{"id":"c8a22387-3b11-4e36-ba3b-077a2ab130f5","arxiv_id":"2504.16636","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DC-NeRF synthesizes all-in-focus novel views from smartphone main and ultra-wide image pairs with view-consistent defocus blur by aligning the pair and learning a defocus-aware fusion mask.","lead":"A smartphone's main camera and ultra-wide camera are combined to train a neural radiance field that produces sharp, all-in-focus views even when every main-camera image shares the same defocus blur. The method, DC-NeRF, aligns the two cameras and learns where to trust the sharp ultra-wide image instead of the blurred main image.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's central assumption that the ultra-wide camera is sharp enough to serve as an all-in-focus reference is not validated and is contradicted by the paper's own failure case, so the all-in-focus claim is conditional at best.","rationale":"The reader's weakest_assumption correctly identifies the unvalidated sharpness of the ultra-wide camera as the load-bearing condition. My stress-test review finds independent support for this in the manuscript itself: the failure case in Sec. 5.7 explicitly states that low-quality ultra-wide regions cannot be fixed by the pipeline, and the fusion losses (Eq. 17, L_fusion) only involve the UWA color, the main color, and the synthetic mask, with no external sharp reference to enforce all-in-focus quality. This is not a mere evaluation gap; it is a structural limitation of the method's core premise. If the UWA is not sharp, the method cannot produce an all-in-focus result in those regions, so the claim 'first framework capable of synthesizing the all-in-focus NeRF' is too strong. However, because the paper does demonstrate working results on several scenes and the limitation is a well-defined, addressable condition (clarify the scope or validate UWA sharpness), the appropriate verdict remains CONDITIONAL, which matches the reader's verdict. I therefore recommend no change to the verdict, and I agree with the reader's identification of the weakest assumption. The concrete test above would settle the concern by quantifying whether the UWA actually provides the sharp reference the method requires.","tokens_in":21490,"tokens_out":5880,"duration_ms":53099,"concrete_test":"On the 7-scene dataset, for each test view compute the main camera defocus mask from the estimated defocus map D_defocus (threshold at 0.5 after normalization). Within that mask, compute the ratio of high-frequency energy (e.g., mean squared gradient magnitude) of the aligned ultra-wide image I_w,a to that of the synthesized all-in-focus ground truth I_gt. For each scene, report this ratio. If any scene has a ratio below 1 (or below a pre-specified margin, e.g., 0.8), the UWA is not sharp enough to serve as the reference in that scene, and the all-in-focus claim fails there. Additionally, retrain DC-NeRF on a variant where I_w,a is blurred with a 1-pixel Gaussian and compare PSNR; a large drop would confirm that the method has no mechanism to recover UWA blur.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DC-NeRF synthesizes an all-in-focus NeRF from a main camera with consistent defocus blur plus an ultra-wide camera. The load-bearing assumption, stated in Sec. 1 and used in Sec. 4.3, is that the ultra-wide image is sharp enough to act as the all-in-focus reference in every region where the main camera is blurred. The paper never measures the residual blur of the ultra-wide camera. More importantly, its own failure case (Sec. 5.7, Fig. 18, row 2) admits that 'the ultra-wide views may have low-quality regions on some details, so the corresponding parts of the main camera cannot be fixed.' Because the fusion module is supervised only by (a) the main camera's blurred color via L_focus and (b) the synthetically mixed Cfuse = M_blend*Cw + (1-M_blend)*Cm in Eq. 17, there is no mechanism in the pipeline that can add detail beyond what the ultra-wide radiance field already contains. If the UWA is blurry in a region, the learned mask weights UWA heavily there, and the rendered output inherits that blur; no loss term can correct it. Thus the 'all-in-focus' property is not guaranteed by the framework; it is conditional on the UWA having sufficient resolution and sharpness. The paper's own experiment demonstrates a failure of this exact condition, so the central claim is overstated as a universal first framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DC-NeRF, a NeRF-based novel-view synthesis framework that inputs a smartphone dual-camera sequence: a main camera with consistent defocus blur (autofocus on the same target in all views) and an aligned ultra-wide camera. The method aligns the ultra-wide image to the main view via homography, optical flow, and histogram matching; trains