{"id":"81abcdcc-bb7a-4b85-818a-03ef0d1f06fe","arxiv_id":"2607.19731","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new CGH pipeline converts multiplane-image stacks into random-phase holograms with wave-optics alpha compositing, matching Gaussian-based hologram quality while being orders of magnitude faster.","lead":"This paper proposes a fast computer-generated-holography pipeline that converts multiplane images (a layered 3D scene representation) into holograms via wave-optics compositing. It claims up to 250,000x speedups over primitive-based hologram rendering with comparable quality, which could make interactive 3D holographic displays more practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The recursive wave-compositing law (Eq. 2) is the load-bearing assumption: no derivation, contrast bound, or speckle analysis is provided for its sqrt(1-α)/sqrt(α) split, and the claimed natural defocus, occlusion, and RPWS-parity quality all depend on it.","rationale":"The reader's weakest_assumption correctly identifies Eq. 2 as the load-bearing physical assumption. The paper gives no derivation, no contrast bound, and no speckle analysis for the random-phase compositing rule; it simply reuses a formula from the authors' prior RPWS work. All of the paper's quality claims—natural defocus, accurate occlusion, and 'comparable to RPWS'—are consequences of this rule holding. If the rule is physically inaccurate, the method's main advantage evaporates, even though the runtime advantage is real. The surrogate validation (MPIs derived by binning Gaussian splats via Eq. 4) is a secondary concern: it limits generality but does not threaten the core algorithm if Eq. 2 is valid. I therefore agree with the reader's choice. The verdict should remain CONDITIONAL: the speed claim is credible and the experiments are suggestive, but the physical model needs a formal derivation or a controlled validation before the quality parity claim can be accepted. A targeted two-layer experiment, both in simulation and on the benchtop, would directly test the compositing law and either confirm or refute the paper's central assumption.","tokens_in":13767,"tokens_out":12728,"duration_ms":143890,"concrete_test":"Construct a canonical two-layer MPI: a uniform back layer with intensity I_b (random phase) and a front layer with constant alpha α and color c (random phase). Synthesize MPI-CGH holograms using Eq. 2 with T = 1, 4, 24, 64. Numerically propagate the time-averaged field to the front and back focal planes and measure the reconstructed intensity I_meas. Compare to the predicted over-compositing result I_pred = (1−α) I_b + α c. Repeat on the benchtop with the described SLM setup for a subset of α values. If I_meas shows a systematic bias (not shrinking as T increases) or if the speckle contrast decays more slowly than 1/sqrt(T), then Eq. 2 is not a faithful compositing law and the core quality claims fail. A null result (bias < 1–2% and speckle scaling consistent with 1/sqrt(T)) would corroborate the law.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quality claim—comparable to RPWS and superior to RGBD holograms—rests entirely on Eq. 2, which combines the accumulated background field u_i with a layer's own random-phase field as u_{i-1} = P( sqrt(1−α_i) u_i + sqrt(c_i) sqrt(α_i) e^{i φ_i}; Δz ). This law is adopted from the authors' RPWS work, but the present paper provides neither a physical derivation nor an error analysis. The sqrt weights reproduce standard intensity-level over compositing only after averaging over random phases; the paper does not bound the residual interference for finite T (they use T=24), nor does it analyze the validity of treating a semi-transparent MPI plane as an amplitude attenuator plus additive emitter. If this compositing law does not faithfully describe propagation through partially transparent layers, then the focal-stack and light-field reconstructions in Figs. 4, 7, and 8—and the qualitative parity with RPWS—are not established. This is independent of the speedup claim, which is structural and credible; the fragility is specifically in the physical optics of the recursion. The paper's own limitation statement ('random phase leads to lower contrast') acknowledges one symptom but provides no quantitative bound, so the concern is not resolved by the experiments as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a wave-optics CGH pipeline that converts multiplane images (MPIs) into random-phase complex holograms via recursive back-to-front propagation and compositing (Eq. 2), with optional time multiplexing (Eq. 3). The claimed contributions are (1) an efficient direct CGH algorithm for MPIs that achieves up to 250,000× speedup over the primitive-based RPWS baseline while maintaining comparable reconstructed focal-stack and light-field quality, and (2) a detailed evaluation against RGBD layer-based and Gaussian-based methods. The method is validated on synthetic and captured benchtop results, including a demonstration with a neural-network MPI predictor (Stereo Magnification). The central claim is that MPI-based CGH offers the speed of RGBD methods and the quality of Gaussian primitive methods.","tokens_in":14050,"tokens_out":4378,"duration_ms":47170,"significance":"If