Skip to content
Computational Pathology Reading Roadmap

Restoration and generative vision

Restoration estimates an image from a degraded observation. Generative modeling estimates or samples from a data distribution. They meet when a learned prior is used to denoise, super-resolve, inpaint, edit, or conditionally synthesize an image, but the scientific claims remain different.

Latest literature discovery audit: 2 October 2026. Existing paper reviews retain their individual source-check dates.

Individual ICLR oral-paper studies retain their 19 September 2026 source-check dates; the later discovery audit does not re-date that evidence.

ECCV 2026 conference additions and metadata checked through 2026-09-12; the broader literature audit boundary above is unchanged.

Diffusion, flow, autoregressive, and pixel-space endpoints appear inside their mechanism families rather than in a dated frontier section.

Prerequisites and objectives

Section titled “Prerequisites and objectives”

You should know probability densities, expectation, KL divergence, maximum likelihood, latent variables, convolution/transformers, and the evaluation contract in the CV overview. By the end, you should be able to:

  • write a forward degradation model and identify what cannot be recovered uniquely;
  • distinguish explicit likelihood, variational bounds, implicit adversarial distributions, score models, flows, and autoregressive factorizations;
  • trace pixel, latent, discrete-token, noise, velocity, and text-conditioning paths;
  • separate reconstruction, fidelity, diversity, alignment, geometry, memorization, and sampling efficiency;
  • design a controlled comparison without treating one perceptual metric as truth.

Restoration is an inverse problem

Section titled “Restoration is an inverse problem”

A common observation model is

y = D(x) + ε

where D may blur, downsample, mask, compress, or otherwise transform a clean image x, and ε is noise. Many x can produce the same y; the estimator necessarily injects assumptions or a learned prior.

Stage Primary reading · year · venue/status Read it for
R1 SRCNN — 2014 · ECCV, peer reviewed mapping bicubic low-resolution images to high-resolution targets with a shallow CNN
R2 DnCNN — 2017 · IEEE TIP, peer reviewed residual noise prediction and one architecture across restoration tasks
R3 EDSR — 2017 · CVPR workshop, peer reviewed removing unnecessary normalization and scaling residual super-resolution models
R4 SwinIR — 2021 · ICCV workshop, peer reviewed shifted-window features for super-resolution, denoising, and compression artifacts
R5 Restormer — 2022 · CVPR, peer reviewed attention designed around high-resolution restoration cost

PSNR and SSIM reward fidelity to one reference under a specified degradation. A plausible texture may lower distortion while inventing content; a high perceptual score does not make invented detail true.

Likelihood and latent-variable foundations

Section titled “Likelihood and latent-variable foundations”

An autoregressive model factorizes

p(x) = ∏_(i=1)^n p(x_i | x_<i)

which provides a normalized likelihood but requires an ordering and potentially many sequential sampling steps.

A variational autoencoder optimizes the evidence lower bound

log p_θ(x) ≥ E_(q_φ(z | x))[log p_θ(x | z)] − D_KL(q_φ(z | x) ‖ p(z))

The latent bottleneck controls compression and reconstruction; the likelihood model defines what “reconstruction error” means.

Family Primary reading · year · venue/status Central mechanism
variational latent Auto-Encoding Variational Bayes — 2014 · ICLR, peer reviewed amortized inference and a variational lower bound
adversarial implicit Generative Adversarial Nets — 2014 · NeurIPS, peer reviewed a generator–discriminator game without tractable likelihood
pixel autoregression Pixel Recurrent Neural Networks — 2016 · ICML, peer reviewed exact sequential likelihood over pixels
discrete latent Neural Discrete Representation Learning — 2017 · NeurIPS, peer reviewed vector-quantized tokens and a learned codebook
invertible flow Real NVP — 2017 · ICLR, peer reviewed exact change of variables through invertible transforms
style-controlled GAN StyleGAN — 2019 · CVPR, peer reviewed mapping/style modulation and multiscale stochastic detail

Likelihood, visual fidelity, coverage, and useful representation are different axes. Do not infer one from another.

Paired and unpaired image translation

Section titled “Paired and unpaired image translation”

Image-to-image translation introduces a second observation domain. The pairing contract is part of the method: aligned pairs can supervise a conditional mapping, whereas unpaired data identify a mapping only through stronger distributional and cycle assumptions.