separate NeRFs for the main and aligned ultra-wide images; learns defocus parameters A and D_f by fitting a scatter-based bokeh renderer to the main-camera observation; and fuses the two radiance fields with a learned blending mask to produce all-in-focus novel views. The paper also contributes a 7-scene smartphone dataset with focal-stack-derived ground truth, ablations of alignment and fusion components, and demonstrations of refocusing and split-diopter effects. The headline claim is that this is the first framework to synthesize all-in-focus NeRF from inputs without manual refocusing under view-consistent defocus blur.","tokens_in":21785,"tokens_out":6902,"duration_ms":60150,"significance":"If the claims hold, this is a well-motivated and useful contribution: it addresses a genuinely under-explored failure mode of NeRF (view-consistent defocus blur), exploits commodity smartphone dual-camera hardware, and the align-and-fuse pipeline is coherent and supported by component ablations. The dataset is a valuable resource, and the DoF applications show practical potential. The paper is honest about its failure cases and computational cost. However, the all-in-focus claim is stronger than the evidence: the ground-truth AiF images are synthesized from a focal stack by an external fusion method; the ultra-wide camera's sharpness is never measured and is explicitly acknowledged to fail in some regions; and the quantitative gains over strong baselines are marginal and not statistically characterized. These issues affect the central claim rather than the peripheral presentation.","major_comments":[{"comment":"The AiF ground truth is synthesized by multi-focus image fusion of only two focal-plane main images; it is not a truly sharp sensor capture. Any artifacts of the external fusion method (e.g., focus-breathing misalignment, halos) are treated as ground truth, and the fusion objective in Eq. (17) is a similar mask-based blending of the two cameras. This creates an evaluation bias toward mask-based fusion and does not directly measure the actual sharpness of the output. Please report a sensitivity analysis with an alternative GT fusion method, or evaluate on scenes with independently known all-in-focus captures.","section":"4.1, Eq. (4)"},{"comment":"The failure case states that 'the ultra-wide views may have low-quality regions on some details, so the corresponding parts of the main camera cannot be fixed.' This is exactly the condition on which the central all-in-focus claim rests: in Eq. (17), Cfuse is a weighted combination of C_w and C_m, so when C_w is blurred in a region, the learned mask has no sharp source to draw from and no loss term can recover the missing detail. The manuscript never measures the residual blur of the ultra-wide camera, and the failure case shows the assumption can fail in practice. Please add a quantitative sharpness/MTF measurement of the UWA inputs and an analysis of how often/where the assumption is violated, or explicitly re-scope the claim to 'all-in-focus when the ultra-wide camera is sharp in the region of interest.'","section":"5.7, Fig. 18"},{"comment":"The quantitative support for the headline claim is thin. In Table 1, DC-NeRF's average PSNR advantage over EasyAIF+NeRF is 0.07 dB (24.31 vs 24.24), and on Dove DC-NeRF has lower PSNR (22.72 vs 22.84); in Table 2, DC-NeRF is not the best on average LPIPS (0.209 vs Deblur-NeRF's 0.184); and in Table 3 the per-scene PSNR is worse than Deblur-NeRF on Stadium (24.53 vs 25.38). All numbers appear to come from a single training run with no variance estimates. Please report standard deviations over multiple seeds/runs and use a statistical test or consistent per-scene wins to support 'compares favorably.'","section":"5.3, Tables 1-3"},{"comment":"The transmittance T_f(t_i) is defined as exp(-Σ_{j<i} σ_m σ_w δ_j). This is not the standard volume-rendering transmittance for two radiance fields, which should involve a sum of densities (or a weighted sum) in the exponential; a product of densities is dimensionally inconsistent and would make occlusion depend on both fields being non-zero. Please either correct the equation to the actual implementation (e.g., T_f = exp(-Σ(η σ_w + (1-η)σ_m)δ) or similar) or explain and justify the product form. This is central to the fused rendering.","section":"4.3, Eq. (12)"},{"comment":"The defocus map is learned by fitting a bokeh renderer to the blurred main camera observation, and this same map defines the fusion target. Because the supervision in Eq. (15) only compares rendered bokeh to the already-blurred main ray, errors in A and D_f are not independently checked; the fusion target in Eq. (17) then inherits these errors. This is a self-supervised fitting loop rather than an independent defocus measurement. Please validate the estimated defocus map