{"id":"d9f7d2b4-3256-4c88-8496-dee803463678","arxiv_id":"2608.06240","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"PRISM is a flow-matching method for unpaired image translation that uses a learned per-feature gate, derived from target-distribution distance, to control both initialization and transport timing, improving realism and preservation.","lead":"PRISM is a new method for translating images between domains without paired examples, such as turning a cat photo into a dog photo or restaining tissue images. It learns a per-region gate that decides what to preserve and what to change, and reports the best FID/KID on four of five benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical claim rests on baselines that are not the published algorithms: SDEdit is re-run on the authors' own frozen flow, EGSDE is retrained far from its reported FID, and no code/splits are released; the AFHQ margin 76.9 vs 107.5 may reflect baseline weakness rather than the per-feature gate.","rationale":"Reader's verdict is CONDITIONAL and I agree with that conclusion, but for a different primary reason. The reader selected the missing distributional guarantee of Eq. 7 as the weakest assumption. I see that as a legitimate and honestly disclosed limitation, but the empirical FID/KID results already provide indirect evidence that the gated trajectories are usefully on-target; if they were badly off-manifold, PRISM would not beat baselines on FID. The more fragile load-bearing assumption is that the baselines are faithful strong controls. The paper is transparent about modifying SDEdit and restricting EGSDE, which is commendable, but transparency does not make the modified baseline a valid comparison for the central claim. A re-implementation of SDEdit on the authors' own frozen flow can be weaker than the original for reasons unrelated to the global-vs-per-feature control distinction; similarly, retraining EGSDE without its original backbone may destroy its guidance. The shared-split protocol ensures internal comparability among the particular checkpoints evaluated, but 'best among evaluated methods' is true by construction and does not transfer to 'better than global-control diffusion translators.' The proposed test is concrete and would settle the concern. If the faithful baselines still lose, the paper's empirical claim is much stronger and the conditional verdict could be upgraded. Until then, CONDITIONAL remains right, so the reader's verdict is unchanged.","tokens_in":22952,"tokens_out":14569,"duration_ms":169764,"concrete_test":"Release code, checkpoints, and split manifests, then run official SDEdit (or an equivalently capable pretrained denoiser) on the same AFHQ cat->dog and breast frozen->permanent splits, sweeping the global noise level and taking the best validation FID; independently rerun EGSDE with its original backbone and guidance weights on AFHQ. If faithful SDEdit's best FID is below PRISM's 76.9/51.8 or within the three-seed std, or if EGSDE's published configuration beats PRISM, the claimed superiority over global-control diffusion translators is unsupported. If PRISM still wins, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PRISM's central claim is comparative: being best among the evaluated methods is supposed to show that a distribution-derived per-feature gate beats global-control translation. For that conclusion, each baseline must be a strong, faithful instance of the method it stands for. This is the least secure condition. The SDEdit baseline is, by the paper's own admission (Section 4), 'an SDEdit-style application of our frozen Stage-1 flow rather than a reproduction of the original method.' Original SDEdit's behavior depends on the pretrained denoiser and on the single global noise level; replacing both with the authors' flow and a per-task corruption makes the baseline a re-implementation, not SDEdit. EGSDE is evaluated on only two of five benchmarks and, after retraining by the authors, gives FID 109.2 on AFHQ and 171.5 on breast, values far from its published operating point; retraining an energy-guided SDE without its original backbone and guidance schedule can arbitrarily weaken it. CycleGAN (115.0 on AFHQ) and CUT (162.3) are also outside the range usually reported on AFHQ. The shared same-split protocol makes the numbers internally comparable, but not externally decisive: 'best among evaluated methods' is guaranteed by the evaluation set. Without released code and split manifests, the reader cannot tell whether the margin to PRISM reflects the gate or the baseline implementations. The Section 5 limitation about missing distributional guarantee for Eq. 7 is real but secondary: if the gated ODE left the manifold badly, FID would presumably suffer; the more load-bearing risk is that the comparison itself is to weak controls.