{"id":"002757eb-7c6b-47ec-b1df-7f80e88394bc","arxiv_id":"2504.14294","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"ConFill modifies diffusion sampling with a context-adaptive Wasserstein-type discrepancy and complexity-based sampling, reporting improved FID, LPIPS, and SSIM on image inpainting benchmarks.","lead":"A diffusion-based image inpainting method, ConFill, adds a context-adaptive discrepancy metric and dynamic sampling to pre-trained diffusion models. It reports consistent quality gains over RePaint, CoPaint, DPS, and other baselines on CelebA-HQ, Places2, and ImageNet-1K.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (14) calibrates CAD guidance using the target/final image and treats a mean CAD as a variance, so the sampling loop is not a well-defined posterior sampler and the reported gains are currently untraceable.","rationale":"The reader's weakest_assumption identifies the same broad area — the CAD posterior approximation in Eqs. (10)-(14) — but my concern is more specific. Eq. (14) is not a variance in the statistical sense: it averages CAD values rather than squared deviations, and it references hat_x_0, which is not available at the time the variance would be used. This makes the guidance scale either circular or a non-adaptive global constant, and Algorithm 2's update does not match Eq. (9). These are internal inconsistencies, not merely disagreement with consensus. The concern is load-bearing because without a valid, implementable calibration the claimed gains cannot be attributed to CAD. I do not think this forces a REJECT: the empirical results could still hold as a heuristic, and the authors could resolve the issue with a corrected derivation, explicit inference-time definitions, and released code. That is exactly the CONDITIONAL posture the reader took, so the verdict is unchanged. Agreement is partial because the reader's phrasing focused on a single variance calibrating all regions, whereas the sharper problem is that the quantity is not computable at inference and is not a variance at all.","tokens_in":23518,"tokens_out":9760,"duration_ms":92287,"concrete_test":"Independently reimplement ConFill from the paper text using a public RePaint/DDPM checkpoint on CelebA-HQ with the reported mask types. At timestep T, compute gamma'^2_T from Eq. (14) in two ways: (a) with the ground-truth x_0 (oracle), and (b) with only r_0 and the one-step estimate g_theta^(T)(hat_x_T), the only quantities actually available at inference. If (b) is undefined because hat_x_0 is not yet known, or if the two gamma'^2_T values differ by more than roughly 30%, then Eq. (14) does not define an implementable calibration and the Table 1 improvements cannot be attributed to CAD. As a second check, run both Eq. (9)'s gradient descent and Algorithm 2's update on a few timesteps and verify whether they minimize the same objective; if they diverge, the method is described by two different algorithms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CAD guidance plus dynamic sampling yields SOTA inpainting, but the derivation connecting CAD to the sampling loop has a concrete defect. Eq. (10) replaces the observation likelihood with a delta, then Eqs. (11)-(14) approximate it by a Gaussian with variance gamma'^2. Eq. (14) sets gamma'^2_T = (1/N) sum_i CAD(s(g_theta^(T)(hat_x_T^(i))), s(hat_x_0^(i))). At timestep T, the completed image hat_x_0 — or the ground-truth x_0, if that is what is meant — is not available; the text says 'N represents r0's dimension' and compares transformed versions of hat_x_T and hat_x_0. If the calibration uses the target image itself, the method is circular; if it uses a training-set average, it is a single global scalar and cannot adapt per region and timestep as claimed. Moreover, a likelihood variance should be a mean squared residual, not a mean CAD value; using mean CAD as gamma'^2 mis-scales the guidance. Algorithm 2 then updates hat_x_{t-1} = g_theta^(t)(hat_x_t) - lambda grad phi_t(hat_x_t), which is not the gradient step on Eq. (9)'s L_CAD, so the implemented loop is not a posterior sampler for the stated model. With no code released, the reported LPIPS/FID gains cannot currently be traced to the CAD mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ConFill, an unsupervised image completion framework built on a pre-trained diffusion model. The method introduces a Context-Adaptive Discrepancy (CAD) model that integrates local context into an optimal-transport-based discrepancy measure, which is used to guide the reverse diffusion process toward satisfying the completion constraint. A dynamic sampling scheme is also proposed to increase sampling density in regions with high