against a measured defocus/disparity (e.g., from the focal stack or an external depth sensor) or show an ablation that uses the ground-truth defocus map instead.","section":"4.3, Eqs. (10), (15), (17)"}],"minor_comments":[{"comment":"The word 'improficiencies' appears to be a typo for 'deficiencies'.","section":"Introduction"},{"comment":"Equation (5) is garbled in the manuscript; the confidence-map formula should be typeset with clear warping notation and thresholding so that the forward-backward consistency check is unambiguous.","section":"4.2, Eq. (5)"},{"comment":"The disparity formula is difficult to read because of the formatting; please rewrite it with explicit parentheses, e.g., D = 1 / (Σ_i T_i (1-exp(-σ δ_i)) t_i).","section":"4.3, Eq. (11)"},{"comment":"The term 'volumef' is used before it is defined; please define the fused volume color explicitly before presenting the rendering equation.","section":"4.3, Eqs. (12)-(13)"},{"comment":"The main text says that the inputs of baselines are only the main-camera images for Figs. 8 and 11, while Figs. 9 and 10 show baselines trained with ultra-wide inputs; please state the input setting in each figure caption explicitly to avoid confusion.","section":"5.2, Figs. 8-11"},{"comment":"The caption acknowledges that Deblur-NeRF achieves better LPIPS and asks readers to consult qualitative results; please report the magnitude of the difference and ideally confidence intervals so the trade-off is interpretable.","section":"5.3, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The novelty claim is plausible in the specific setting of multi-view consistent defocus blur, but the authors should consider softening the 'first framework capable' phrasing to make explicit the dependence on ultra-wide camera sharpness. The comparison to DC2 is weakened by the unavailability of DC2 code; this is acceptable if stated transparently. The main work needed is evaluative: independent or more rigorous ground truth, sharpness validation of the ultra-wide camera, and uncertainty-aware quantitative comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: this is the first NeRF method I know that takes the smartphone main/ultra-wide pair as multi-view input to handle view-consistent defocus blur. The pipeline is sensible — homography plus optical flow for alignment, histogram matching for color, then a learnable bokeh renderer that predicts a defocus map and a blending mask to fuse the two radiance fields. The ablations are fairly thorough and the comparisons against Deblur-NeRF, DoF-NeRF, DP-NeRF, Restormer+NeRF, and EasyAIF+NeRF are appropriate. I'd credit the authors for showing failure cases and stating limitations honestly.\n\nThat said, the paper's central claim — \"the first framework capable of synthesizing the all-in-focus NeRF from inputs without manual refocusing\" — is overstated. The load-bearing assumption is that the ultra-wide camera is sharp everywhere the main camera is blurred. The paper never measures the ultra-wide's residual blur, and its own failure case (Fig. 18, row 2) says low-quality ultra-wide regions cannot be fixed. Given Eq. 17, the fusion target is essentially a blend of the ultra-wide radiance field and the blurred main field, so there is no mechanism to recover detail beyond what the ultra-wide already contains. The all-in-focus property is conditional on the ultra-wide being good enough, not guaranteed by the framework.\n\nThere are also evaluation-level weaknesses. The ground-truth AiF images are synthesized from two focal-plane main images via an external fusion method, not captured by a genuinely sharp sensor. The dataset is private (7 scenes, 24 views each), no code or data are released, and the quantitative tables show a single run without error bars. That makes the reported margins over EasyAIF somewhat fragile. The defocus parameters A and D_f are learned by fitting a bokeh renderer to the blurred main observations, then the resulting defocus map supervises the fusion mask — that's a fitting loop, not an independent measurement, so the defocus map's accuracy is not independently validated.