Stage Primary reading · year · venue/status Read it for
T1 Pix2Pix — 2017 · CVPR, peer reviewed paired conditional GAN training, an L1 reconstruction term, and a local PatchGAN discriminator
T2 CycleGAN — 2017 · ICCV, peer reviewed unpaired adversarial mappings constrained by forward–backward cycle consistency

Neither objective guarantees semantic preservation. A cycle can be satisfied while fine detail is hidden, altered, or hallucinated; paired supervision can still inherit registration error. Validate the structures required by the downstream claim, not only visual plausibility. The generative and virtual pathology route applies this distinction to stain translation and whole-slide evidence.

Diffusion and score models

Section titled “Diffusion and score models”

A forward diffusion process gradually corrupts x₀:

q(x_t | x₀) = N(√ᾱ_t x₀, (1 − ᾱ_t)I)

A network may predict noise ε, clean data x₀, score ∇_(x_t) log p_t(x_t), or velocity. The scheduler and parameterization alter optimization and sampling. Classifier-free guidance combines conditional and unconditional predictions, often improving prompt adherence while reducing diversity or creating artifacts at high guidance.

Stage Primary reading · year · venue/status Read it for
F1 Denoising Diffusion Probabilistic Models — 2020 · NeurIPS, peer reviewed forward corruption, learned reverse transitions, and variational training
F2 Score-Based Generative Modeling through SDEs — 2021 · ICLR, peer reviewed continuous-time score learning, reverse SDEs, and probability-flow ODEs
F3 Latent Diffusion Models — 2022 · CVPR, peer reviewed separating perceptual compression from conditional denoising
F4 Elucidating the Design Space of Diffusion-Based Generative Models — 2022 · NeurIPS, peer reviewed separating preconditioning, loss weighting, noise sampling, network design, and sampler choices
F5 Scalable Diffusion Models with Transformers — 2023 · ICCV, peer reviewed replacing a convolutional denoiser with a patch-token transformer
F6 Consistency Models — 2023 · ICML, peer reviewed learning consistent points on a probability-flow trajectory for few-step sampling
F7 PixelDiT — 2026 · CVPR, peer reviewed a hierarchical pixel-space diffusion transformer without a separately trained autoencoder
F8 Progressive Distillation — 2022 · ICLR, peer reviewed repeatedly halving diffusion sampling steps while retaining teacher-generated trajectory targets
F9 Scaling Rectified Flow Transformers — 2024 · ICML, peer reviewed rectified-flow training and multimodal diffusion-transformer scaling for high-resolution text-to-image synthesis

Latent generation inherits the autoencoder’s reconstruction ceiling. Pixel-space generation avoids that bottleneck but pays a severe sequence and compute cost; the comparison must include the full encoder/decoder and sampling pipeline.

Flow matching and modern autoregression

Section titled “Flow matching and modern autoregression”

Flow matching learns a time-dependent vector field v_θ(x_t, t) along chosen probability paths:

dx_t / dt = v_θ(x_t, t)

The path, target velocity, solver, and number of function evaluations are part of the method.

Branch Primary reading · year · venue/status Design question
continuous flow Flow Matching for Generative Modeling — 2023 · ICLR, peer reviewed how conditional probability paths make vector-field regression simulation-free
conditional / OT coupling Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport — 2024 · TMLR, peer reviewed how conditional-path choice and independent versus minibatch-optimal-transport endpoint coupling change trajectory complexity
next-scale autoregression Visual Autoregressive Modeling — 2024 · NeurIPS, peer reviewed coarse-to-fine next-scale prediction instead of raster-order next-token prediction
continuous-token autoregression Autoregressive Image Generation without Vector Quantization — 2024 · NeurIPS, peer reviewed diffusion loss over continuous tokens in autoregressive and masked-autoregressive models
bidirectional few-step flow BiFM — 2026 · CVPR, peer reviewed jointly learning image-to-noise inversion and noise-to-image generation under few-step constraints

Flow-related speed claims must disclose whether inference savings required a teacher, an additional distillation stage, or substantially more training. Keep continuous image transport separate from Discrete Flow Matching, which defines categorical probability paths for language and other discrete data. Within continuous flow, also report the endpoint coupling: minibatch OT can simplify training trajectories, but its result depends on batch construction and is not an exact population-level optimal-transport solution.

