas are glasses' is unclear; please specify which mask is used and how T is aligned with the object regions in Fig. 7.","section":"Sec. 5.3"},{"comment":"The paper would benefit from a release of the compositing code and the exact intrinsic-channel inputs used for the five test scenes, since the method is deterministic and reproducible only if those inputs are specified.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful idea and a clear presentation, but the transmission Fresnel error is a genuine physical flaw that affects the central contribution, not a cosmetic issue. The empirical section is also too thin for the strength of the claimed advantage over RGB↔X. I would ask for the transmission model to be corrected and validated against path-traced references for glass at varying incidence angles, and for a more controlled quantitative comparison, before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's transmission term is the load-bearing piece, and it's wrong. Eq. 8 defines the BTDF with the same Fresnel reflectance F used for reflection, and Eq. 21 then scales the transmitted background by (AF0+B), the precomputed reflectance integral from the split-sum. For dielectric glass at normal incidence, that factor is about 0.04, whereas physical transmission through a thin slab is ~0.92. So the model predicts dark glass at normal view and brighter glass at grazing, the opposite of reality. This is not just a missing energy-conservation term; it's a qualitative inversion of the physics.\n\nThat said, the conceptual core is genuinely new: putting a transparency channel T into the screen-space intrinsic map and compositing with a deterministic, blur-based rule, no Monte Carlo. The double-convolution approximation for eNDF transmission is a reasonable trick. The paper is honest about the thin-surface assumption and other simplifications, which I appreciate.\n\nThe empirical side is thin: five scenes, hand-adjusted M and T, no error bars, no code, and the only baseline is RGB↔X. The LPIPS table shows moderate gains, but with that setup I wouldn't give it much weight.\n\nThe stress-test note is correct. The Fresnel error makes the central 'glass' result unsupported as written. The fix is straightforward—use (1-F) for transmission and recompute the preintegrated factor—but until then the paper's main claim doesn't stand. The rest of the framework, especially the specular reflection compositing, is plausible but not independently verified.\n\nFor readers in intrinsic-image editing or real-time screen-space rendering, the compositing rule could be useful if the physics were corrected. As is, a serious referee would likely reject, but the error is localized and the idea has merit. It deserves reviewer time, not a desk rejection.\n\nSkeptical but not dismissive.","headline":"The transparency term is physically wrong, so the paper's central claim about synthesizing glass is unsupported, though the compositing idea is worth a fix.","tokens_in":12831,"tokens_out":4794,"would_cite":false,"duration_ms":44987,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding one transparency channel to intrinsic image maps turns single-image compositing into deterministic, editable glass synthesis.","keywords":["intrinsic image decomposition","physically based rendering","transparent materials","specular transmission","thin surface model","screen-space compositing","ePBR materials","image editing"],"falsifier":"Render a thick transparent object, such as a water glass or bottle, in front of a textured background both with a path tracer and with this compositing rule using the same roughness and transparency; if the twice-blurred background shows visible displacement, double imaging, or distortion that grows with thickness or viewing angle, the zero-thickness premise fails. A simpler screen-space variant is to compare Eq. (20) against a Monte Carlo reference for one oblique-angle glass slab and look for background misalignment beyond one pixel.","tokens_in":11807,"feed_emoji":"🪟","tokens_out":8578,"duration_ms":73930,"temperature":0.7,"pith_summary":"The paper tries to establish that a single extra scalar channel—a transparency map $T$—is enough to extend standard intrinsic-image/PBR material representations from opaque surfaces to transparent ones such as glass and windows. It derives a thin-surface BSDF whose transmission term is a microfacet lobe, then shows that in screen space this term becomes two applications of the same blur kernel to the background radiance. Combining the resulting layers with the explicit rule $I=(1-T)(1-M)I_{\\mathrm{diff}}+I_{\\mathrm{spec}}+TI_{\\mathrm{tran}}$ gives deterministic, interpretable image synthesis that needs no Monte Carlo sampling. Against the diffusion-based baseline on the five tested indoor scenes, the paper reports lower LPIPS over the whole composed image, with the largest gains in mirror-like regions. If this holds, glass-like materials become directly editable in single images by changing $T$, roughness, or albedo and re-applying basic image operations.","feed_headline":"One transparency channel makes glass editable in single images","feed_subtitle":"Screen-space blur replaces path tracing for glass, and the composed image beats a diffusion baseline on perceptual error.","key_machinery":"The central mechanism is the thin-surface BSDF together with its screen-space reduction: specular transmission is treated as the same microfacet lobe used for reflection but evaluated with an extended normal distribution function (eNDF), and under the zero-thickness assumption this collapses to applying the same GGX blur kernel twice to the background radiance map. The compositing identity $I=(1-T)(1-M)I_{\\mathrm{diff}}+I_{\\mathrm{spec}}+TI_{\\mathrm{tran}}$ then assembles the final image from diffuse irradiance, screen-space ray-traced mirror reflections, and the twice-blurred background, with transparency $T$ stored in the unused blue channel of a standard three-channel PBR map.","core_discovery":"On the paper's own terms, the central discovery is that the transmission behavior of a transparent thin surface can be folded into an intrinsic-image compositing equation. Whereas existing intrinsic/PBR representations only model diffuse and specular reflectance, ePBR adds specular transmittance through an extended normal distribution function and expresses the transmitted radiance as the original