textural or structural complexity. The authors report state-of-the-art performance on CelebA-HQ, Places2, and ImageNet-1K across several mask types, including LPIPS improvements of 10.9% over RePaint and 6.7% over CoPaint, better FID on all three datasets, and higher human preference in a user study.","tokens_in":23831,"tokens_out":7134,"duration_ms":59854,"significance":"If the method worked as described, it would be a meaningful advance in training-free image completion: a context-adaptive discrepancy metric that replaces hand-tuned guidance in diffusion-based inversion, with empirical gains over strong baselines. The experimental section is broad, covering many baselines, mask types, ablations, and a human study, and the authors report multiple metrics. However, the theoretical derivation of the CAD-guided posterior is the core of the contribution, and it contains serious formal flaws that undermine the link between the CAD mechanism and the reported results. The paper does not release code, and the described algorithm is not precisely specified, so the empirical claims cannot currently be traced to the proposed mechanism. As presented, the central claim is therefore not established.","major_comments":[{"comment":"The variance γ'^2_T is calibrated using the same CAD values and the same target r0 that the completion constraint is meant to enforce. At inference time the full target is unavailable, and if the calibration uses the masked known region alone, a single global scalar cannot adapt per region and timestep as claimed. Moreover, a likelihood variance should be a mean squared residual, not a mean CAD value; using mean CAD as γ'^2 mis-scales the guidance. The sampling loop is therefore not a valid posterior sampler under the stated model, and the reported gains cannot be traced to CAD.","section":"Sec. 3.2, Eq. (14)"},{"comment":"The second term applies CAD to a single argument, 'CAD( r0 − s(g_theta^(t−1)(hat_x_{t−1})) )^2', but CAD is defined in Eq. (16) as a sum over pointwise discrepancies between two distributions. The difference of two images is not a distribution, so this expression is undefined. The derivation of Eq. (19) is therefore incomplete and cannot support the claimed gradual approximation.","section":"Sec. 3.4, Eq. (19)"},{"comment":"The update hat_x_{t−1} = g_theta^(t)(hat_x_t) − λ∇φ_t(hat_x_t) is not the gradient step on L_CAD as defined in Eq. (9), which would be hat_x_{t−1} = hat_x_t − λ∇_{hat_x_t} L_CAD. Additionally, the loss described in line 6 (minimizing 'CAD(hat_x_t, r0) + λ_t ||∇φ_t(hat_x_t) − hat_x_t||') does not match Eq. (9). This mismatch makes it impossible to identify which objective is actually optimized in the experiments, so the connection between the posterior derivation and the reported performance is broken.","section":"Algorithm 2, lines 6 and 9"},{"comment":"The replacement of the observation likelihood with a delta function and the subsequent Gaussian approximation are asserted without a limiting argument. The statement following Eq. (11) that the approximation becomes increasingly accurate as γ_T → 0 is in tension with Eq. (14), which sets γ' to be 'sufficiently large'. This inconsistency needs to be resolved for the posterior derivation to be coherent.","section":"Sec. 3.2, Eqs. (10)–(13)"},{"comment":"The hyperparameters are selected on the first five validation images, and Eq. (14) is evaluated per-sample using r0. It is unclear whether r0 in Eq. (14) is the full ground-truth image or the observed (masked) region only. If the former, the method leaks test-time information; if the latter, the global scalar γ' cannot provide the per-region adaptation claimed. Please clarify the exact input used in the calibration.","section":"Sec. 4.1 and Eq. (14)"}],"minor_comments":[{"comment":"There are numerous typos and inconsistent notations, including 'PowePaint' (Sec. 4.4), 'COPaint' (Sec. 4.4), 'p′θ(ˆxT|OT)' (Sec. 3.4), and the overloaded use of N in Eqs. (14) and (15) (as dimension, sample count, and pixel count).","section":"Throughout"},{"comment":"The definition of L_diffusion is not used elsewhere in the paper; consider removing it or explicitly connecting it to the CAD loss.","section":"Sec. 3.2, Eq. (4)"},{"comment":"The transport map f computed in Algorithm 1 is not referenced in the update equations of Algorithm 2, so its role in the overall method is unclear.","section":"Algorithm 1"},{"comment":"The figure is not referenced in the text, and the curves are not clearly labeled; the caption says 'mean pixel-wise discrepancy' but it is unclear whether this is computed for CAD, Wasserstein, or Euclidean measures, and over which images.","section":"Fig. 2"},{"comment":"The paper states that J=1 and ´t=10 are used in the experiments, but Table 4 does not show a row for J=1 and ´t=15 or other combinations; please report the full grid used for selection.","section":"Sec. 4.6, Table 4"},{"comment":"The phrase 'setting a new benchmark' overstates the contribution of a single paper; a more cautious wording, such as 'state-of-the-art results on the evaluated datasets', would be more appropriate.