validated, this is a significant advance: it links readily available neural-rendering representations (MPIs) to holographic displays, potentially enabling interactive CGH from casually captured scenes. The paper is clearly written and contains reproducible elements (source code/to-be-released datasets, PyTorch implementations, benchtop captures). The experimental captures in Fig. 8 are noteworthy, and the inclusion of a neural-MPI predictor in Fig. 6 strengthens the claim of representation agnosticism. However, the quantitative support for the headline speed/quality tradeoff is not fully rigorous, and the physical model underlying the recursion is imported from prior work without derivation or error analysis in this manuscript.","major_comments":[{"comment":"The recursive compositing law u_{i-1} = P( sqrt(1−α_i) u_i + sqrt(c_i) sqrt(α_i) e^{i φ_i}; Δz ) is the physical core of the paper, yet it is adopted from “recently proposed” work [Chao et al. 2025] with no derivation, no statement of the domain of validity for semi-transparent MPI planes, and no error bound. The paper itself acknowledges in Sec. 5 that “random phase leads to lower contrast” but does not quantify this or provide a contrast/speckle analysis vs. T or α. Since the claims of natural defocus, occlusion, and parity with RPWS all rest on this law, the paper should either derive it directly from wave optics or provide a quantitative validation (e.g., comparing against brute-force wave-propagation through a discretized volume for representative scenes, or bounding the interference error for finite T). As written, the physical correctness of the central quality claim is not self-c","section":"Fig. 3 and Sec. 4.2.1"},{"comment":"The headline 250,000× speedup is computed from the CGH propagation/compositing runtime only. The time to extract MPIs from Gaussian splats using Eq. 4, and the SGD-based phase-only encoding step (Sec. 4.1.4), are excluded. If the comparison is intended to reflect end-to-end hologram synthesis from a scene representation, the preprocessing costs should be included or explicitly excluded with a clear statement. The paper should report total pipeline times for a representative scene for all methods, so the reader can assess the practical speedup.","section":"Fig. 3 and Sec. 4.2.1"},{"comment":"Quantitative metrics (PSNR, SSIM) are reported only as averages over “selected scenes,” without per-scene values, standard deviations, or statistical significance tests. The claim that MPI “matches RPWS in SSIM” is not supported if the difference is within measurement noise. Please provide a per-scene breakdown (e.g., table or box plot) and state the number of scenes and selection criteria. This is essential to substantiate the core quality parity claim.","section":"Sec. 4.1.2"},{"comment":"The evaluation uses MPIs extracted from pre-optimized Gaussian splats via Eq. 4, not MPIs produced by the neural predictors that motivate the work. Fig. 6 shows only a qualitative example with Stereo Magnification. The paper should either provide quantitative results on neural-network MPIs (even a few scenes) or discuss explicitly how the Gaussian-splat-derived MPIs may differ in opacity/color statistics and therefore limit the generalization of the quality claims.","section":"Sec. 4.1.2"}],"minor_comments":[{"comment":"Typo: “addtionally” should be “additially” or “additionally.”","section":"General"},{"comment":"The 250,000× speedup corresponds to the lowest layer count (4) and should be reported together with the scaling: at 512 layers the speedup is far smaller. Please clarify in the abstract or the figure caption that this is a best-case value.","section":"Abstract and Fig. 3"},{"comment":"The phase-only encoding step (SGD with complex loss) is an additional cost that is common to all compared methods. Please state explicitly that the reported speedup applies to complex-valued hologram generation, not to the phase-encoding step.","section":"Sec. 4.1.4"},{"comment":"The paper fixes T=24 for time multiplexing without an ablation. Since the random-phase compositing law converges to the target intensity only in expectation, please provide a brief study of the effect of T on image quality and runtime, or at least a sentence justifying the choice.","section":"Sec. 4.2.1"},{"comment":"The caption states “matching RPWS in SSIM,” but the plotted MPI SSIM appears slightly lower than RPWS. Clarify whether “matching” is meant perceptually or statistically, and adjust the wording accordingly.","section":"Fig. 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid contribution to a top graphics venue, but the revision should address the physical grounding of Eq. 2 and the completeness of the evaluation. The stress-test concern about Eq. 2 is legitimate: because the paper does not derive the law itself and provides no error analysis, the central quality claim rests on an assertion. The speedup is credible but needs proper scoping. I recommend major revision rather than rejection, as the issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this paper proposes converting multiplane images into random-phase holograms by sequentially propagating and compositing each RGBA layer with a weighted random-phase field, plus optional time multiplexing. The genuinely new piece is that sequential compositing recipe for MPI layers and the demonstration that it gives RPWS-like focal stacks and parallax with far fewer layers. That is a real subfield advance: it gives CGH a practical interface to neural rendering, since MPIs are compact, feed-forward trainable, and naturally front-facing. The speedup claim is structurally credible—processing tens of layers instead of millions of primitives should be orders of magnitude faster—and the simulation and benchtop captures consistently show MPI beating RGBD and approaching RPWS.