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"PRISM is a two-stage, GAN-free flow-matching framework for unpaired image-to-image translation. Stage 1 trains a frozen domain-conditional flow; Stage 2 trains a small gate predictor, a norm-constrained correction network, and an optional learned noise generator. A per-feature gate tau, supervised by a distribution-distance prior (DDtau) computed from DINOv2 feature distances to target statistics, controls both the ODE initialization (Eq. 4) and the per-feature wake-up timing during integration (Eq. 7). The t=0 corruption is task-matched, interpolating between content-anchored AdaIN corruption and isotropic noise. Experiments on five benchmarks report the best Inception FID and KID among the compared methods on four benchmarks, a competitive result on virtual staining, and a nuclei-count ratio closest to ideal on breast histopathology, together with ablations, a realism-faithfulness analysis, and inference-time local gate overrides.","tokens_in":23342,"tokens_out":9962,"duration_ms":107789,"significance":"If the comparative results hold up, PRISM is a meaningful contribution: it replaces a single global noise or guidance level with a learned per-feature preservation gate that is coupled to both the initialization and the transport timing, and it does so without adversarial training. The task-matched corruption idea is well motivated, the gate-granularity ablation is well designed, and the paper is unusually candid in its limitations section, including the lack of a formal distributional guarantee for Eq. 7 and the proxy nature of the pathology metrics. The shared same-split protocol, three-seed variance studies on two benchmarks, and bootstrap uncertainty estimates are strengths. However, the central comparative claim is only as strong as the baseline implementations, and the current protocol makes the baselines weaker and less faithful than the published methods they stand for.","major_comments":[{"comment":"The comparative conclusion 'best among evaluated methods' is under-determined by the baseline protocol. The SDEdit baseline is, in the paper's own words, 'an SDEdit-style application of our frozen Stage-1 flow rather than a reproduction of the original method', which substitutes the authors' flow for the pretrained denoiser and the per-task corruption for SDEdit's single global noise level; this is a new method, not SDEdit. EGSDE is evaluated on only two of the five benchmarks, and after retraining yields FID 109.2 on AFHQ and 171.5 on breast, values the paper acknowledges are far from its published operating point; CycleGAN and CUT also report FID values (115.0 and 162.3 on AFHQ, 74.6 on breast) outside the ranges usually associated with those methods. Because the headline claim is explicitly comparative, each baseline must be a faithful instance of the published algorithm on the same splits, or the claims must be restricted to 'our re-implementations' with the original code released. Please provide original-implementation comparisons or a sensitivity analysis showing the conclusions are robust to faithful baseline tuning, and report variance for all baselines.","section":"§4, 'Baselines and protocol'; Tables 5, 7, 8, 9"},{"comment":"The second contribution, 'the same gate controls both initialization and transport timing', is not tested by any ablation that isolates Eq. 7. Table 12 varies gate granularity and Table 13 varies the DDtau prior, corruption, joint training, and correction, but no experiment keeps Eq. 4 and disables the wake-up factor in Eq. 7, or applies a global timing schedule instead of the per-feature one. Since the paper's Limitations state that the gated dynamics 'do not inherit a formal distributional guarantee from the frozen flow', an ablation of the wake-up term is the minimal experiment needed to establish that per-feature timing, rather than the gated initialization and the correction network, is responsible for the reported FID/KID improvements.","section":"§3.3 (Eq. 7) and §4.10"},{"comment":"The paper states that 'the training source code, trained weights, split manifests, generated test outputs, and self-contained evaluation code can be released upon publication', but the evaluation protocol is highly custom: all baselines are retrained on split manifests that are not provided, EGSDE is retrained under a different schedule, and the FID/KID pipeline is unified. The 'best among evaluated methods' claim cannot be checked without these artifacts. Please release the code, weights, exact split definitions, and baseline configurations as part of the revision.","section":"Data availability; §4 protocol"}],"minor_comments":[{"comment":"At tk=1 the wake-up factor for tau=1 is sigma(0)=0.5, so a fully preserved feature still receives half the final velocity step; clarify whether this endpoint behavior is intended and whether the last step should be handled separately.","section":"§3.3, Eq. (7)"},{"comment":"The norm-clipping procedure says the clip is applied 'first using the channelwise norm at each spatial location and then using the norm over the full tensor'; state precisely which norm is used and in which order.","section":"§3.5, Eq. (9)"},{"comment":"The