{"id":"7cf2af5d-f9a9-470c-a6a1-eeab108fe8f1","arxiv_id":"2608.07176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"REVEAL, a flow-matching generative model pretrained on 4.8 million endoscopic frames with domain-adapted representation alignment, produces high-fidelity endoscopy images and features that beat EndoViT and Endo-FM on classification benchmarks.","lead":"A new AI model called REVEAL, trained on nearly five million endoscopy images from eight Dutch hospitals, generates realistic gastrointestinal images and, without any label training, matches or beats specialized endoscopy AI models at detecting disease.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No patient-level deduplication is reported between GN-5M pretraining and the BE/POLAR evaluation sets; if test-patient frames appear in pretraining, the reported superiority over EndoViT/Endo-FM and the FID gains reflect memorization, not generalization.","rationale":"The reader's weakest assumption identified in-domain teacher memorization as a key risk; my concern is a concrete and more severe mechanism for that memorization, namely patient-level overlap between GN-5M pretraining and the BE/POLAR test cohorts. If the overlap check comes back empty, the reader's original conditionality still applies, including the unspecified FID reference, missing error bars, and the self-contradictory linear-probing description, but the central comparative claim survives. If overlap exists, the claims of exceeding EndoViT and Endo-FM and of large FID improvements from alignment are not evidence of generalization. I keep the verdict conditional rather than rejecting outright because the required evidence is obtainable from the dataset owners and the paper is otherwise careful and well-ablated. The specific, falsifiable check is the patient-disjoint re-evaluation described above.","tokens_in":12226,"tokens_out":11063,"duration_ms":98319,"concrete_test":"Request a patient-level intersection analysis: for every patient in the POLAR and BE test sets, determine whether any GN-5M frame originates from the same patient or the same endoscopic procedure, using hospital, procedure ID, and timestamps, or a de-identified overlap check from the dataset owners. Then re-run the Table 2 linear-probing protocol and the Table 1 FID computation after removing all GN-5M frames from overlapping patients. If the REVEAL-versus-EndoViT/Endo-FM gaps and the alignment-versus-no-alignment FID gains largely disappear after this exclusion, the reported advantages are leakage artifacts rather than representation quality. If patient IDs are unavailable, a fallback is to retrain the GN-5M-adapted encoders on all non-Amsterdam hospitals and test on the Amsterdam UMC BE cohort, approximating a patient-disjoint external evaluation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central comparative claims in Tables 2 and 3 (REVEAL and the GN-5M-adapted encoders exceed EndoViT and Endo-FM) and the FID gains in Table 1 presuppose that the evaluation benchmarks are independent of the pretraining distribution. The paper never states this. GN-5M (Section 4.1) is a multicenter dataset from eight Dutch hospitals, including the authors' Amsterdam UMC; the private BE set originates from Amsterdam UMC, and POLAR is an Amsterdam-based public benchmark. If any BE or POLAR test patient's frames, or frames from the same endoscopic procedure, are also in GN-5M, the DINOv2/DINOv3 (GN-5M) teachers and REVEAL can memorize per-patient anatomy. This would inflate the linear-probing AUCs in Table 2 and Table 3. It could also lower the FID in Table 1 if the unspecified FID reference set (Section 4.3) includes GN-5M, since the generator could reproduce training-distribution frames. The conclusion that large-scale generative pretraining yields clinically useful representations thus rests on an untested independence assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces REVEAL, a latent diffusion transformer (SiT) trained with representation alignment on GN-5M, a 4.82M-frame multicenter endoscopic dataset. The model aligns noisy hidden states with a domain-adapted DINOv3-B teacher using an iREPA-style objective, and the authors ablate VAE backends, teacher encoders, model scale, dataset size, and training duration. They report FID improvements for generation and linear-probing AUC on the POLAR benchmark, a private Barrett's Esophagus dataset (BE), and a corrupted variant (BE-C). The central claims are that REVEAL is the largest generative foundation model for