Paper Slam 4/25: When AI Starts to "See" — Diagnosing Lesions in Video and Reversing Hallucinations in Photos

> arXiv:2604.21814 | Divide-then-Diagnose: Weaving Clinician-Inspired Contexts for Ultra-Long Capsule Endoscopy Videos > arXiv:2604.21879 | Addressing Image Authenticity When Cameras Use Generative AI

---

Introduction: Two Questions About "Seeing"

Picture yourself as a gastroenterologist. Today you must review not an X-ray but a movie — an 8-to-12-hour film shot by a capsule camera a patient swallowed, capturing a few frames per second along the entire intestine. That can total ~100,000 frames, of which fewer than 10 have real diagnostic value. The rest are normal mucosa, motion blur, bubble reflections, food residue, rapidly shifting viewpoints. Your job: find those few frames among the hundred thousand, characterize the lesions, write a report.

Now switch scenes. You take a photo at night with your smartphone. The camera's AI ISP automatically performs low-light enhancement — brightening shadows, sharpening detail. But looking closely at a street sign, you notice the AI has *changed* the letters: characters too blurry to read have been "guessed" into different ones. Or portrait mode turns brown eyes green. The camera output this image directly, and you naturally assume it is "real." It isn't.

These scenarios look unrelated — one is medical diagnosis, the other consumer photography. But they point to the same deep question: when AI enters the process of "seeing," what is it actually doing? Is it helping us discover the truth, or manufacturing illusions we need to guard against?

Two papers appearing on arXiv on April 23, 2026 stand at opposite ends of this question.

The first, from HKUST and Renji Hospital (Shanghai Jiao Tong University School of Medicine), asks: how can AI "read" an ultra-long capsule endoscopy video the way a clinician does, extracting reliable diagnoses from sparse, noisy evidence? This is not video compression or object detection — it is decision-making under extreme information sparsity and high uncertainty.

The second, from the University of Toronto and Samsung AI Center, asks: when an in-camera AI ISP hallucinates, can we "roll back" the hallucination after the shot and recover the unmodified image? This is not classic denoising or super-resolution — it is inverting an AI-modified image without access to the camera's internal ISP.

Both papers tackle the hardest case in AI vision: the evidence is sparse, ambiguous, unreliable — but the conclusion matters and is irreversible. One searches for faint signal in a sea of noise; the other recovers tampered truth from AI-"beautified" images. Let me dissect each, then compare them — what core ideas they share, and where they diverge.

---

Chapter 1: DiCE — Mimicking How Doctors Read Exams

1.1 First, Understand the Problem

Capsule endoscopy (CE): the patient swallows a pill-sized camera that travels naturally down the digestive tract, capturing tens of thousands of images wirelessly. No intubation, no anesthesia. For small-bowel screening it is nearly irreplaceable.

But there is a huge clinical pain point: a typical CE video runs 8-12 hours with tens to hundreds of thousands of frames. Clinicians review frame by frame, averaging over 1 hour per case. The DiCE paper cites deployment studies from 2023-2024 showing that even with AI-assisted frame-level analysis, review time still exceeds 1 hour — and only ~8% of AI-screened images contain meaningful lesions.

Why doesn't AI help more? Because existing methods work almost entirely at the frame level: given a single frame, decide if it contains a lesion. That works on curated datasets like Kvasir-Capsule, but real videos contain motion blur, bubbles, reflections, residue. A lesion may appear for only seconds, surrounded by normal tissue. Frame-by-frame classifiers produce false positives or miss lesions whose frames happen to be blurred.

The DiCE authors did something important: they redefined the task. Not "video classification" or "anomaly detection" but "diagnosis-driven CE video summarization" — extract from the whole video the key evidence frames that support the final clinical diagnosis, with accurate lesion classification.