LNG experiment reports FID 75.0 versus 75.3 but no table or run configuration; add a table with the seed and training budget.","section":"§4.2"},{"comment":"Removing the DDtau prior leaves FID essentially unchanged (103.5 vs 103.6) while changing NPS; the text should discuss explicitly why the realism metric does not move.","section":"§4.10, Table 13"},{"comment":"The qualifier 'among the evaluated methods' is used inconsistently; ensure it appears in every sentence that claims the best FID/KID, since the protocol is not directly comparable to published numbers.","section":"Abstract and §4"},{"comment":"The phrase 'can be released upon publication' should be a firm commitment to release code, weights, and split manifests with links.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising but the comparative protocol is the main risk. I would push the authors either to compare with original implementations on the same splits or to explicitly frame the numbers as re-implementation-relative, and to release all artifacts. The missing Eq. 7 ablation is the most important technical gap; without it, the contribution of the transport-timing mechanism is not established. The self-citation pattern is not problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the mechanism is the interesting part. PRISM's per-feature gate, derived from a distribution-distance prior and applied to both initialization and transport timing, is not in the cited literature. That coupling is the real contribution, and the paper deserves credit for a clear architecture: frozen flow backbone, zero-initialized gate predictor, norm-bounded correction, and an honest set of ablations.\n\nWhat it does well: the gate-granularity experiment (global vs. spatial vs. per-feature) supports the central design claim internally. The component ablations are careful and the limitations section is straight about what is and is not guaranteed. The inference-time override from text or a detector is a useful practical extension. The authors are not overselling within their own controlled protocol.\n\nWhere it gets soft: the comparison to baselines is the load-bearing weakness. The paper says plainly that the SDEdit baseline is \"an SDEdit-style application of our frozen Stage-1 flow rather than a reproduction of the original method,\" so those numbers do not represent SDEdit. EGSDE is evaluated on only two benchmarks and, after retraining, lands far from its published operating point (109.2 AFHQ, 171.5 breast). CycleGAN and CUT on AFHQ also look weak at 115 and 162. That makes \"best among evaluated methods\" a statement about this evaluation set, not about the field. The unified protocol makes the numbers internally comparable but not externally decisive, and without released code and split manifests no outside reader can tell whether the margin comes from the gate or from weak controls. Add the single-run estimates on several benchmarks, and the empirical case is thinner than the tables suggest.\n\nThe admitted lack of a formal distributional guarantee for Eq. 7 is a secondary concern; the FID values on real outputs suggest the trajectory mostly stays on manifold, so I would not treat that as the main risk.\n\nWho this is for: researchers working on controllable unpaired translation, especially those interested in per-feature control and medical image harmonization. The mechanism is worth knowing about. It should go to peer review, but the bar for acceptance should include code release, a faithful original-SDEdit baseline, a stronger or published-checkpoint EGSDE comparison, and error bars on all benchmarks. As it stands, it is a promising core idea with a comparison that needs substantial hardening.","headline":"A genuinely new per-feature gate mechanism, but the baseline comparisons are too weak to support the headline FID claims.","tokens_in":23895,"tokens_out":2549,"would_cite":false,"duration_ms":29678,"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":"PRISM replaces the single global noise knob of diffusion-based unpaired translation with a learned per-feature gate, and claims this gate, applied to both initialization and transport timing, improves realism without sacrificing structure.","keywords":["unpaired image-to-image translation","flow matching","controllable generation","per-feature transport","distribution distance","structure-preserving translation","medical image harmonization"],"falsifier":"Retrain PRISM on breast frozen-to-permanent translation with the gate frozen to a constant mid-value in both Equation (4) and Equation (7), keeping all losses and budget fixed; if FID and the nuclei-count ratio stay at the reported level (51.8 and 0.93), the distribution-derived spatial pattern is not what drives the balance. On the released test outputs, also check that per-feature displacement decreases monotonically with predicted $\\tau$: a large fraction of high-$\\tau$ features with large displacement would contradict the claimed selective