endoscopy to date, that domain-adapted representation alignment improves synthesis fidelity, and that its frozen features are competitive with or exceed endoscopy-specific discriminative foundation models such as EndoViT and Endo-FM.","tokens_in":12356,"tokens_out":7337,"duration_ms":67267,"significance":"If the claims hold, REVEAL would be a valuable contribution: it is the first large-scale generative pretraining for gastrointestinal endoscopy with a domain-adapted teacher, and the paper includes systematic ablations (Table 1) that disentangle latent space, teacher, scale, and data size. The public release of code and weights is a concrete strength, and the use of held-out external datasets (POLAR, BE, BE-C) for the classification claims is a positive feature. However, the synthesis-fidelity claim currently rests on an underspecified FID protocol, the evaluation sets are not shown to be patient-disjoint from the pretraining corpus, and the linear-probing protocol lacks patient-level stratification and variance reporting. These gaps prevent full verification of the paper's central comparative statements.","major_comments":[{"comment":"The FID protocol is underspecified. The text states only that \"FID scores are computed from samples generated using the Euler solver with 50 function evaluations (NFEs),\" without reporting the reference image set, its size, its preprocessing, or the number of generated samples used. Since the central fidelity claims (unguided 13.55 vs. aligned 10.18 at 150k iterations; 5.32 at 250k on the full set) depend entirely on these numbers, the authors must specify the reference set, confirm that it is disjoint from the training data, and state the sample count.","section":"Sec. 4.3, Table 1"},{"comment":"No patient-level or procedure-level deduplication is reported between GN-5M pretraining and the POLAR, BE, and BE-C evaluation sets. GN-5M was collected from eight Dutch hospitals including Amsterdam UMC; the private BE set originates from Amsterdam UMC, and POLAR is an Amsterdam-based public benchmark. Without an explicit exclusion statement, the reported AUC gains over EndoViT and Endo-FM, and the FID improvements if the FID reference is drawn from GN-5M, could reflect memorization of training patients rather than generalization. Please provide a patient-level overlap check or describe how the evaluation sets were kept disjoint from pretraining.","section":"Secs. 4.1-4.3, Tables 2-3"},{"comment":"The linear-probing protocol is ambiguous: \"We evaluate on the test sets of both benchmarks with five-fold cross-validation\" does not state whether the folds are patient-stratified or image-level, whether the probe is trained on the official training split and evaluated on the test split, or how the reported mean is obtained. No per-fold variance or confidence interval is given. Given the small test sets (273 BE images; 588 POLAR images), the reported margins (e.g., REVEAL vs. DINOv2, BE AUC 0.786 vs. 0.772) need variance estimates to support the comparative claims. Please report patient-level stratified cross-validation with per-fold results and confidence intervals.","section":"Sec. 4.3, Tables 2-3"},{"comment":"The benefit of the in-domain teacher for synthesis is currently supported only by FID numbers whose reference distribution is unspecified and likely overlaps with GN-5M; the teacher, the student, and the FID reference all derive from the same multicenter corpus. The held-out POLAR and BE benchmarks validate the discriminative features but do not validate the synthesis claim itself. I recommend either computing FID against an external endoscopic reference set (e.g., one of the hospitals not used in training, or POLAR frames) or adding a quantitative external evaluation (e.g., a reader study or a downstream segmentation utility test) to rule out the possibility that the fidelity gains reflect reproduction of training-distribution statistics.","section":"Sec. 4.2 and Sec. 5.4"}],"minor_comments":[{"comment":"The claim \"largest generative foundation model for endoscopy to date\" is not substantiated; please specify whether \"largest\" refers to parameter count, training data size, or both, and compare with prior generative endoscopy models such as Endora and Polyp-DDPM.","section":"Abstract and Sec. 1"},{"comment":"Typographical artifacts should be corrected: \"underrealisticimagingcorruptions\" in the abstract and \"T able 1\", \"T able 2\", and \"T able 3\" in the body appear to be formatting errors.","section":"Abstract and