Why does this matter? An "anomalous" frame isn't necessarily diagnostic. A bubbly stretch may look odd but would never enter the report. What matters is lesions that directly influenced the final diagnostic decision. DiCE's output is not "this frame is anomalous" but "this lesion event is an ulcer, located here in the video, supported by these consecutive frames."

To support this task they built VideoCAP: 240 full-length CE videos from two clinical centers at Renji Hospital, annotated not frame-by-frame but directly from clinical diagnostic reports, each annotation corresponding to a lesion that actually influenced diagnosis, with precise timestamps. Twelve lesion categories: ulcer, erosion, vascular malformation, mucosal erythema, protruding lesions, hematoma, lymphangiectasia, lymphoid follicular hyperplasia, polyp, parasite, intestinal fluid retention, normal small-bowel mucosa.

240 videos, 7.24 million frames total. Perhaps 1 diagnostically valuable frame per 1,000. That is the essence of the problem: extreme sparsity.

1.2 DiCE's Core Idea: Divide, then Diagnose

Clinicians reviewing CE video don't jump to frame 35,842 at random. Their standard workflow: 1. Fast skim — flag suspicious time segments; 2. Local close reading — return to flagged segments, compare consecutive frames, confirm the lesion type; 3. Synthesis — integrate confirmed lesions into a report.

DiCE translates this into a three-stage framework:

Stage 1: Selector (candidate screening) — A lightweight binary classifier (frozen DINOv3 backbone + MLP head) scores each frame for potential diagnostic relevance. Low threshold (τs=0.5) — recall is king. This shrinks 100,000 frames to a manageable candidate set.

Stage 2: Context Weaver — DiCE's core innovation. Candidates are grouped by anatomical position via two-level hierarchical clustering:

• Coarse (Anatomical Context Anchoring): time-visual affinity groups candidates into "anatomical contexts," ensuring different gut regions are covered rather than repeatedly sampling one photogenic area;
• Fine (Lesion Context Refinement): within each anatomical context, finer clustering groups observations of the same lesion together and separates different lesions.

A clever prior: within a "lesion context," most frames should support the same lesion type. This "dominant-label prior" isn't explicitly supervised but guides downstream evidence fusion.

Stage 3: Evidence Converger — Given clustered contexts, what lesion does each represent? The authors found single-frame predictions are extremely unstable: strong baselines (e.g., ViLAMP+LingShu) produce contradictory diagnosis labels on roughly half of adjacent keyframe pairs within 30 seconds (Figure 4). Evidence Converger fuses evidence within each context: 1. Multi-frame fusion: sum the per-frame class distributions so consistent evidence reinforces and noise cancels; 2. Internal consistency refinement: drop frames disagreeing with the context consensus (e.g., one bubble-obscured frame saying "erosion" when everything else says "ulcer"); 3. Cross-context pruning: remove contexts that remain "normal" or low-confidence after refinement.

Each surviving lesion context contributes one representative frame (feature-space medoid) with the aggregated label — the final video summary.

1.3 What the Experiments Say

DiCE uses a 0.2B-parameter DINOv3 backbone, far smaller than 8B MLLMs (Qwen3-VL, InternVL3.5). The results are striking:

• Lesion Detection Rate (LDR): 44.12% vs. best baseline 35.29% (ViLAMP+LingShu);
• Sensitivity: 85.29%, far above baselines;
• Temporal error: 54.86s, the most precise;
• Redundancy: 77.81%, the lowest (little repetition in the summary);
• Diagnostic Yield: 20% — the fraction of patients whose *entire* clinical report lesion set was detected. Baselines manage only 12.5-15%.

Ablations are even more convincing:

• Removing Context Weaver (replaced by simple 300-second time windows): LDR drops 22.76 points, sensitivity drops 46.46 points. Temporal proximity alone cannot organize diagnostic context;
• Removing Evidence Converger (single highest-confidence frame instead): sensitivity drops 55.88 points and diagnostic yield falls from 20% to 2.5%. Single-frame prediction is nearly unusable in CE.