transport.","tokens_in":22705,"feed_emoji":"🎨","tokens_out":13305,"duration_ms":122653,"temperature":0.7,"pith_summary":"The paper argues that the central difficulty in unpaired image-to-image translation is deciding, per image and per region, what to change and what to preserve, and that existing diffusion-based translators fail because they control this with a single global noise or guidance value. It proposes PRISM, a GAN-free method built on flow matching (a generative model that learns straight-line paths from noise to data), whose learned per-feature gate $\\tau\\in[0.05,1]$ is derived from each latent feature's standardized distance to the target distribution, so features already close to the target are preserved and distant features are freed. The same gate sets the ODE initial mixture $z(0)=\\tau\\odot z_A+(1-\\tau)\\odot\\varepsilon_\\alpha$ and the per-feature wake-up timing of transport. On five benchmarks spanning structure-preserving and structure-changing tasks, PRISM reports the best Inception FID and KID (Fréchet Inception Distance and Kernel Inception Distance) among the compared methods on four, a competitive result on the fifth, and a histopathology nuclei-count ratio closest to ideal, supporting the claimed realism-preservation balance.","feed_headline":"Per-feature gate wins image translation on 4 of 5 benchmarks","feed_subtitle":"Flow-matching model decides, per region, what to keep and what to redraw, beating global-noise baselines.","key_machinery":"The load-bearing object is the per-feature preservation gate $\\tau\\in[0.05,1]$, one scalar per latent location, whose spatial pattern is supervised by DD$\\tau$: each patch's standardized distance to the target feature distribution, quantile-normalized so distant patches receive low $\\tau$ (freed) and target-like patches high $\\tau$ (preserved). The gate carries the argument by being applied at both coupling points of the frozen-flow sampler: Equation (4) initializes each feature as $\\tau\\odot z_A+(1-\\tau)\\odot\\varepsilon_\\alpha$, mixing the real source latent with a task-matched corruption (AdaIN content anchor for structure-preserving tasks, partial isotropic anchor for structure-changing ones), and Equation (7) gates the composite velocity by the soft per-feature wake-up $\\sigma((t_k-\\tau)/T)$, so each feature only moves after its own threshold time. Both uses are smooth functions of $\\tau$, making the gate trainable end to end against GAN-free realism objectives and a $\\tau$-gated structure loss, and overridable at inference from a text prompt or a detector map without retraining.","core_discovery":"The paper's central claim is that unpaired translation can be recast as selective per-feature transport to the target distribution, controlled by a distribution-informed gate $\\tau\\in[\\tau_{\\min},\\tau_{\\max}]$ with the same shape as the latent. The gate's spatial prior comes from DD$\\tau$, the standardized distance of each source patch to the target feature distribution under a frozen DINOv2 encoder, so target-consistent patches receive high $\\tau$ and distant patches low $\\tau$. The same gate enters the sampler twice: the initialization $z(0)=\\tau\\odot z_A+(1-\\tau)\\odot\\varepsilon_\\alpha$ mixes the real source latent with a task-matched corruption (AdaIN-style content anchoring for structure-preserving tasks, a partially anchored isotropic blend at $\\alpha=0.5$ for structure-changing ones), and the integration applies a soft per-feature wake-up $\\sigma((t_k-\\tau)/T)$ so high-$\\tau$ features stay source-anchored and low-$\\tau$ features translate earlier. Over a frozen flow-matching backbone with a norm-constrained residual correction, PRISM reports the best Inception FID and KID on AFHQ cat-to-dog (76.9), CelebA-HQ man-to-woman (90.6), breast frozen-to-permanent (51.8) and day-to-night (85.9), competitive results on virtual staining (50.8 versus 49.2), and the nuclei-count ratio closest to 1.0 (0.93) on histopathology, arguing this reflects selective transport rather than under-translation.","pith_inferences":["The same checkpoint could expose the anchor strength $\\alpha$ as an interactive realism-faithfulness dial at inference; the paper reports the frontier in its ablation tables but does not claim the user can change it per image after training.","If the wake-up schedule does what the paper argues, preserved regions should stay nearly unchanged when the number of ODE steps is increased far beyond the default of 16, because high-$\\tau$ features never activate; this is a checkable consequence the paper does not state.","The DD$\\tau$ prior is computed from a frozen encoder and target statistics only, so it could act as a plug-in preservation map for other latent-space editors beyond PRISM's own sampler, as long as those editors accept a per-feature start-point mixture."],"forward_implications":["On four of the five benchmarks the gated sampler achieves the best Inception FID and KID among the compared methods, with three-seed training variability ($77.4\\pm0.7$ on AFHQ) well below the gap to the next-best method.","A single trained checkpoint can be steered locally at inference, for example by a text prompt such as 'preserve eyes' or by a nuclei-detector map, trading source preservation against target realism without any retraining.","Task-matched corruption avoids resynthesizing shared geometry: content-anchored initialization keeps structure while the gate transports only appearance, and on histopathology the generated-to-source nuclei-count ratio is closest to the ideal 1.0 among compared methods (0.93 versus, for example, EGSDE's 3.32).","The same gate field implies a per-feature compute budget: idealized active-step counts drop by 82.4% on the preservation-dominant breast task, though realizing this as wall-clock savings requires sparse execution that the paper leaves to future work."],"supporting_citations":[{"why":"Supplies the flow-matching objective and the frozen straight-path generative flow that Stage 1 trains and Stage 2 reuses.","marker":"[20]"},{"why":"Defines the SDEdit baseline whose single global noise knob Equation (4) generalizes per feature.","marker":"[5]"},{"why":"CycleGAN is the principal adversarial baseline that PRISM reports beating on every benchmark.","marker":"[2]"},{"why":"CUT is the contrastive baseline compared on all five datasets.","marker":"[3]"},{"why":"UNSB is the Schrödinger-bridge diffusion baseline compared on all five datasets.","marker":"[17]"},{"why":"EGSDE is the energy-guided SDE baseline compared on AFHQ and frozen-to-permanent translation.","marker":"[15]"},{"why":"AdaIN provides the content-anchored corruption fixed point used at t=0 for structure-preserving tasks.","marker":"[41]"},{"why":"DINOv2 is the frozen encoder behind both the DDτ distribution-distance prior and the k-NN local realism matcher.","marker":"[43]"}],"fun_headline_variants":["Per-feature gate beats global-noise baselines on 4/5 translation tasks","Flow-matching gate preserves what matters, redraws the rest","Distribution-gated flow matching wins 4/5 unpaired translation benchmarks","One gate per feature: better unpaired translation on 4/5"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that multiplying the flow's velocity by a per-feature wake-up sigmoid still keeps the trajectory on the target-image manifold, even though the gated dynamics deviate from the frozen flow's training distribution and the paper states they carry no formal distributional guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Per-feature gate beats global-noise baselines on 4/5 translation tasks","Flow-matching gate preserves what matters, redraws the rest","Distribution-gated flow matching wins 4/5 unpaired translation benchmarks","One gate per feature: better unpaired translation on 4/5"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2875,"prompt_tokens":1163,"completion_tokens":1712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":779,"completion_tokens_details":{"reasoning_tokens":1632}},"tokens_in":779,"tokens_out":1712,"duration_ms":14323,"temperature":1.0,"reasoning_tokens":1632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:38:18.313769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain PRISM on breast frozen-to-permanent translation with the gate frozen to a constant mid-value in both Equation (4) and Equation (7), keeping all losses and budget fixed; if FID and the nuclei-count ratio stay at the reported level (51.8 and 0.93), the distribution-derived spatial pattern is not what drives the balance. On the released test outputs, also check that per-feature displacement decreases monotonically with predicted $\\tau$: a large fraction of high-$\\tau$ features with large displacement would contradict the claimed selective transport.","supporting_citations":[{"cited_title":"Lipman, R","cited_arxiv_id":null,"evidence_quote":"Supplies the flow-matching objective and the frozen straight-path generative flow that Stage 1 trains and Stage 2 reuses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the SDEdit baseline whose single global noise knob Equation (4) generalizes per feature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CycleGAN is the principal adversarial baseline that PRISM reports beating on every benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CUT is the contrastive baseline compared on all five datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UNSB is the Schrödinger-bridge diffusion baseline compared on all five datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EGSDE is the energy-guided SDE baseline compared on AFHQ and frozen-to-permanent translation."},{"cited_title":"Huang, S","cited_arxiv_id":null,"evidence_quote":"AdaIN provides the content-anchored corruption fixed point used at t=0 for structure-preserving tasks."},{"cited_title":"Oquab, T","cited_arxiv_id":null,"evidence_quote":"DINOv2 is the frozen encoder behind both the DDτ distribution-distance prior and the k-NN local realism matcher."}],"review_version":1}