\n\nIs the paper worth a serious referee? Yes, I think so. The application is new, the formulation is coherent, and the limitations are acknowledged. But I'd send it back with a demand for the dataset and code, plus a quantitative sharpness check on the ultra-wide inputs and ideally some real sharp-reference evaluation rather than a synthesized one. The central idea is plausible; it's just not yet demonstrated as robustly as the abstract claims.","headline":"A genuine new application of dual-camera smartphone input for consistent-defocus NeRF, but the all-in-focus claim is conditional on the unmeasured sharpness of the ultra-wide camera, and the evaluation is too thin to fully support it.","tokens_in":22361,"tokens_out":1115,"would_cite":true,"duration_ms":12030,"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":"A smartphone's second camera yields all-in-focus 3D scenes from blurry video-like captures.","keywords":["neural radiance field","all-in-focus","dual-camera","defocus blur","novel view synthesis","bokeh rendering","smartphone photography","image alignment"],"falsifier":"Capture a scene where the ultra-wide camera itself shows visible defocus in a region where the main camera is also blurred, such as a low-light close-up that exceeds the ultra-wide's depth of field, and check whether DC-NeRF renders that region all-in-focus; if the region stays soft, the claim that the ultra-wide camera supplies the missing sharp reference is refuted.","tokens_in":21298,"feed_emoji":"📷","tokens_out":6260,"duration_ms":52906,"temperature":0.7,"pith_summary":"The paper claims to be the first to build an all-in-focus neural radiance field from inputs captured without manual refocusing. When a smartphone's main camera stays focused on one object in every view, the resulting defocus blur is consistent across views, so ordinary NeRF training has no sharp reference to recover the blurred regions. The paper's solution, DC-NeRF, adds the phone's ultra-wide camera, whose deeper depth of field keeps the whole scene sharper, and fuses it with the main camera through spatial alignment and defocus-aware blending. If the claim is right, smartphone owners can capture a scene once and later synthesize sharp novel views, refocus, or apply cinematic depth effects without any special focusing effort.","feed_headline":"A phone's second camera unlocks all-in-focus 3D scenes","feed_subtitle":"DC-NeRF blends a sharp ultra-wide lens with a blurry main camera to rebuild sharp new viewpoints.","key_machinery":"The load-bearing object is the defocus-aware fusion module, which turns dual-camera fusion into a learned blending problem. It renders a bokeh image from the aligned ultra-wide radiance field using a scatter operation whose circle-of-confusion radius is $r = a f |D_f - D|$, learns the blur intensity $A = af$ and focused disparity $D_f$ by matching the rendered bokeh color to the main-camera color with an SSIM loss, and computes a defocus map $D_{\\text{defocus}} = A|D - D_f|$ from the NeRF disparity. That map supervises an MLP blending mask $\\eta$ that weights the ultra-wide and main-camera fields during weighted volume rendering, so the sharp ultra-wide content replaces only the blurred main-camera regions. The alignment pipeline — SIFT-based homography, RAFT optical flow, a forward-backward consistency check, and histogram matching — makes the fusion meaningful across the two physically different cameras.","core_discovery":"On the paper's own terms, the central discovery is that view-consistent defocus blur — the case where every input view is blurred in the same places because the camera never refocuses — can be resolved by replacing the single-camera assumption of NeRF with a two-camera model. DC-NeRF takes a high-resolution main image and a lower-resolution ultra-wide image of the same view, aligns them with homography, optical flow, and histogram matching, then learns defocus parameters (blur intensity and focal disparity) by rendering bokeh from the ultra-wide field and matching it to the main-camera colors. A blending weight field predicted from the resulting defocus map fuses the sharp, deep-focus ultra-wide radiance field with the high-detail main-camera field inside volume rendering, producing all-in-focus novel views. The same learned parameters let the user refocus the rendered scene or apply split-diopter effects.","pith_inferences":["This suggests the same align-and-fuse design could be adapted to other paired-camera configurations, such as dual-pixel sensors or wide-plus-tele setups, whenever one camera's depth of field reliably exceeds the other's.","Because the fusion mask is supervised only by the defocus map, the method may implicitly learn a scene-depth prior that could be reused for editing focal effects or estimating disparity from novel viewpoints.","A natural stress test is a low-light scene where the ultra-wide image is noisy or itself blurred; the paper's own failure case already hints that such a boundary would break the assumption of a sharp ultra-wide reference."],"forward_implications":["Users can capture an all-in-focus, multi-view-consistent NeRF with a standard smartphone without manually refocusing the main camera between shots.","The learned defocus parameters enable post-capture refocusing, letting the focal plane be shifted to any disparity plane and the blur intensity be adjusted.","Split-diopter-style rendering, where foreground and background stay sharp while the middle region blurs, becomes a byproduct of the same representation.","Because DC-NeRF