Another key finding: DiCE cut adjacent-keyframe label inconsistency from baselines' 46-65% down to 5.1-10.7% — an 8-9x improvement at 30s and 60s thresholds. The summaries are temporally coherent, so doctors aren't left wondering "is this an ulcer or an erosion?"

1.4 A Feynman-Style Critique

First, naming ≠ understanding. Behind the jargon — Context Weaver, Evidence Converger, dominant-label prior — are three simple things: screen suspicious regions, group related frames, vote within groups. If you can say that to someone who hasn't read the paper, you understand DiCE.

Second, what does it actually solve? DiCE's core contribution is not a new architecture or loss. It is redefining the task, then solving the redefined problem with a plain three-stage pipeline (screen–group–vote). It beats 8B MLLMs not because it is "smarter" but because it hews closer to how clinicians actually work.

Third, any self-deception? A potential blind spot: VideoCAP is annotated retrospectively from clinical reports, which themselves may contain errors or omissions. Trained on VideoCAP, the model may learn "how to match existing reports" rather than "how to make the most accurate independent diagnosis." The paper acknowledges this and lists prospective validation with clinicians as future work.

Fourth, cargo-cult check: no giant models, no stacked attention, no RLHF. A 0.2B backbone, an MLP head, hierarchical clustering, and voting — beating every 8B MLLM baseline. DiCE understood the problem, so it needed no fancy packaging.

---

Chapter 2: Camera Hallucinations — Reversing the AI "Beautification"

2.1 The Essence: Camera Output Is No Longer Trustworthy

We traditionally treat a camera's output as "real" — a record of the optical information in front of the lens. Yes, there's Photoshop and deepfakes, but those are post-capture human edits. RAW/JPEG straight from the camera counted as raw evidence.

That is changing. Modern smartphone ISPs increasingly embed deep-learning modules, especially in:

• AI digital zoom / super-resolution: at 10x zoom, the ISP's model "guesses" missing details. The paper shows a license plate where blurry "R" and "M" became "A" and "N";
• Low-light enhancement / night mode: AI brightens and sharpens, potentially inventing textures and text that were never there.

Crucially, this happens at capture time, inside the hardware, invisibly to the user. Some pixels in your photo were not "recorded" — they were "generated."

Not all AI enhancement is bad; most of it (edge sharpening, texture) is harmless or helpful. The problem arises when AI changes semantics — altering a license-plate character or garbling a dim QR code.

Existing image-forensics work focuses almost entirely on post-hoc tampering. The paper cites one related work: Punnappurath et al. (2024), which computes a binary mask of "possibly hallucinated" pixels at capture time and stores it in JPEG metadata. But it only tells you "this pixel is suspicious," not "what it should look like." It detects; it doesn't restore. This paper restores — reconstructing an "unhallucinated" version from an already-processed image.

2.2 The Method: 180KB of Metadata Magic

Problem: given an AI-ISP-processed image y, recover the pre-processing image x. Key difference from classic restoration: we don't know which AI model the ISP used — it may be a third-party or proprietary black box, and we have no access to ISP internals.

The method rests on two insights:

Insight 1: Hallucinations typically occur at the end of the ISP pipeline. Early stages (denoising, demosaicing, white balance) aim to faithfully restore sensor signal. Later enhancement stages (tone mapping, super-resolution, low-light boost) are trained with generative losses (GAN, perceptual) and are the main hallucination source. So there is an intermediate point — the input to the AI enhancement module — where image x is still "real." If the camera vendor can capture this intermediate image (the paper assumes they can), the (x, y) pairs let us learn to recover x from y.

Insight 2: Each image needs an image-specific decoder. A general image-to-image model (e.g., a 64MB NAFNet) fails on tasks like low-light enhancement — an inherently many-to-one mapping: the same dark image could have been brightened in many ways, so from y you cannot uniquely infer how dark x was. Blind restoration is fundamentally ambiguous here.