Do not compare only best FID. Record tokenizer/reconstruction quality, training compute, sampling steps, wall-clock time, guidance, rejection or reranking, and the number of generated samples used to compute the metric. The original FID paper fits Gaussians to Inception features; sample count, feature implementation, resizing, reference split, and domain mismatch can change the result.

Conditioning, control, and editing

Section titled “Conditioning, control, and editing”

Conditioning may enter through concatenation, adaptive normalization, cross-attention, an input image, spatial maps, masks, depth, edges, or retrieved references. For any conditional model, ask:

  1. Is the condition present in training, inference, or both?
  2. Which components are frozen?
  3. Does the output preserve non-target content?
  4. Can the conditioning source leak the answer or reference image?
  5. Is “editing” measured against an intended change and preservation mask?

Text alignment, attractive samples, and grounded spatial compliance require different evaluation.

Branch Primary reading · year · venue/status Mechanism and audit
spatial condition ControlNet — 2023 · ICCV, peer reviewed a trainable copied branch conditions a frozen diffusion backbone; test condition adherence separately from prompt fidelity
instruction editing InstructPix2Pix — 2023 · CVPR, peer reviewed synthetic edit-instruction pairs and dual conditioning; measure requested change and preservation separately
few-step inversion/editing BiFM — 2026 · CVPR, peer reviewed bidirectional average velocities for fast inversion and generation; disclose pretrained-model access and function evaluations

Controlled lab

Section titled “Controlled lab”

Mechanism lab

Section titled “Mechanism lab”

First implement a tiny VAE and diffusion/flow model on a two-dimensional mixture. Plot samples, density or trajectories, mode coverage, and sampling cost. This makes collapse and path geometry visible without image-scale confounds.

Image lab

Section titled “Image lab”

On CIFAR-10 or another redistributable dataset:

  1. train a small VAE, GAN, and diffusion model within stated parameter and compute ranges;
  2. preserve the same train set and duplicate policy;
  3. hold out nearest-neighbour and memorization-audit images;
  4. report reconstruction for autoencoding stages separately;
  5. compute FID or KID with the same feature extractor, sample count, and reference;
  6. add precision/recall or density/coverage, nearest-neighbour inspection, class coverage, and measured sampling time;
  7. blind and randomize any human comparison;
  8. for restoration, test the trained synthetic degradation and an unmatched real degradation.

Failure modes

Section titled “Failure modes”
  • a restoration model hallucinates diagnostically or semantically meaningful detail;
  • a GAN drops rare modes while producing sharp common samples;
  • an autoencoder discards small structures before latent diffusion starts;
  • FID changes with implementation, sample count, crop, or reference statistics;
  • CLIP-based alignment rewards text-compatible stereotypes rather than image fidelity;
  • generated or benchmark images occur in web-scale training data;
  • cherry-picked samples replace distribution-level evaluation;
  • guidance improves prompt score while reducing diversity;
  • a faster sampler inherited expensive distillation that is omitted from the cost;
  • an editor changes protected background, identity, or geometry.

Self-check and exit task

Section titled “Self-check and exit task”
  1. What information does the VAE evidence lower bound trade off?
  2. Why can a GAN have sharp samples and poor coverage?
  3. Which part of latent diffusion limits recoverable spatial detail?
  4. How do score, noise, velocity, and flow targets differ?
  5. Why can lower FID coexist with more memorization or less diversity?
  6. What must be included in a fair sampling-speed comparison?

Exit task: prepare a generative model card containing the data/license, representation or tokenizer, objective, conditioning, sampling algorithm, guidance, compute, fidelity and coverage evidence, memorization audit, failure slices, and intended/non-intended uses.

Pathology bridge

Section titled “Pathology bridge”

Restoration and generation can normalize appearance, synthesize patches, augment rare classes, or connect morphology to text—but can also invent structures and leak patients. Never use a plausible synthetic feature as clinical evidence. Evaluate downstream tasks, pathologist-relevant morphology, privacy, cohort overlap, and external scanners. Connect to WSI engineering, pathology VLMs, and evaluation and translation.