\n\nWhat the paper does well: the pipeline is simple, clearly explained, and the experiments are reasonably thorough for a graphics paper. The ablation over layer counts is useful. The Stereo Magnification demo (Fig. 6) shows the method works with off-the-shelf neural MPIs, not just the binning pipeline, which addresses part of the generality concern. The time-multiplexing and eyebox discussion is a nice practical touch. Credit where due: the paper ships no code yet, but the method is straightforward to reimplement and the promises are concrete.\n\nSoft spots, in proportion to how soft they actually are. The main one is Eq. 2: the sqrt(1−α)/sqrt(α) split between the propagated background and the layer’s own random-phase field is adopted from the authors' own RPWS paper and given no derivation or error bound here. The random-phase averaging argument is plausible, but the finite-T behavior (they use T=24) and the physical fidelity for partially transparent layers are not analyzed. The paper's own limitation statement admits random phase lowers contrast, but gives no quantitative bound. That is a real gap, though I would not call it fatal: the experimental parity with RPWS and the improvement over RGBD are consistent across scenes and captured results, so the law is at least empirically reasonable. Still, a referee should push for a derivation or at least a speckle/contrast analysis for finite T.\n\nSecond, the 250,000× speedup is the most favorable corner (4 layers) and excludes MPI extraction and phase-encoding costs. That is common in this literature, but the abstract's phrasing makes it easy to over-read. The paper does disclose the layer-count dependence in the runtime plot, so I see this as a framing issue, not a deceptive one.\n\nThird, metrics are averages over selected scenes with no per-scene breakdown or error bars. Minor, given the qualitative figures are convincing.\n\nThe circularity concern is mild. Yes, Eq. 2 comes from the authors' own work and the main baseline is their own RPWS, but the independent Stereo Magnification demo gives genuine external grounding. Self-citation is not a sin when the cited result is the actual physical model.\n\nWho this is for: anyone working on CGH for near-eye displays, especially people bridging neural rendering and holography. It deserves a serious referee; the central idea is useful and the validation is mostly sound. My recommendation: send it out, but require a derivation or bounds for Eq. 2, per-scene numbers, and a clearer statement of what the speedup includes before acceptance.\n\nBest,\n[You]","headline":"A plausible and useful pipeline paper: MPI-to-hologram via random-phase wave compositing, with a credible speedup claim and decent validation, but the load-bearing compositing law is adopted without derivation or error bounds and the headline speedup excludes upstream costs.","tokens_in":14640,"tokens_out":828,"would_cite":true,"duration_ms":11807,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new rendering pipeline converts multiplane images into random-phase holograms, matching primitive-based CGH quality while running up to 250,000× faster.","keywords":["multiplane images","computer-generated holography","random-phase holograms","wave-optics rendering","angular spectrum method","time multiplexing","light field","focal stack"],"falsifier":"Render a scene with a sharp occlusion boundary as both an MPI hologram and a primitive-based reference hologram, then capture focal slices at several depths on a benchtop display and compare the intensity profile across the edge; visible ringing or light leakage in the MPI case but not in the reference would refute the compositing law.","tokens_in":13572,"feed_emoji":"🥽","tokens_out":7182,"duration_ms":72768,"temperature":0.7,"pith_summary":"Multiplane images (MPIs) are a layered 3D representation—a stack of RGBA images at discrete depths—that neural rendering has made easy to acquire, yet until now they have been viewed as 2D pictures. This paper proposes a wave-optics rendering pipeline that treats each MPI layer as a random-phase amplitude mask and sequentially propagates and composites the layers into a complex hologram at the display plane. The central claim is that this MPI-based CGH achieves the image quality of primitive-based computer-generated holography (which must composite millions of individual Gaussian or polygon wavefronts) while running up to 250,000× faster, and that it avoids the defocus and occlusion artifacts of RGBD-based layer CGH. This matters because holographic near-eye displays need an efficient path from casually captured or neural-rendered scene content to the patterns on a spatial light