So the authors turn to implicit neural representations, with a twist: optimize a lightweight MLP decoder per image. Two components: 1. Modality-specific encoder Φ: pretrained per hallucination type (natural-image SR, text SR, low-light enhancement), NAFNet-based, ~31.75K parameters (127KB). Maps input y to a k-dimensional latent feature map w; 2. Image-specific MLP decoder Θ: two hidden layers of 64 neurons each, ~13K parameters (53KB).

At capture time: Φ is frozen; Θ is fine-tuned for the specific image. Feed y through Φ to get w; concatenate pixel coordinates with w; the MLP predicts a residual; the recovered image is x̂ = y - Θ([x, y, w]). Loss is plain reconstruction ||x - x̂||² — no GAN or perceptual loss, since those would make the decoder hallucinate too. Fine-tuning takes 1000 iterations, ~3 seconds on a V100. The total 180KB of parameters is embedded in standard JPEG/HEIC metadata fields.

At viewing time: extract the parameters from metadata, re-run encoder + MLP on y, and get the de-hallucinated x̂. No camera-ISP access needed.

2.3 Why It Works

• Encoder + MLP separation, not pure coordinate MLPs (SIREN/NeRF): those need slow per-image optimization from scratch (100K steps, minutes — unacceptable at capture). Pretraining an encoder to produce a useful intermediate representation lets a small MLP fine-tune reach or exceed that quality in 3 seconds.
• Residual prediction: y and x usually differ little, so the MLP only learns to "correct" problem regions rather than reconstruct the whole image.
• Modality-specific encoders: one general encoder for all tasks loses 1.5dB PSNR — text SR and natural-image SR have different hallucination patterns. The camera knows which AI module is active (digital zoom vs. night mode) and picks the matching encoder.

2.4 Results

Evaluated on three tasks:

• Natural-image SR: DIV2K, 4x, RealESRGAN as hallucination source;
• Text SR: MARCONet;
• Low-light enhancement: LOL, AutoDIR.

Results (Table 1):

• DIV2K: 32.96dB, beating SIREN (28.75), NeRF (29.46), hashgrid (29.20), and blind NAFNet (32.25);
• MARCONet text SR: 31.26dB, far above all baselines (best: hashgrid at 30.32);
• LOL low-light: 36.34dB, again best.

Compared with JPEG-compressed residuals: at 180KB only 27.44dB; even 416KB (QF=30) gives 34.67dB, still below the method's 35.12dB (Table S1). And the metadata size is resolution-independent — always 180KB regardless of image size, while residual images grow linearly with resolution.

Qualitative results: an AI SR turned "GQB" into "608"; the method recovered the original characters. Low-light enhancement turned "i" into "l"; the method restored the "i".

2.5 A Feynman-Style Critique

First, strip the names: this is simply learning, for each image, a mapping from "AI-modified version" to "pre-modification version." Every photo is modified differently (content, lighting, camera settings), so one big model can't cover all cases. Solution: a big encoder to understand the image, a small per-image fine-tuned model — 180KB.

Second, cargo cult? No. A 2-layer MLP (53KB) plus a light encoder (127KB). The innovation is "we found a compact, practical way to store and reverse AI tampering," not "we used the newest technique."

Third, what's left unsaid?

• It assumes capture-time access to the pre-enhancement intermediate x for paired training — fine for camera vendors, useless for third parties analyzing an existing JPEG;
• Metadata security: could maliciously altered 180KB parameters fabricate "restored" images that are themselves fake? Encryption and watermarking are declared out of scope;
• Only SR and low-light enhancement are validated. Future end-to-end AI ISPs may have no clean "intermediate point x" — will the method survive?

Fourth, how small is 180KB? An ordinary iPhone JPEG is 2-5MB. So 3-9% extra storage per photo buys a "de-AI-ify" button — a trade-off entirely acceptable in real products.