trains directly on the target scene, it avoids the extra defocus-deblurring dataset and pretraining that single-view dual-camera deblurring baselines require.","Existing NeRF-based deblurring methods that assume inconsistent blur across views fail on consistent blur, so the dual-camera fusion is a necessary ingredient rather than a modest improvement."],"supporting_citations":[{"why":"Supplies the base neural radiance field representation that DC-NeRF extends from one camera to two.","marker":"[4]"},{"why":"The main defocus-blur baseline; its sparse-kernel blur model assumes inconsistent focus, which DC-NeRF targets.","marker":"[5]"},{"why":"Depth-of-field baseline with controllable DoF rendering that also assumes inconsistent focusing planes.","marker":"[6]"},{"why":"Physical-scene-prior deblurring baseline that fails on view-consistent blur.","marker":"[7]"},{"why":"Source of the scatter-based bokeh rendering used to learn the defocus parameters.","marker":"[35]"},{"why":"Single-view dual-camera deblurring baseline that requires extra pretraining; DC-NeRF compares against it.","marker":"[51]"},{"why":"The RAFT optical flow model used for pixel-level alignment of the camera pair.","marker":"[54]"},{"why":"Multi-focus image fusion used to synthesize the all-in-focus ground truth from the two main-camera focal planes.","marker":"[55]"},{"why":"Structure-from-motion tool that estimates the camera intrinsics and extrinsics used during optimization.","marker":"[63]"}],"fun_headline_variants":["Dual-camera NeRF achieves all-in-focus 3D scenes","DC-NeRF: dual lenses remove defocus for sharp novel views","Two cameras better than one: all-in-focus NeRF","Smartphone dual cameras enable all-in-focus neural rendering","All-in-focus novel views from dual-camera NeRF"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ultra-wide camera's image is sharp in every region where the main camera is blurred; if the ultra-wide shot is itself blurry or too low quality in a region, the fusion has no sharp source to draw from and the all-in-focus result degrades.","fun_headline_variants_meta":{"raw":{"variants":["Dual-camera NeRF achieves all-in-focus 3D scenes","DC-NeRF: dual lenses remove defocus for sharp novel views","Two cameras better than one: all-in-focus NeRF","Smartphone dual cameras enable all-in-focus neural rendering","All-in-focus novel views from dual-camera NeRF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1384,"prompt_tokens":985,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":309}},"tokens_in":601,"tokens_out":399,"duration_ms":3670,"temperature":1.0,"reasoning_tokens":309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:58:31.094190+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Capture a scene where the ultra-wide camera itself shows visible defocus in a region where the main camera is also blurred, such as a low-light close-up that exceeds the ultra-wide's depth of field, and check whether DC-NeRF renders that region all-in-focus; if the region stays soft, the claim that the ultra-wide camera supplies the missing sharp reference is refuted.","supporting_citations":[{"cited_title":"Structure-from-motion revis- ited,","cited_arxiv_id":null,"evidence_quote":"Structure-from-motion tool that estimates the camera intrinsics and extrinsics used during optimization."},{"cited_title":"Nerf: Representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the base neural radiance field representation that DC-NeRF extends from one camera to two."},{"cited_title":"Deblur-nerf: Neural radiance fields from blurry images,","cited_arxiv_id":null,"evidence_quote":"The main defocus-blur baseline; its sparse-kernel blur model assumes inconsistent focus, which DC-NeRF targets."},{"cited_title":"Dof-nerf: Depth-of-field meets neural radiance fields,","cited_arxiv_id":null,"evidence_quote":"Depth-of-field baseline with controllable DoF rendering that also assumes inconsistent focusing planes."},{"cited_title":"Bokehme: When neural rendering meets classical rendering,","cited_arxiv_id":null,"evidence_quote":"Source of the scatter-based bokeh rendering used to learn the defocus parameters."},{"cited_title":"Point-and- shoot all-in-focus photo synthesis from smartphone camera pair,","cited_arxiv_id":null,"evidence_quote":"Single-view dual-camera deblurring baseline that requires extra pretraining; DC-NeRF compares against it."},{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow,","cited_arxiv_id":null,"evidence_quote":"The RAFT optical flow model used for pixel-level alignment of the camera pair."},{"cited_title":"Guided filter-based multi- focus image fusion through focus region detection,","cited_arxiv_id":null,"evidence_quote":"Multi-focus image fusion used to synthesize the all-in-focus ground truth from the two main-camera focal planes."}],"review_version":1}