throughout"},{"comment":"The inpainting and outpainting evidence is purely visual. Since the paper claims \"robust structural coherence\" in these edits, please either add quantitative metrics (e.g., LPIPS or PSNR on held-out masks) or explicitly describe these results as anecdotal demonstrations.","section":"Sec. 5.4"},{"comment":"The comparison with EndoViT and Endo-FM should acknowledge the input-resolution mismatch: those models were pretrained and evaluated at 224x224, while REVEAL features are extracted at 256x256, which may contribute to the reported differences.","section":"Sec. 5.2"},{"comment":"The two-stage DINOv3-B (GN-5M) training curriculum does not state whether each 115,000-iteration stage uses the full GN-5M dataset or the 250k-image subset; please clarify for reproducibility.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a straightforward application of iREPA-style representation alignment to endoscopy with a domain-adapted teacher. The empirical scale and the systematic ablations are the main contributions. The evaluation-leakage concern is the most serious issue: without a patient-level exclusion statement, the comparisons against EndoViT and Endo-FM cannot be interpreted. The FID protocol and the linear-probing details are also easily fixable but currently prevent a full assessment. The authors should also verify that the \"largest generative foundation model\" claim is supportable against concurrent work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: REVEAL is a credible first large-scale generative foundation model for endoscopy, and the ablations support the main direction. The biggest problem is an untested independence assumption: no patient-level dedup is reported between GN-5M pretraining and the BE/POLAR evaluation sets, so the headline gains could partly reflect memorization.\n\nThe genuinely new content is empirical. The architecture follows REPA/iREPA with one meaningful change: the teacher is a DINOv3-B encoder continued-pretrained on the target distribution. The result that the in-domain teacher beats ImageNet teachers on FID and downstream probing is useful, and the Table 1 ablations are systematic (VAE, teacher, scale, data size). The paper is also honest about several limitations, e.g., the resampling-based inpainting artifacts.\n\nSoft spots, in rough order of importance. First, the data-leakage question. GN-5M includes eight Dutch hospitals including Amsterdam UMC; the private BE set is from Amsterdam UMC, and POLAR is an Amsterdam-based public benchmark. The paper never states that test-patient frames were excluded from pretraining. If any overlap exists, the AUC gains in Tables 2-3 and the FID gains in Table 1 are inflated. This is not a derivation error; it is a missing protocol statement that the authors can fix. Second, FID is reported without specifying the reference image set, sample count, or whether the reference is the training distribution. FID 5.32 is not externally interpretable. Third, all numbers are point estimates with no error bars or significance tests, on a small private test set. Fourth, the linear-probing protocol in Section 4.3 is internally inconsistent: it says accuracy in one place and AUC in another, and the feature layer selection is under-specified. Fifth, the code release has no commit hash, and the pretraining corpus is private; this limits reproducibility, though the claim of released weights softens it.\n\nThe 'surpasses all general-purpose encoders' claim in Section 5.2 is fine under a strict reading; DINOv3 (GN-5M) is not general-purpose. Section 5.3 similarly holds. So the overclaim concern is minor.\n\nWho is this for? Anyone working on medical generative models or endoscopy foundation models. The paper deserves serious refereeing: the empirical contribution is significant, and the main concerns are addressable with more transparency about dedup, FID protocol, and statistical rigor. I would send it to review with a request for major revision.","headline":"Credible first large-scale generative foundation model for endoscopy with systematic ablations, but the headline comparisons rest on an untested patient-level dedup assumption that needs explicit clarification.","tokens_in":13118,"tokens_out":3980,"would_cite":true,"duration_ms":33326,"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":"REVEAL is a generative foundation model for endoscopy trained on 4.82M endoscopic frames that produces high-fidelity images and features that beat endoscopy-specific encoders in several