Timeline

Every paper in the general-domain atlas by date, with the line it belongs to. Medical imaging has its own separate timeline of ideas. The point of reading it in order is to see how long an idea sits unused before something makes it work: generating in a semantic space was tried in 2022 and abandoned, then became the most active question in the field three years later.

2021

Before the shift. Latent diffusion and masked autoencoders arrive separately; nobody yet connects them.

CLIPlandmark · Contrastive language-image pretraining

CDMlandmark · Cascaded and multiscale pixel diffusion

MAElandmark · Self-distillation representations

LDM / Stable Diffusionlandmark · VAE-latent diffusion

2022

Latent diffusion becomes the default, and CLIP-space generation is tried once and set aside.

Flamingolandmark · Cross-attention into a frozen language model

unCLIP / DALL-E 2landmark · Representation-space generation

Imagenlandmark

Classifier-Free Guidancelandmark · Guidance and sampling

Prompt-to-Promptlandmark · Training-free attention manipulation

Re-Imagenlandmark · Agentic and search-augmented generation

Rectified Flowlandmark · Flow matching and rectified flow

Flow Matchinglandmark · Flow matching and rectified flow

InstructPix2Pixlandmark · Editing in the VAE latent

Plug-and-Play · Training-free attention manipulation

DiTlandmark · VAE-latent diffusion

2023

The backbone becomes a transformer, and self-supervised features get good enough to matter.

BLIP-2landmark · Cross-attention into a frozen language model

I-JEPA · Self-distillation representations

simple diffusion · Cascaded and multiscale pixel diffusion

ControlNetlandmark · Adapter-based conditioning

T2I-Adapter · Adapter-based conditioning

SigLIP · Contrastive language-image pretraining

DINOv2landmark · Self-distillation representations

LLaVAlandmark · Encoder plus projector

MasaCtrl · Training-free attention manipulation

Otter · Cross-attention into a frozen language model

MagicBrush · Editing in the VAE latent

IP-Adapterlandmark · Adapter-based conditioning

ViT Registers · Self-distillation representations

Idea2Img · Agentic and search-augmented generation

Emu Edit · Editing in the VAE latent

Emu2 · Representation-space generation, Unified understanding and generation

GIVT · Continuous-token autoregression

AM-RADIO · Agglomerative multi-teacher distillation

2024

Flow matching wins the objective argument. Autoregression mounts a serious challenge. Alignment losses appear.

HDiT · Cascaded and multiscale pixel diffusion

InstantID · Adapter-based conditioning

SiT · Flow matching and rectified flow, VAE-latent diffusion

SD3 / MMDiTlandmark · Flow matching and rectified flow, VAE-latent diffusion

Guidance interval · Guidance and sampling

PuLID · Adapter-based conditioning

VARlandmark · Discrete-token autoregression

Chameleonlandmark · Unified understanding and generation

Autoguidancelandmark · Guidance and sampling

CFG++ · Guidance and sampling

MARlandmark · Continuous-token autoregression

GenArtist · Agentic and search-augmented generation

Theia · Agglomerative multi-teacher distillation

LLaVA-OneVision · Encoder plus projector

Transfusionlandmark · Continuous-token autoregression, Unified understanding and generation

UNIC · Agglomerative multi-teacher distillation

Emu3 · Unified understanding and generation

NVLM · Cross-attention into a frozen language model

Qwen2-VL · Native-resolution vision-language models

Adaptive Projected Guidance · Guidance and sampling

DC-AE · VAE-latent diffusion

Fluidlandmark · Continuous-token autoregression

Janus · Unified understanding and generation

REPAlandmark · Representation-aligned latents

RF-Inversion · Inversion for flow models

Shortcut Models · Natively few-step objectives

SiD2landmark · Cascaded and multiscale pixel diffusion

AIMv2 · Contrastive language-image pretraining

Edify Image · Cascaded and multiscale pixel diffusion

OminiControllandmark · Adapter-based conditioning, In-context editing

RF-Solver · Inversion for flow models

Stable Flow · Training-free attention manipulation

FireFlow · Inversion for flow models

RADIOv2.5 · Agglomerative multi-teacher distillation

TarFlowlandmark · Normalizing-flow revival

TokenFlow · Semantic-plus-detail hybrids

2025

The latent itself becomes the contested question: align it, replace it, or delete it.