background convolved twice with the same GGX kernel. With the transparency map $T$ modulating the transmission layer and $(1-T)(1-M)$ modulating diffuse, the screen-space rule reproduces the path-traced references in high-specular regions and, on LPIPS, beats the diffusion-based recomposition for all five tested scenes.","pith_inferences":["A natural stress test is to repeat the five-scene comparison with fully automatic intrinsic estimation, since the reported results use manually corrected metallic and transparency maps; if the LPIPS advantage shrinks, the explicit compositing still wins only when its input channels are trustworthy.","The same double-blur trick could be extended to front/back roughness by replacing the identical second blur with a second kernel parameterized by a different roughness, which the paper notes is physically more accurate for glass.","Because the composition is differentiable and parameter-light, it could serve as a cheap generative prior for training decomposers that output $T$ directly, avoiding expensive path-traced transparent-object datasets."],"forward_implications":["Setting $T$ and recomposing with the blur kernels edits a window's transparency and roughness in one image, with no path tracing.","Because $T$ occupies the blue channel that older PBR maps leave unused, ePBR materials fit existing intrinsic-image pipelines without extra memory.","The composition is deterministic and uses only basic image operations, so resolution and GPU memory no longer limit the synthesis step.","The method's validity is limited to thin, nearly flat transparent surfaces; thick or strongly refractive objects require geometry the screen-space representation does not provide.","If ePBR maps replace PBR maps in inverse-rendering and decomposition data, learning-based systems could inherit explicit transparency control, a direction the paper names as future work."],"supporting_citations":[{"why":"Supplies the extended normal distribution function and joint spherical warping that turn thin-slab transmission into a microfacet lobe.","marker":"[13]"},{"why":"Supplies the split-sum factorization that separates each radiance integral into a material preintegral and a lighting convolution, enabling blur-kernel compositing.","marker":"[21]"},{"why":"Provides the GGX distribution and microfacet refraction treatment that justify using a normalized blur kernel whose shape tracks roughness.","marker":"[42]"},{"why":"Defines the diffusion-based decomposition/synthesis baseline and provides the estimated irradiance map used in the comparison.","marker":"[50]"},{"why":"Supplies the five test scenes, ground-truth intrinsic maps, and the screen-space ray-tracing method used to build the mirror-reflection layer.","marker":"[55]"},{"why":"Extends the Disney BRDF to a full BSDF with transmission, the appearance model this work simplifies into the ePBR material.","marker":"[4]"}],"fun_headline_variants":["Screen-space blur replaces path tracing for glass editing","ePBR adds a transparency channel to intrinsic images","Double GGX blur bakes transmission into single images","Intrinsic images now model glass with a second blur","Editable glass without ray tracing via extended PBR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the thin-surface model: the transparent object is treated as an infinitely thin slab with two parallel faces, negligible refraction offset, no internal reflection, and equal roughness on both sides; if the target is thick, bent, or viewed at steep angles, the transmitted layer is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Screen-space blur replaces path tracing for glass editing","ePBR adds a transparency channel to intrinsic images","Double GGX blur bakes transmission into single images","Intrinsic images now model glass with a second blur","Editable glass without ray tracing via extended PBR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1799,"prompt_tokens":814,"completion_tokens":985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":910}},"tokens_in":430,"tokens_out":985,"duration_ms":8540,"temperature":1.0,"reasoning_tokens":910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:50:35.203209+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a thick transparent object, such as a water glass or bottle, in front of a textured background both with a path tracer and with this compositing rule using the same roughness and transparency; if the twice-blurred background shows visible displacement, double imaging, or distortion that grows with thickness or viewing angle, the zero-thickness premise fails. A simpler screen-space variant is to compare Eq. (20) against a Monte Carlo reference for one oblique-angle glass slab and look for background misalignment beyond one pixel.","supporting_citations":[{"cited_title":"Render- ing thin transparent layers with extended normal distribution functions","cited_arxiv_id":null,"evidence_quote":"Supplies the extended normal distribution function and joint spherical warping that turn thin-slab transmission into a microfacet lobe."},{"cited_title":"Real shading in Unreal Engine","cited_arxiv_id":null,"evidence_quote":"Supplies the split-sum factorization that separates each radiance integral into a material preintegral and a lighting convolution, enabling blur-kernel compositing."},{"cited_title":"Robust Monte Carlo methods for light transport simulation","cited_arxiv_id":null,"evidence_quote":"Provides the GGX distribution and microfacet refraction treatment that justify using a normalized blur kernel whose shape tracks roughness."},{"cited_title":"Stablenormal: Reducing diffusion variance for stable and sharp normal","cited_arxiv_id":null,"evidence_quote":"Defines the diffusion-based decomposition/synthesis baseline and provides the estimated irradiance map used in the comparison."},{"cited_title":"Dreammat: High-quality PBR material generation with geometry-and light-aware diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the five test scenes, ground-truth intrinsic maps, and the screen-space ray-tracing method used to build the mirror-reflection layer."},{"cited_title":"Extending the Disney BRDF to a BSDF with integrated subsurface scattering","cited_arxiv_id":null,"evidence_quote":"Extends the Disney BRDF to a full BSDF with transmission, the appearance model this work simplifies into the ePBR material."}],"review_version":1}