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The paper's theoretical core—the CAD-guided posterior sampling—has load-bearing errors that make the method ill-defined and the empirical results untraceable. The calibration in Eq. (14) is circular if it uses the full target, and statistically incorrect if it uses a mean CAD as a variance. Eq. (19) is undefined, and Algorithm 2 does not implement the stated objective. These are not presentation issues but fundamental flaws that would require a substantial re-derivation and likely a rethinking of the method. Even with the extensive experiments, I cannot see how the current manuscript could be revised within the scope of a normal revision to be made sound. I recommend rejection, with the possibility of considering a future version that provides a rigorous derivation and released code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely new discrepancy term and a lot of solid empirical work, but the derivation tying that term to the sampling loop is circular enough that the headline gains are not yet traceable to the mechanism. I'd still send it to review, because the experiments are substantial and the issue is fixable.\n\nWhat's actually new: the CAD term in Eqs. (6)-(7), which combines a Brenier-potential transport map with a context-sensitive exponential weighting, does not appear in the cited prior work, including DPM-OT. That's a reasonable extension, and the dynamic sampling strategy, while not novel on its own, is a sensible complement. The paper does many things right: three datasets, seven baselines, a 46-participant user study, ablations on the main components, and a test on unseen mask types. The reported gains are consistent and sometimes large—10.9% LPIPS improvement over RePaint on ImageNet, for instance. If the method works as described, it's a useful contribution to a busy subfield.\n\nThe soft spots are real. First, the posterior derivation in Sec. 3.2 is asserted rather than derived. Eq. (10) uses a delta for the observation likelihood, and Eqs. (11)-(14) approximate it with a Gaussian. Eq. (14) estimates the variance gamma'^2_T as the mean CAD between predicted and target images at timestep T. That is circular if the target image is used—it's the thing you're trying to infer—and even as a training-set average it's mis-scaled: a likelihood variance should be a mean squared residual, not a mean CAD value. Second, Algorithm 2's update x_{t-1} = g_theta^{(t)}(x_t) - lambda grad phi_t(x_t) is not a gradient step on the L_CAD defined in Eq. (9), so the implemented loop is not sampling from the stated posterior. Third, Eq. (19) feeds a difference vector r0 - s(...) into CAD, which Eq. (16) never defines. These are fixable in revision, but they're load-bearing because the paper's whole claim is that CAD drives the gains. Also missing: no code, no error bars, and the hyperparameters psi and m from the dynamic sampling section are never given concrete values.\n\nBottom line: the empirical work deserves a serious referee, but the method is under-specified as written. If the authors release code or tighten the derivation, this could be a solid paper. For now, I'd treat the numbers as provisional.","headline":"A plausible new discrepancy term with substantial experiments, but the posterior derivation is circular enough that the headline gains are not yet traceable to the mechanism.","tokens_in":24417,"tokens_out":3847,"would_cite":false,"duration_ms":29446,"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":"ConFill claims a context-adaptive discrepancy term, inserted into a fixed pre-trained diffusion model, produces more coherent image completions than existing inpainting methods.","keywords":["image completion","image inpainting","diffusion models","context-adaptive discrepancy","optimal transport","dynamic sampling","generative inverse problems","unsupervised inpainting"],"falsifier":"A direct check is to replace Eq. (14)'s variance with a constant and re-run the ImageNet experiment; if the LPIPS gain over RePaint shrinks by less than half, the adaptive calibration is not responsible for the reported improvement.","tokens_in":23259,"feed_emoji":"🧩","tokens_out":6871,"duration_ms":55206,"temperature":0.7,"pith_summary":"ConFill is an