---

Chapter 3: Head-to-Head

3.1 Shared Problem Structure: Sparse Evidence + High Uncertainty + Irreversible Conclusions

DiCE: fewer than 10 meaningful frames in 100,000, often degraded by blur and bubbles, yet the diagnosis must be reliable. De-hallucination: most pixels are "real," but a small AI-fabricated fraction can flip the image's semantics.

Common traits:

• Extremely low SNR: DiCE's useful signal is ~0.01% of the data; tampered pixels may be ~5% of an image but concentrated on text, faces, key markings;
• High cost of error: a missed ulcer delays treatment; a wrong plate character invalidates legal evidence;
• Conventional methods fail: single-frame classifiers are unstable; general-purpose restoration models can't handle per-image ambiguity.

3.2 Shared Strategy: Context + Aggregation

Both papers say: don't trust a single observation; aggregate in context. DiCE's Context Weaver + Evidence Converger: one frame is unreliable, but a temporally and visually coherent group provides solid evidence. The de-hallucination encoder learns a modality-level "context" (how this class of AI enhancement modifies images), then the MLP adapts per image — single pixels are unreliable, but global features plus coordinates can infer what was changed.

3.3 Key Divergence: Data-Oriented vs. Model-Oriented

DiCE is data-oriented: its core contribution is redefining the task and data (VideoCAP), proving that a simple screen–group–vote pipeline beats 8B MLLMs once the task is right. The bottleneck is task definition and evaluation, not model capacity.

De-hallucination is model-oriented: its contribution is a concrete deployable scheme (encoder + MLP + metadata), proving 180KB can invert ISP hallucinations. The bottleneck is compactly representing the per-image inverse mapping.

3.4 Another Divergence: Clinical Loop vs. Consumer Product

DiCE is inherently clinical: it needs diagnosis-report annotations, its metrics (Diagnostic Yield, patient-level detection) map to clinical utility, and its goal is freeing doctors from 1-hour reviews. De-hallucination is inherently a consumer feature: 180KB fits JPEG metadata, 3-second fine-tuning runs on-device, and every photographer gets a button for the "AI-untouched" version. Their techniques don't transfer: hierarchical clustering is meaningless for a single static photo; 3 seconds × 100,000 frames is infeasible.

3.5 Deeper Contrast: The Direction of Trust

DiCE essentially says: "doctors are exhausted; let AI help diagnose." Its premise is trust in AI-assisted diagnosis, provided the method mimics clinician workflow. De-hallucination says: "AI is secretly altering things in your camera; you need a tool to fight it." Its premise is distrust of AI 'beautification' and a user's right to see the untouched version.

Together the two papers (which appear in the same CVPR 2026 session) form one narrative: AI vision systems can help us see what was previously invisible — and can make us see what doesn't exist. Good engineering should amplify the former and contain the latter.

---

Chapter 4: Deeper Reflections — What Problem Are We Actually Solving?

4.1 Medical AI's Special Predicament: From "Assisted" to "Replacement"

DiCE's numbers are encouraging, but a Feynman-style question: what do they mean? DiCE is better at "matching lesions described in clinical reports." Is that the same as "better at clinical diagnosis"? Not necessarily — reports themselves err, and VideoCAP inherits their blind spots: a lesion missed by the report gives the model no incentive to find it.

Deeper: DiCE's Diagnostic Yield is only 20% — for 80% of patients, the model missed at least one reported lesion. In practice, doctors must still review every video; DiCE provides a summary, not a replacement. To DiCE's credit, the authors report this honestly. But it exposes an industry-wide self-deception: papers report AUC, sensitivity, specificity — clinicians ask "can I skip half the video without missing anything?" At 20% diagnostic yield, AI is still far from that goal in CE.

