black-forest-labs/FLUX.2-dev

Image-to-Image • 32B • Updated Feb 17 • 891k • • 2.05k

Note Current default for text-to-image quality. Heavy — plan on offloading the text encoder on a 24GB card.

black-forest-labs/FLUX.1-dev

Text-to-Image • 12B • Updated Jun 27, 2025 • 595k • • 14.2k

Note Still the most-forked base in the ecosystem, so it has by far the deepest LoRA/ControlNet support. Pick it when you need the accessories, not the raw quality.

Comfy-Org/flux1-dev

Updated 3 days ago • 373k • 654

Note The ComfyUI-repackaged single-file weights. Saves you assembling the components by hand.

black-forest-labs/FLUX.1-schnell

Text-to-Image • 12B • Updated Aug 16, 2024 • 375k • • 5.57k

Note Few-step, Apache-2.0. What to use for prompt iteration before committing to a slow high-res run.

comfyanonymous/flux_text_encoders

Updated Nov 20, 2024 • 1.39k

Note The CLIP/T5 encoders Flux needs as separate loads. The fp8 T5 is the difference between fitting and not.

Qwen/Qwen-Image-Edit-2511

Image-to-Image • 20B • Updated Dec 23, 2025 • 242k • • 1.28k

Note The one to reach for when you want a real instruction edit — 'change the jacket to red' — rather than an img2img reinterpretation. This is the distinction most people get wrong.

Qwen/Qwen-Image-Edit-2509

Image-to-Image • 20B • Updated Sep 22, 2025 • 462k • • 1.23k

Note Previous edit revision. Worth keeping for comparison; behaviour on multi-image conditioning differs.

Qwen/Qwen-Image-Edit

Image-to-Image • 20B • Updated Aug 25, 2025 • 117k • • 2.48k

Note Original edit release, for reference on how fast this line moved.

Qwen/Qwen-Image

Text-to-Image • 20B • Updated Aug 18, 2025 • 238k • • 2.58k

Note The text-to-image base. Best-in-class text rendering inside images, which is still Flux's weak spot.

Comfy-Org/Qwen-Image-Edit_ComfyUI

Updated 3 days ago • 1.21M • 458

Note ComfyUI-ready packaging for the edit models. Start here rather than the diffusers repos.

Comfy-Org/Qwen-Image_ComfyUI

Updated 3 days ago • 1.81M • 470

Note ComfyUI-ready Qwen-Image weights, split by component.

stabilityai/stable-diffusion-xl-base-1.0

Text-to-Image • 3B • Updated Oct 30, 2023 • 1.56M • • 8.05k

Note The ancestor of the whole fine-tune ecosystem below. Keep it for LoRA compatibility testing.

purplesmartai/pony-v7-base

Text-to-Image • 7B • Updated Oct 26, 2025 • 40.5k • 101

Note Pony V7. Tag-driven (score_9 / rating_* prefixes) rather than natural-language — prompt it like a booru query, not like Flux. Uses a Qwen-family text encoder, which surprises people migrating from V6.

AstraliteHeart/pony-diffusion-v6

Updated Jan 23, 2024 • 70

Note Pony V6, the SDXL-based predecessor. Vastly more community LoRAs exist for it than for V7.

xinsir/controlnet-union-sdxl-1.0

Text-to-Image • 1B • Updated Jul 30, 2024 • 112k • 1.81k

Note One ControlNet covering most conditioning modes for SDXL. Replaces a folder of single-purpose ControlNets.

madebyollin/taesdxl

2.45M • Updated Nov 7, 2024 • 3.44k • 64

Note Tiny VAE for fast previews. Cheap quality-of-life win on long batch runs — decode previews without paying full VAE cost.