image-completion framework that leaves the diffusion model itself untouched and instead changes how the denoising loop measures and enforces agreement with the known pixels. Its central claim is that existing diffusion inpainting methods fail because their discrepancy metrics, Euclidean distance or plain Wasserstein distance, ignore local context, so the evolving distribution of the missing region never aligns with the surrounding image. The paper replaces those metrics with a Context-Adaptive Discrepancy (CAD) that scales an optimal-transport cost by local edge, texture, and semantic differences, and adds a Dynamic Sampling rule that allocates more denoising steps to complex regions. If the claim is right, a pre-trained diffusion model becomes the best tested unsupervised image completer: on ImageNet-1K the reported LPIPS is 10.9% lower than RePaint and 6.7% lower than CoPaint, with better FID on all three datasets.","feed_headline":"Context-adaptive diffusion beats the leading image-completion baselines","feed_subtitle":"A fixed pre-trained diffusion model, guided by context-aware discrepancy and dynamic sampling, becomes the best tested unsupervised…","key_machinery":"The load-bearing object is the Context-Adaptive Discrepancy, $W_C(x,y;C)=\\|\\nabla\\varphi(x)-y\\|^2 \\cdot f(C(x),C(y))$, where $\\nabla\\varphi$ is the gradient of a Brenier potential giving the optimal transport map between latent distributions, and $f(C(x),C(y))=1+\\upsilon e^{-\\tau\\|C(x)-C(y)\\|}$ is a scaling factor that up-weights the cost when local contextual features differ. The contextual features $C(\\cdot)$ combine handcrafted edge and texture filters with a pre-trained semantic feature extractor. This term does two jobs: it replaces the Euclidean discrepancy in the diffusion loss, and it defines the completion constraint in the approximate posterior of Eq. (13). The supporting mechanism is Dynamic Sampling, whose adaptability function $\\varrho(x)=\\hat{\\alpha}\\exp(-\\psi\\,\\mathrm{Var}(x)N/(N-1))+\\hat{\\beta}\\,\\mathrm{ED}(x)$ sets the number of samples per region, concentrating computational effort where textural and structural complexity is highest.","core_discovery":"ConFill's core discovery is that the completion constraint can be pushed into the very first latent $\\hat{x}_T$ instead of being enforced by replacement or blending at every step. The posterior $p_\\theta(\\hat{x}_T \\mid O)$ is written as the prior times a CAD-aware delta function (Eq. 10), and its log-density is approximated by a Gaussian whose variance $\\gamma'^2_T$ is estimated from the CAD values of the target image (Eqs. 13-14). During denoising, each step minimizes a loss that combines the standard transition term with $\\gamma_t \\cdot W_C(s(g_\\theta(x_t)), r_0; C)$, where $W_C$ is the context-scaled optimal-transport discrepancy of Eqs. (6)-(7). A Dynamic Sampling schedule then raises the sample density in high-variance, high-edge-density regions. Taken together, the method progressively reduces the discrepancy between known and unknown regions at every timestep, which the paper argues is what produces the reported gains in fidelity and coherence.","pith_inferences":["An implication the paper leaves implicit is that the variance calibration in Eq. (14) is estimated from the same CAD values the method is trying to minimize; a natural test is to replace $\\gamma'^2_T$ with a fixed constant and measure whether the LPIPS gains survive, which would isolate whether the adaptive scaling or the CAD term itself is doing the work.","Because the CAD formulation only assumes a differentiable generator and a known-region operator $s(\\cdot)$, the same guidance could in principle be applied to other inverse problems such as super-resolution or deblurring; that extension is not tested in the paper.","The per-mask tables show the largest gains on the hardest masks (Expand, Half); on Narrow masks the gap to RePaint and CoPaint shrinks considerably, suggesting ConFill's advantage is concentrated where the completion problem is underdetermined."],"forward_implications":["With ConFill, a pre-trained diffusion model can be converted into a strong image completer without retraining on masks; the paper reports best average FID, LPIPS, and SSIM over all tested baselines on CelebA-HQ, Places2, and ImageNet-1K.","The method generalizes to mask types never seen during tuning: on image-expansion and nearest-neighbour masks, ConFill achieves FID 19.63 versus 23.87 for CoPaint, a 17.8% improvement.","Dynamic sampling makes the framework faster than the resampling baselines: about 109 seconds per image on unseen masks versus roughly 211 for RePaint and 453 for DPS.","Human preference