modulator, and MPIs are a compact, readily available intermediate format for front-facing scenes. The paper supports the claim with simulated focal stacks and light fields, benchtop captures on a phase-only SLM, and a demonstration with neural-network-predicted MPIs.","feed_headline":"MPI-based holography: Gaussian-level quality at 250,000x speed","feed_subtitle":"The layered scene format neural rendering already produces becomes a practical input for real-time holographic displays.","key_machinery":"The load-bearing object is the sequential wavefront compositing rule in Eq. 2, adopted from the authors' earlier random-phase wave-splatting work. It defines how the accumulated field from deeper layers is attenuated by the current layer's transmittance √(1−α_i) and amplitude-weighted by √α_i with a random phase e^{iφ_i} before both are propagated by the angular spectrum method across the inter-layer gap. The rule is what turns an RGBA stack into a complex hologram in a single forward pass, and its correctness is what makes the natural-defocus, occlusion, and RPWS-parity results follow. The accompanying time-multiplexing rule (Eq. 3) averages intensities over T independent random-phase rende","core_discovery":"At the paper's core is the recursive compositing equation u_{i-1} = P( √(1−α_i) u_i + √c_i √α_i e^{iφ_i}; Δz ), which takes the accumulated complex field from the far side of the stack, propagates it across the inter-layer gap with the angular spectrum method, and blends in the current layer's color weighted by the square root of its opacity and multiplied by a layer-specific random phase. Iterating from the backmost to the frontmost layer produces the full complex wavefront at the SLM plane. Repeated with different random phase maps and time-averaged, the pipeline reconstructs 3D focal stacks and 4D light fields with natural defocus, parallax, and occlusion. The paper's discovery is that a","pith_inferences":["The paper's quantitative MPI evaluations are built mostly from binning optimized Gaussian splats (Eq. 4), so the claim that the pipeline is agnostic to MPI acquisition rests on a single qualitative neural-network demo; a systematic benchmark with predicted MPIs would directly test the motivating use case.","If Eq. 2 holds as a physical model, the same compositing scheme should transfer to other layered representations—layered depth images, alpha-matted RGBD, or learned layered radiance fields—with the same cost model, but the paper does not test these extensions.","Because the per-layer cost is essentially two Fourier transforms (propagation and composite), pairing the pipeline with a neural MPI predictor that emits 16–30 planes is the most direct route to real-time CGH; the paper identifies this direction but does not demonstrate it.","The paper inherits a known weakness of random-phase holograms—reduced image contrast—and mentions learning-based calibration as a remedy, but offers no quantitative measurement of the contrast cost or eyebox gain as T varies; a characterization experiment would help system designers choose T."],"forward_implications":["Scenes already represented as MPIs—whether from optimized splats, stereo networks, or single-image predictors—can be turned into holograms without per-scene optimization.","Runtime scales about linearly with the number of depth planes, and perceptual quality saturates near 30 layers, making a ~30-layer MPI a practical operating point for interactive displays.","Time multiplexing of independent random-phase renders reduces speckle contrast and expands the eyebox, improving space–bandwidth product utilization.","MPI holograms eliminate the halo, light-leakage, and dark-border occlusion artifacts that RGBD-based layer CGH exhibits at depth transitions."],"fun_headline_variants":["MPI wave-optics: 250,000x faster holograms","Holographic display rendering via multiplane images","Multiplane-image CGH achieves 250,000x speedup","Wave-optics MPI: real-time holographic rendering"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on Eq. 2: that a partially transparent MPI layer behaves in wave optics as an amplitude mask with transmittance √α carrying a random phase—a law taken from earlier work without derivation or direct physical measurement in this paper.","fun_headline_variants_meta":{"raw":{"variants":["MPI wave-optics: 250,000x faster holograms","Holographic display rendering via multiplane images","Multiplane-image CGH achieves 250,000x speedup","Wave-optics MPI: real-time holographic rendering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1553,"prompt_tokens":790,"completion_tokens":763,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":692}},"tokens_in":534,"tokens_out":763,"duration_ms":7560,"temperature":1.0,"reasoning_tokens":692,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:51:54.589126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a scene with a sharp occlusion boundary as both an MPI hologram and a primitive-based reference hologram, then capture focal slices at several depths on a benchtop display and compare the intensity profile across the edge; visible ringing or light leakage in the MPI case but not in the reference would refute the compositing law.","supporting_citations":[],"review_version":1}