ACE++ · In-context editing

Janus-Pro · Unified understanding and generation

VA-VAE / LightningDiT · Representation-aligned latents

KV-Edit · Training-free attention manipulation

MAETok · Representation-aligned latents

Qwen2.5-VL · Native-resolution vision-language models

SigLIP 2 · Contrastive language-image pretraining

ICE-Bench · Evaluation and benchmarks

IMM · Natively few-step objectives

Lumina-Image 2.0 · VAE-latent diffusion

SANA-Sprint · Natively few-step objectives

GigaTok · Discrete-token autoregression

ICEdit · In-context editing

InternVL3 · Native-resolution vision-language models

MetaQuery · Encoder plus projector, Unified understanding and generation

Perception Encoder · Contrastive language-image pretraining

PixelFlow · Cascaded and multiscale pixel diffusion, Flow matching and rectified flow

REPA-E · Representation-aligned latents

RISEBench · Evaluation and benchmarks

Step1X-Edit · Editing in the VAE latent, In-context editing

Web-SSL · Self-distillation representations

BAGEL · Editing inside a unified model, Unified understanding and generation

BLIP3-o · Unified understanding and generation

Flow-GRPO · Reinforcement learning and preference alignment

HiDream-I1 · VAE-latent diffusion

ImgEdit · Evaluation and benchmarks

KRIS-Bench · Evaluation and benchmarks

MeanFlowlandmark · Natively few-step objectives

UniEval · Evaluation and benchmarks

FLUX.1 Kontextlandmark · Editing in the VAE latent, In-context editing

OmniGen2 · Editing inside a unified model, In-context editing, Unified understanding and generation

RefEdit

Show-o2 · Unified understanding and generation

STARFlow · Normalizing-flow revival

Transition Matching · Transition matching

V-JEPA 2 · Self-distillation representations

PixNerdlandmark · Single-stage pixel transformers

REG · Representation-aligned latents

Resurrect MAR · Continuous-token autoregression

DINOv3 · Self-distillation representations

NextStep-1 · Continuous-token autoregression

Qwen-Image · VAE-latent diffusion

DiffusionNFT · Reinforcement learning and preference alignment

EditScore · Reinforcement learning and preference alignment

GLASS Flows · Guidance and sampling

HunyuanImage 3.0 · Unified understanding and generation

Seedream 4.0 · In-context editing, VAE-latent diffusion

Tokenizer Post-Training · Discrete-token autoregression

Emu3.5 · Editing inside a unified model, Unified understanding and generation

Neon · Reinforcement learning and preference alignment

There is No VAE · Single-stage pixel transformers

Pico-Banana-400K

RAElandmark · Representation-space generation

Rectified-CFG++ · Guidance and sampling

SVG · Representation-space generation

Demystifying TM · Transition matching

VFM-VAE · Representation-space generation

DiP · Single-stage pixel transformers

JiT · Single-stage pixel transformers

PixelDiT · Single-stage pixel transformers

Qwen3-VL · Native-resolution vision-language models

GenEval 2 · Evaluation and benchmarks

PS-VAE · Editing in a representation latent

REPA spatial-structure study · Representation-aligned latents

SVG-T2I · Representation-space generation

TM Design Space · Transition matching

2026

Consolidation and new paradigms. Hybrids soften the VAE-versus-representation split while one-step objectives attack from a different direction.

C-RADIOv4 · Agglomerative multi-teacher distillation

DiDAE · Editing in a representation latent

Pixel MeanFlow · Natively few-step objectives, Single-stage pixel transformers

Scale-RAE · Representation-space generation

TM Distillation · Transition matching

Drifting Modelslandmark · Natively few-step objectives

FlatDINO · Semantic-plus-detail hybrids

Latent Forcing · Semantic-plus-detail hybrids

LV-RAE · Semantic-plus-detail hybrids

Masked Bit Modeling · Discrete-token autoregression

PixelREPA · Single-stage pixel transformers, Representation-space generation

RPiAE · Editing in a representation latent

DecQ · Semantic-plus-detail hybrids

HiDream-O1-Image · Editing inside a unified model, Single-stage pixel transformers

HyperDiT · Single-stage pixel transformers

PAE · Semantic-plus-detail hybrids

Qwen-Image 2.0 · VAE-latent diffusion

RAEv2 · Representation-space generation

STARFlow2 · Normalizing-flow revival, Unified understanding and generation

CrossFlow · Semantic-plus-detail hybrids, Natively few-step objectives

Distilling Drifting Transformers · Natively few-step objectives, Representation-space generation

Latent Diffusability · Semantic-plus-detail hybrids

Parallel Rollout Approximation · Single-stage pixel transformers

Chimera · VAE-latent diffusion

dRAE · Representation-space generation, Discrete-token autoregression

Parallel Decoding Distillation · Natively few-step objectives

Pixel-space survey · Single-stage pixel transformers

SearchGen · Agentic and search-augmented generation

Self-Sample Guidance · Single-stage pixel transformers, Guidance and sampling

Three-Body Scattering

Any-OPD · Natively few-step objectives

ContextBias · Evaluation and benchmarks

GenFirst · Semantic-plus-detail hybrids

MOSAIK · Single-stage pixel transformers

Observation Operators · Single-stage pixel transformers

Pixel T2I empirical study · Single-stage pixel transformers

Second-Order Drifting · Natively few-step objectives

UniSpace · Editing in a representation latent

Agentic Visual Generation · Agentic and search-augmented generation

LLaDA-Image · VAE-latent diffusion

PixSGR · Single-stage pixel transformers