4.2 The Legal and Ethical Dimensions of Camera Hallucination

After an accident, a phone photo shows plate "608"; the restored image shows "GQB." Which is evidence in court? What if the 180KB metadata was tampered with? What if vendors refuse to release the pre-enhancement image for third-party training? These questions exceed pure technique, but technical choices shape the answers: encrypted signatures prevent metadata tampering; legislation could mandate unhallucinated versions. This paper lays the technical groundwork for that discussion.

4.3 "Context" as a General Antidote

Both papers lean on context — Context Weaver weaves diagnostic context; the encoder provides modality context. Is lack of contextual understanding a common root of AI vision failures? A single-frame classifier doesn't know where the frame sits in the video; a general restoration model doesn't know whether it's processing text or a natural scene. In information-theoretic terms, context is conditioning: P(lesion | one frame) is low, P(lesion | 30s of similar content around it) is far higher. Context isn't magic — it costs computation and pretraining — but both papers show the price is worth paying.

4.4 The Feynman Question: Do You Really Understand?

For DiCE, the real understanding is: why does this simple pipeline beat 8B MLLMs? Because MLLMs are trained for "general video understanding," implicitly assuming roughly uniform information density — an assumption CE videos violently violate. DiCE succeeds by abandoning generality and embracing domain specificity.

For de-hallucination: why is per-image fine-tuning better than one big model? Because ISP hallucinations aren't uniformly distributed — they depend on content, lighting, noise. A big model "averages" over cases and necessarily fails on fundamentally ambiguous ones (how dark was it originally?). Per-image fine-tuning says: "I won't guess universal rules; I'll learn this photo's specific inverse mapping."

4.5 Cargo-Cult Detection

A growing cargo cult in AI research: bigger models, more data, parameters = capability. DiCE beat 8B rivals with 0.2B parameters. De-hallucination solved what big models couldn't with 180KB. The joint message: on some problems, problem structure matters more than model scale. Large models aren't useless — but before chasing scale, ask: do I truly understand the problem? If yes, 0.2B parameters may suffice. If not, even 64B is just building a grander control tower out of bigger bamboo.

---

Conclusion: Two Revolutions in "Seeing"

DiCE represents constructive AI — entering medicine not to replace doctors but to free them from mechanical mass screening, encoding expert workflows into computable steps. De-hallucination represents defensive AI — recognizing that AI itself can be the problem, so another AI must supervise and correct it, giving users the choice between "beautified" and "real."

Both directions matter; neither can be abandoned. From a Feynman perspective, the best thing about both papers is that they honestly face the nature of their problems: DiCE doesn't pretend CE videos are like YouTube clips; de-hallucination doesn't pretend one general model can invert every mapping. Both choose to confront the problem's specificity with sufficiently simple, targeted solutions.

Neither paper is perfect — DiCE's 20% diagnostic yield leaves room; de-hallucination's metadata security and generalization await validation. But both take solid steps in the right direction.

> "The first principle is that you must not fool yourself — and you are the easiest person to fool." > > As AI vision embeds itself ever deeper into our lives, this may matter more than ever. Whether diagnosing patients or taking photos, we must trust AI's abilities while staying clear-eyed about its limits. > > Don't fool yourself. Start by understanding the problem.

---

Paper index

• Bowen Liu, Li Yang, Shanshan Song, Mingyu Tang, Zhifang Gao, Qifeng Chen, Yangqiu Song, Huimin Chen, Xiaomeng Li. "Divide-then-Diagnose: Weaving Clinician-Inspired Contexts for Ultra-Long Capsule Endoscopy Videos." arXiv:2604.21814 [cs.CV], April 2026.
• Umar Masud, Abhijith Punnappurath, Luxi Zhao, David B. Lindell, Michael S. Brown. "Addressing Image Authenticity When Cameras Use Generative AI." arXiv:2604.21879 [cs.CV], April 2026.

*Written by Xiaokai using a Feynman-style thinking framework. Any misinterpretations are my own.*