Generative Vision Atlas

Which model does what

The question most people arrive with is whether a given model can both generate and edit, and what it is working on top of. This table answers both. The basis column is the space the model actually operates in, which is what determines its failure modes.

10 of 40 systems here do both generation and editing. 6 also do image understanding in the same model.

Almost all of them sit on a VAE latent. The exception worth knowing is HiDream-O1-Image, which drops the autoencoder entirely and embeds raw pixel patches, text and condition tokens into one shared transformer, so editing and generation are literally the same forward pass.

ModelGenerateEditMulti-refUnderstandTokenizeBuilt onAvailabilityDate
LLaDA-Image··VAE latentopen weights2026-09
UniSpace··not recordedunclear2026-08
HiDream-O1-Image··Pixel spaceopen weights2026-05
Qwen-Image 2.0···VAE latentunclear2026-05
RAEv2····Representation latentcode only2026-05
Drifting Models····not recordedcode only2026-02
Scale-RAE···Representation latentopen weights2026-01
SVG-T2I···Representation latentopen weights2025-12
JiT····Pixel spaceopen weights2025-11
Emu3.5··not recordedopen weights2025-10
RAE····Representation latentopen weights2025-10
SVG····Representation latentopen weights2025-10
HunyuanImage 3.0···not recordedopen weights2025-09
Seedream 4.0··VAE latentAPI only2025-09
NextStep-1····Continuous tokensopen weights2025-08
Qwen-Image···VAE latentopen weights2025-08
FLUX.1 Kontext··not recordedopen weights2025-06
OmniGen2··not recordedopen weights2025-06
Show-o2···not recordedopen weights2025-06
STARFlow····not recordedunclear2025-06
BAGEL··not recordedopen weights2025-05
BLIP3-o···not recordedopen weights2025-05
HiDream-I1····VAE latentopen weights2025-05
MeanFlow····not recordedcode only2025-05
ICEdit····not recordedopen weights2025-04
MetaQuery···not recordedcode only2025-04
Step1X-Edit····not recordedopen weights2025-04
Lumina-Image 2.0····VAE latentopen weights2025-03
ACE++····not recordedopen weights2025-01
Janus-Pro···not recordedopen weights2025-01
MAR····Continuous tokensopen weights2024-06
VAR····Discrete tokensopen weights2024-04
SD3 / MMDiT····VAE latentopen weights2024-03
SiT····VAE latentopen weights2024-01
Emu2···Representation latentopen weights2023-12
Emu Edit····not recordedclosed2023-11
DiT····VAE latentopen weights2022-12
InstructPix2Pix····not recordedopen weights2022-11
unCLIP / DALL-E 2····Representation latentclosed2022-04
LDM / Stable Diffusion····VAE latentopen weights2021-12

A blank basis means the paper is a method or analysis contribution rather than a system built on one particular latent. Medical models are listed separately under medical imaging.