Reasoning budgets, executable outputs and scientific edits

Test-time scaling studies how extra reasoning changes perception and mathematics; EmbedCopilot checks whether plausible embedded-system answers actually execute; GRADE requires domain-correct and readable image edits. Together they motivate measuring compute, implementation success and constraint satisfaction rather than judging prose alone.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
GRADE: Benchmarking Discipline-Informed Reasoning in Image EditingMingxin Liu, Ziqian Fan, Zhaokai Wang, Leyao Gu, Zirun Zhu, Yiguo He, Yuchen Yang, Changyao Tian, Xiangyu Zhao, Ning Liao, Shaofeng Zhang, Qibing Ren, Zhihang Zhong, Xuanhe Zhou, Junchi Yan, and Xue Yang2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #2706 checked on 2026-09-12. Publisher version of record was not independently verified.GRADE evaluates whether an image editor can use disciplinary knowledge to make a scientifically correct visual change, preserve relevant context, and keep the result logically readable. Its 520 image–instruction–reference triplets span ten academic disciplines. The central distinction is between drawing a plausible image and rendering the right answer to a domain-specific problem.ECCV Spotlight program · Conference manuscript · Official project or implementation

Reading question: What external test can falsify a persuasive explanation: a perception answer, a running circuit, or a scientifically valid image?

Evidence boundary: More tokens can reduce perceptual accuracy. Judge scores and physical execution use different protocols; GRADE strict accuracy is a conjunction of three maxima, not its mean reasoning score.

Low-resolution student inference versus resolution-cascaded generation

GAP-Track removes teacher and residual reconstruction branches at inference. RMD instead retains high-resolution generative refinement. Both exploit cross-resolution supervision but solve different deployment problems.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Cross-Resolution Distribution Matching for Diffusion DistillationFeiyang Chen, Hongpeng Pan, Haonan Xu, Xinyu Duan, Zhefeng Wang, and Yang Yang2026 · ECCV · acceptedECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. RMD retains high-resolution refinement at inference; it is not low-resolution-only recognition. Reported speedups depend on the single-A100 baseline, CFG and NFE. The conference PDF author order differs from the arXiv record and is preserved here.RMD distills a diffusion generator into a few-step trajectory that begins at low resolution and ends at high resolution. It aligns resolution-specific noise states using logSNR, projects student distributions into the teacher’s resolution, and uses distribution matching plus noise reinjection to train the cascade. The output remains a high-resolution generated image or video.Official ECCV poster record

Reading question: Is high resolution absent at deployment, or only used for fewer steps? What is the full-resolution reference?

Evidence boundary: Synthetic downsampling does not reproduce all optical changes between microscope magnifications; lower-resolution gains need not equal full-resolution performance.

Masked generation and diffusion-friendly latent spectra

MIMFlow combines masked image modeling with normalizing flows; SSVAE changes the spectral structure of video latents to improve downstream diffusion. Separate tokenizer reconstruction quality, likelihood or generation objectives, and final sample quality.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image GenerationYang Chen, Xiaowei Xu, Shuai Wang, Xinwen Zhang, Qiushi Guo, Tiezheng Ge, and Limin Wang2026 · ECCV · acceptedECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #2272. Publisher proceedings/version of record was not verified on 2026-09-12.MIMFlow makes a masked visual tokenizer and its normalizing-flow prior learn together. A fixed 128-token bottleneck extracts global information from incomplete images; the invertible flow models those semantic latents, while a separate decoder reconstructs detailed pixels. The important change is where modeling capacity is spent, not simply making a larger flow.ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript
Delving into Latent Spectral Biasing of Video VAEs for Superior DiffusabilityShizhan Liu, Xinran Deng, Zhuoyi Yang, Jiayan Teng, Xiaotao Gu, and Jie Tang2026 · ECCV · acceptedECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7577. Publisher proceedings/version of record was not verified on 2026-09-12. The initial arXiv manuscript appeared in December 2025; year follows the 2026 conference.SSVAE trains video latents to be easier for a diffusion model to learn, not merely easy for a decoder to reconstruct. It promotes two separate spectral biases: a few dominant channel modes and greater low-frequency spatio-temporal energy. Latent Masked Reconstruction (LMR) and Local Correlation Regularization (LCR) implement those targets without requiring a new downstream diffusion objective.ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript · Model checkpoints

Reading question: Which representation is easier to model, and what fidelity or sampling cost is exchanged for that advantage?