benchmarks.","keywords":["endoscopy","generative foundation model","representation alignment","flow matching","diffusion transformer","image inpainting","image outpainting","linear probing"],"falsifier":"Compute FID on a held-out endoscopic image set from hospitals that did not contribute to GN-5M and run a blinded clinical reader test in which experts try to distinguish REVEAL samples from real frames; if FID degrades sharply or experts reliably spot synthesis artifacts, the claim that aligned latents preserve generalizable anatomy would be weakened.","tokens_in":11890,"feed_emoji":"🩺","tokens_out":6927,"duration_ms":57893,"temperature":0.7,"pith_summary":"REVEAL is a generative foundation model for endoscopy trained on GastroNet-5M (GN-5M), a multicenter collection of 4,820,653 unlabeled endoscopic frames. The paper argues that representation alignment with an encoder pretrained on the same endoscopic distribution—rather than a general-purpose natural-image encoder—is what makes synthesis faithful, and that the resulting frozen transformer is itself a competitive visual feature extractor. On linear-probing benchmarks for polyp and Barrett's esophagus neoplasia, REVEAL's layer-8 features exceed EndoViT and Endo-FM, and its FID on generated GN-5M samples reaches 5.32 after 250k iterations with the SiT-L/2 backbone. If true, large-scale generative pretraining offers a viable alternative path to clinical visual representations, alongside masked-image-modeling encoders, with a single backbone for synthesis, editing, and downstream classification.","feed_headline":"REVEAL: 4.82M-frame generative model beats EndoViT and Endo-FM","feed_subtitle":"Representation-aligned training yields high-fidelity synthesis and competitive clinical features.","key_machinery":"The load-bearing mechanism is the representation-alignment objective in its iREPA variant: during training the student's hidden patch tokens $h_t$ are pushed toward the teacher's patch embeddings $y^*$ by maximizing cosine similarity, $$\n\\mathcal{L}_{\\text{REPA}} = -\\mathbb{E}_{x,\\epsilon,t}\\left[\\frac{1}{N}\\sum_{n=1}^N \\operatorname{sim}(y^*[n], h_\\$\\varphi$(h_t[n]))\\right],\n$$ added to the denoising loss with weight $\\lambda=1$. Two details make it work for endoscopy: a convolutional projection head that keeps neighboring tokens spatially coherent, and an instance-style spatial normalization applied to the teacher features that suppresses global variance and boosts local anatomical contrast. The teacher is a DINOv3 ViT-B/16 checkpoint continued on GN-5M, and the student is a SiT flow-matching transformer operating on a frozen SD2 VAE latent space. This mechanism injects domain-specific patch-level structure into the generative trajectory while leaving the trained transformer usable as a feature extractor.","core_discovery":"The paper's central claim is that domain-adapted representation alignment is the primary driver of synthesis fidelity in endoscopic diffusion transformers. Aligning the hidden states of a Scalable Interpolant Transformer (SiT) with patch-level features from a DINOv3-B encoder pretrained on GN-5M lowers FID from 13.55 (no alignment, SD2 VAE, SiT-B/2) to 10.18 with the in-domain teacher, and scaling the backbone to SiT-L/2 on the full dataset reaches 5.32. The same frozen backbone, probed at layer 8 with a linear head, achieves BE AUC 0.786 and POLAR AUC 0.758, beating both EndoViT (0.629/0.632) and Endo-FM (0.764/0.683) and all general-purpose encoders tested; it also keeps higher AUC than both endoscopy-specific models under synthetic imaging corruptions. REVEAL also demonstrates spatially coherent inpainting and outpainting via RePaint-style resampling, which the paper reads as evidence that the aligned latent space respects gastrointestinal morphology rather than just pixel statistics.","pith_inferences":["Editorial inference: the same representation-alignment recipe should transfer to other clinical imaging domains that have a large unlabeled archive and a domain-pretrained teacher, since nothing in the method is endoscopy-specific.","Editorial inference: because the teacher, the student, the FID reference, and the pretraining distribution all come from GN-5M, the headline fidelity gain may partly reflect familiarity with the training manifold; an external FID set or a blinded clinician study would separate anatomical realism from memorization.","Editorial inference: using intermediate diffusion timesteps $t>0$ for feature extraction, which the authors note naturally denoise corrupted inputs, is