tracks the metric gains: ConFill was 21% more likely to be selected over CoPaint on Places2 at mask ratio 0.8-1.0 and 29% more likely over RePaint on ImageNet at mask ratio 0.4-0.6."],"supporting_citations":[{"why":"Supplies the resampling baseline method that ConFill is compared against and whose time-travel mechanism ConFill adapts.","marker":"[17]"},{"why":"Supplies the CoPaint baseline, the strongest unsupervised DDIM-based competitor, against which ConFill reports a 6.7% LPIPS improvement on ImageNet.","marker":"[27]"},{"why":"Supplies DPS, a Bayesian posterior-sampling baseline that ConFill's CAD-guided posterior approximation extends.","marker":"[25]"},{"why":"Supplies the pre-trained DDPM forward-reverse process that ConFill uses as the base generative model.","marker":"[13]"},{"why":"Supplies the DDIM deterministic sampling framework that ConFill builds on for efficient reverse sampling with time travel.","marker":"[14]"},{"why":"Supplies the Brenier-potential optimal transport idea that CAD adapts to be locally context-aware.","marker":"[23]"},{"why":"Supplies the posterior-sampling view of diffusion restoration that motivates the posterior approximation in Eqs. (10)-(14).","marker":"[55]"},{"why":"Supplies the null-space blending strategy, an alternative completion constraint that CAD is designed to replace.","marker":"[15]"}],"fun_headline_variants":["ConFill: Context-adaptive diffusion for seamless image completion","Context-adaptive discrepancy drives diffusion image completion to SOTA","Adaptive sampling and context alignment sharpen diffusion inpainting","ConFill's CAD and dynamic sampling refine image completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that one number, a variance estimated from how well the model currently matches the known pixels, can correctly calibrate the completion constraint for every part of the image at every denoising step; if this single calibration is wrong, the guidance is miscalibrated and the reported gains do not follow.","fun_headline_variants_meta":{"raw":{"variants":["ConFill: Context-adaptive diffusion for seamless image completion","Context-adaptive discrepancy drives diffusion image completion to SOTA","Adaptive sampling and context alignment sharpen diffusion inpainting","ConFill's CAD and dynamic sampling refine image completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1496,"prompt_tokens":960,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":470}},"tokens_in":576,"tokens_out":536,"duration_ms":4998,"temperature":1.0,"reasoning_tokens":470,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:52:15.393076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check is to replace Eq. (14)'s variance with a constant and re-run the ImageNet experiment; if the LPIPS gain over RePaint shrinks by less than half, the adaptive calibration is not responsible for the reported improvement.","supporting_citations":[{"cited_title":"Repaint: Inpainting using denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the resampling baseline method that ConFill is compared against and whose time-travel mechanism ConFill adapts."},{"cited_title":"Towards coherent image inpainting using denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Supplies the CoPaint baseline, the strongest unsupervised DDIM-based competitor, against which ConFill reports a 6.7% LPIPS improvement on ImageNet."},{"cited_title":"Diffusion posterior sampling for general noisy inverse problems,","cited_arxiv_id":null,"evidence_quote":"Supplies DPS, a Bayesian posterior-sampling baseline that ConFill's CAD-guided posterior approximation extends."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained DDPM forward-reverse process that ConFill uses as the base generative model."},{"cited_title":"Denoising diffusion implicit models,","cited_arxiv_id":null,"evidence_quote":"Supplies the DDIM deterministic sampling framework that ConFill builds on for efficient reverse sampling with time travel."},{"cited_title":"Dpm-ot: a new diffusion probabilistic model based on optimal transport,","cited_arxiv_id":null,"evidence_quote":"Supplies the Brenier-potential optimal transport idea that CAD adapts to be locally context-aware."},{"cited_title":"Denoising diffusion restoration models,","cited_arxiv_id":null,"evidence_quote":"Supplies the posterior-sampling view of diffusion restoration that motivates the posterior approximation in Eqs. (10)-(14)."},{"cited_title":"Zero-shot image restoration using denoising diffusion null-space,","cited_arxiv_id":null,"evidence_quote":"Supplies the null-space blending strategy, an alternative completion constraint that CAD is designed to replace."}],"review_version":1}