Evidence boundary: Good reconstruction alone does not establish good diffusability, and normalizing flows are not interchangeable with flow matching.

Causal audiovisual generation and variable-step distillation

OmniForcing transfers a bidirectional audio-video generator to causal streaming; AnyFlow learns on-policy flow maps supporting different sampling budgets. Preserve training-stage, self-rollout, synchronization, step-count and hardware qualifiers.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
OmniForcing: Unleashing Real-time Joint Audio-Visual GenerationYaofeng Su, Yuming Li, Zeyue Xue, Jie Huang, Siming Fu, Haoran Li, Haoyang Huang, and Nan Duan2026 · ECCV · acceptedECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #9084. Publisher proceedings/version of record was not verified on 2026-09-12. Author list follows the conference manuscript and program; the current repository citation lists two additional authors. Review separates the paper's three-stage method from the later simplified two-stage code release.OmniForcing converts the 19B LTX-2 bidirectional audio–video generator into a streaming model through causal distillation. Its defining problem is not simply fewer diffusion evaluations: video and audio have different latent time scales, yet both must be generated causally and remain synchronized. The paper combines physically aligned one-second blocks, stabilized causal attention and joint self-forcing. It reports a 0.7-second first chunk and 25 FPS on one H200, while retaining explicit quality tradeoffs against the slower teacher.ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript · Author project · Model checkpoint
AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map DistillationYuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, Song Han, Han Cai, and Mike Zheng Shou2026 · ECCV · acceptedECCV 2026 Spotlight verified in the official program; scientific claims use the conference-hosted manuscript #9062. Publisher proceedings/version of record was not verified on 2026-09-12.AnyFlow distills a video generator that can use different sampling budgets without retraining. Instead of learning only a noisy-state-to-clean endpoint map, it learns transitions between two arbitrary diffusion times. A second, on-policy stage trains on the student's own trajectories using shortcut flow-map simulation, improving few-step quality while retaining useful refinement at more steps.ECCV Spotlight program · Conference manuscript · Official implementation · arXiv manuscript · Author project · Model collection

Reading question: What teacher information disappears at inference, and does quality remain stable as the generated history or step budget changes?

Evidence boundary: Real-time and any-step claims depend on resolution, duration, hardware and guidance; source-code revisions may change the training recipe.

Generative scene representations and coordinate privacy

Autoregressive Gaussian scene generation and inversion of scene-coordinate regressors expose different consequences of learned 3D representations: what a model can generate, and what its outputs reveal about a private scene.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
GaussianGPT: Towards Autoregressive 3D Gaussian Scene GenerationNicolas von Lützow, Barbara Rössle, Katharina Schmid, and Matthias Nießner2026 · ECCV · acceptedAccepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified.GaussianGPT treats a 3D Gaussian scene as a sequence of discrete spatial decisions. A sparse autoencoder compresses Gaussian attributes into quantized voxel features, and a causal Transformer alternately predicts the next occupied position and its feature code. Prefix conditioning then supports generation, completion, and outpainting with the same generative model.Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project

Reading question: Which view, geometry or coordinate information is accessible, and do appearance quality and geometric fidelity agree?

Evidence boundary: Image FID does not certify 3D coverage; coordinate-output access is stronger than a pose-only service interface.

Cross-mode distillation into explicit 3D representations

FlashWorld combines a multi-view teacher, rendered-Gaussian student and learned fake-score network. Study distribution matching, consistency regularization and when OOD batches must omit adversarial targets.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
FlashWorld: High-quality 3D Scene Generation within SecondsXinyang Li, Tengfei Wang, Zixiao Gu, Shengchuan Zhang, Chunchao Guo, Liujuan Cao2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. A frozen multi-view teacher guides a few-step rendered-Gaussian student using learned fake scores, consistency and OOD co-training. Shared-scene rendering is not guaranteed correct geometry. Preserve different benchmark timing protocols, proprietary-data and training-code access limitations. No WSI or cross-magnification evidence.FlashWorld distills a high-quality multi-view diffusion mode into a few-step generator that directly produces a 3D Gaussian scene. Its central idea is cross-mode distribution matching: transfer appearance quality from the multi-view teacher while rendering the student’s shared 3D representation during denoising.Official proceedings record · OpenReview · Official implementation

Reading question: What does changing the student representation preserve, and which visual failures escape the numerical metrics?