a promising cheap route to further robustness gains without retraining."],"forward_implications":["A single generative backbone trained without labels can double as a classification feature extractor, outperforming dedicated endoscopy encoders on polyp and Barrett's esophagus linear probing.","Using an in-domain teacher matters more than model size or VAE choice: the DINOv3-B (GN-5M) teacher gives the best FID and POLAR results among all alignment targets tested.","Training on the full 4.82M-frame distribution yields the largest single gain in both fidelity and representation quality, and continued training past 200k iterations gives diminishing returns.","The same frozen model supports unconditional generation, inpainting, and outpainting without architectural changes, so synthesis and editing tools can be built directly on the released weights.","Public release of weights lowers the compute threshold for downstream conditional synthesis, segmentation, and out-of-distribution detection in gastroenterology."],"supporting_citations":[{"why":"Introduces the representation-alignment objective and the 17x training speedup that REVEAL extends to endoscopy.","marker":"[38]"},{"why":"Shows spatial structure rather than global semantics drives generation and supplies the convolutional projection and spatial normalization used here.","marker":"[32]"},{"why":"Provides GastroNet-5M, the 4.82M-frame multicenter corpus used for pretraining encoders, training the student, and computing FID.","marker":"[16]"},{"why":"Defines the Scalable Interpolant Transformer backbone that REVEAL trains with alignment.","marker":"[25]"},{"why":"Supplies the DINOv3 checkpoint the authors continue training on GN-5M to obtain the in-domain teacher.","marker":"[31]"},{"why":"EndoViT is a baseline endoscopy-specific foundation model REVEAL compares against on classification.","marker":"[1]"},{"why":"Endo-FM is the masked-image-modeling endoscopy foundation model REVEAL exceeds on BE and POLAR linear probing.","marker":"[35]"},{"why":"RePaint's resampling strategy is used for the inpainting and outpainting structural-coherence probes.","marker":"[24]"}],"fun_headline_variants":["REVEAL: 5M-frame endoscopy model beats EndoViT and Endo-FM","Aligning latents with in-domain encoder cuts FID to 5.32","REVEAL’s frozen backbone outclasses tuned endoscopy models","Domain-aligned latents yield high-fidelity endoscopy synthesis and features","In-domain alignment: better synthesis and robust features for endoscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the in-domain teacher, trained on the same corpus the generative student learns, transfers generalizable anatomical structure rather than memorized statistics of GN-5M.","fun_headline_variants_meta":{"raw":{"variants":["REVEAL: 5M-frame endoscopy model beats EndoViT and Endo-FM","Aligning latents with in-domain encoder cuts FID to 5.32","REVEAL’s frozen backbone outclasses tuned endoscopy models","Domain-aligned latents yield high-fidelity endoscopy synthesis and features","In-domain alignment: better synthesis and robust features for endoscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0014,"raw_usage":{"total_tokens":5704,"prompt_tokens":1031,"completion_tokens":4673,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":4571}},"tokens_in":647,"tokens_out":4673,"duration_ms":34002,"temperature":1.0,"reasoning_tokens":4571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:30:44.758544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute FID on a held-out endoscopic image set from hospitals that did not contribute to GN-5M and run a blinded clinical reader test in which experts try to distinguish REVEAL samples from real frames; if FID degrades sharply or experts reliably spot synthesis artifacts, the claim that aligned latents preserve generalizable anatomy would be weakened.","supporting_citations":[{"cited_title":"In: The Thirteenth International Conference on Learning Representations (2025)","cited_arxiv_id":null,"evidence_quote":"Introduces the representation-alignment objective and the 17x training speedup that REVEAL extends to endoscopy."},{"cited_title":"Gastroenterology (2025)","cited_arxiv_id":null,"evidence_quote":"Provides GastroNet-5M, the 4.82M-frame multicenter corpus used for pretraining encoders, training the student, and computing FID."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Endo-FM is the masked-image-modeling endoscopy foundation model REVEAL exceeds on BE and POLAR linear probing."}],"review_version":1}