Evidence boundary: A shared rendered scene does not certify correct geometry, fewer parameters or cross-magnification transfer; preserve hardware and benchmark protocol differences.

Real-data correction in generative distillation

RealUID unifies matching-model distillation and routes real examples through an auxiliary matching model. Read the alternating optimization and coefficient assumptions before interpreting the one-step generator results.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)Nikita Kornilov, David Li, Tikhon Mavrin, Aleksei Leonov, Nikita Gushchin, Evgeny Burnaev, Iaroslav Koshelev, Aleksandr Korotin2026 · ICLR · peer-reviewedOfficial ICLR 2026 oral and full proceedings manuscript checked. Three-network training yields a one-step generator. No GAN discriminator does not mean no min–max optimization. Selected low-resolution image results include additional fine-tuning; teacher-error correction has restrictive assumptions. No cross-magnification or pathology evaluation.RealUID adds real-data supervision to a unified inverse-distillation formulation for matching models. A frozen teacher, a one-step generator and a learned fake matching model interact through matching losses. No GAN discriminator is added, but alternating min–max optimization remains: “No GANs” does not mean no adversarially structured optimization.Official proceedings record · OpenReview

Reading question: Where does real data enter the teacher–student system, and which extra training stages are included in the comparison?

Evidence boundary: No GAN discriminator does not mean no min–max problem; low-resolution generation does not establish pathology validity.

Generative geometry: separate pairing cost, endpoint normalization and trajectory

SOT-CFM changes optimal-transport pairing to an angular cost. Spherical Flow Matching additionally places noise and data on a common sphere and regresses geodesic tangent velocities. These interventions expose whether the gain comes from pairing, removed radial variation, or the path itself.

Primary paper · year · evidenceCore idea and why to read itOfficial resources
Geometry-Aware Image Flow MatchingJunho Lee, Kwanseok Kim, and Joonseok Lee2026 · ICML · peer-reviewedThis paper asks whether image generation can benefit from separating direction from magnitude. SOT-CFM changes the optimal-transport pairing cost to an angular distance; Spherical Flow Matching additionally projects both endpoints to a sphere and learns geodesic tangent dynamics. These are two different interventions, not two names for the same normalization trick.Final proceedings PDF · Author manuscript

Reading question: Which geometric assumption changes, and what information could normalization discard?

Evidence boundary: ImageNet results use separately tuned guidance and radius; precision need not improve with gFID. Extreme norms and domains where magnitude carries meaning require explicit checks before transfer.

Complete maintained paper route

Section titled “Complete maintained paper route”

48 canonical readings · 2014–2026· route reviewed through October 2, 2026.

This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.

Show all 48 readings in maintained order

  36. #362026arXivPreprintL3

    On-Policy Self-Distillation in Diffusion Models

    Wei Zhou, Xiongwei Zhu, Lingdong Kong, Bo Chen, Lei Zhang, Yongyuan Liang, Xiaoxia Hou, Ye Tian, Xian Sun, Yingshuo Wang, Linfeng Li, Shengqiong Wu, Leigang Qu, Feng Li, Wei Liu, Julian McAuley, and Tat-Seng Chua

    arXiv v1 technical report submitted 25 August 2026; reported reward gains rely on the chosen automated evaluators and two tested backbones.

    Official artifacts: Official project · Official implementation

    Metadata checked

  39. #392026ECCVAccepted · proceedings pendingL3

    MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image Generation

    Yang Chen, Xiaowei Xu, Shuai Wang, Xinwen Zhang, Qiushi Guo, Tiezheng Ge, and Limin Wang

    ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #2272. Publisher proceedings/version of record was not verified on 2026-09-12.

    Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript

    Metadata checked

  40. #402026ECCVAccepted · proceedings pendingL3

    Delving into Latent Spectral Biasing of Video VAEs for Superior Diffusability

    Shizhan Liu, Xinran Deng, Zhuoyi Yang, Jiayan Teng, Xiaotao Gu, and Jie Tang

    ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #7577. Publisher proceedings/version of record was not verified on 2026-09-12. The initial arXiv manuscript appeared in December 2025; year follows the 2026 conference.

    Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript · Model checkpoints

    Metadata checked

  41. #412026ECCVAccepted · proceedings pendingL3

    AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

    Yuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, Song Han, Han Cai, and Mike Zheng Shou

    ECCV 2026 Spotlight verified in the official program; scientific claims use the conference-hosted manuscript #9062. Publisher proceedings/version of record was not verified on 2026-09-12.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Official implementation · arXiv manuscript · Author project · Model collection

    Metadata checked

  42. #422026ECCVAccepted · proceedings pendingL3

    OmniForcing: Unleashing Real-time Joint Audio-Visual Generation

    Yaofeng Su, Yuming Li, Zeyue Xue, Jie Huang, Siming Fu, Haoran Li, Haoyang Huang, and Nan Duan

    ECCV 2026 Oral verified in the official program; scientific claims use the conference-hosted manuscript #9084. Publisher proceedings/version of record was not verified on 2026-09-12. Author list follows the conference manuscript and program; the current repository citation lists two additional authors. Review separates the paper's three-stage method from the later simplified two-stage code release.

    Official artifacts: ECCV Oral program · Conference manuscript · Official implementation · arXiv manuscript · Author project · Model checkpoint

    Metadata checked

  43. #432026ECCVAccepted · proceedings pendingL3

    GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation

    Nicolas von Lützow, Barbara Rössle, Katharina Schmid, and Matthias Nießner

    Accepted to ECCV 2026; oral presentation verified in the official program. This record and review use the conference-hosted manuscript checked 2026-09-12; a separate Springer proceedings record was not verified.

    Official artifacts: Official ECCV oral presentation · Official ECCV paper record · arXiv record · Official code · Official project

    Metadata checked

  44. #442026ECCVAccepted · proceedings pendingL3

    GRADE: Benchmarking Discipline-Informed Reasoning in Image Editing

    Mingxin Liu, Ziqian Fan, Zhaokai Wang, Leyao Gu, Zirun Zhu, Yiguo He, Yuchen Yang, Changyao Tian, Xiangyu Zhao, Ning Liao, Shaofeng Zhang, Qibing Ren, Zhihang Zhong, Xuanhe Zhou, Junchi Yan, and Xue Yang

    ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #2706 checked on 2026-09-12. Publisher version of record was not independently verified.

    Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation

    Metadata checked

  45. #452026ECCVAccepted · proceedings pendingL3

    Cross-Resolution Distribution Matching for Diffusion Distillation

    Feiyang Chen, Hongpeng Pan, Haonan Xu, Xinyu Duan, Zhefeng Wang, and Yang Yang

    ECCV 2026 poster verified from the official linked program event. Review uses the conference-hosted manuscript checked 2026-09-12; a separate Springer version of record was not verified. RMD retains high-resolution refinement at inference; it is not low-resolution-only recognition. Reported speedups depend on the single-A100 baseline, CFG and NFE. The conference PDF author order differs from the arXiv record and is preserved here.

    Official artifacts: Official ECCV poster record

    Metadata checked

  46. #462026ICLRPeer reviewedL3

    Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)

    Nikita Kornilov, David Li, Tikhon Mavrin, Aleksei Leonov, Nikita Gushchin, Evgeny Burnaev, Iaroslav Koshelev, Aleksandr Korotin

    Official ICLR 2026 oral and full proceedings manuscript checked. Three-network training yields a one-step generator. No GAN discriminator does not mean no min–max optimization. Selected low-resolution image results include additional fine-tuning; teacher-error correction has restrictive assumptions. No cross-magnification or pathology evaluation.

    Official artifacts: Official proceedings record · OpenReview

    Metadata checked

  47. #472026ICLRPeer reviewedL3

    FlashWorld: High-quality 3D Scene Generation within Seconds

    Xinyang Li, Tengfei Wang, Zixiao Gu, Shengchuan Zhang, Chunchao Guo, Liujuan Cao

    Official ICLR 2026 oral and full proceedings manuscript checked. A frozen multi-view teacher guides a few-step rendered-Gaussian student using learned fake scores, consistency and OOD co-training. Shared-scene rendering is not guaranteed correct geometry. Preserve different benchmark timing protocols, proprietary-data and training-code access limitations. No WSI or cross-magnification evidence.

    Official artifacts: Official proceedings record · OpenReview · Official implementation

    Metadata checked