Model watch: which model should Kokon run next?

[GabrieleLozupone]

Kokon's idea is simple: one plug-and-play editor, and we pick the best model for you. You shouldn't have to compare checkpoints, samplers and quantizations to get a good image. So Kokon runs one model at a time, chosen by us, and the choice is made in the open here.

This thread is where we keep track of new models and where you can propose one.

What the model has to do

Kokon is not a showcase for one task, so a candidate has to be good at all of these, on hardware people actually have:

1. Text to image: prompt following, text rendering, photos and illustrations.
2. Editing an image: change what you ask for and keep everything else (faces, layout, light, details).
3. Editing with several images: combine a person, a product, a background or a style from 2–3 references.
4. Brush edits, expand, remove: the inpaint and outpaint actions in the app.
5. Resources: VRAM / unified memory, time per image on a 12–16 GB GPU and on a Mac, download size. We count the whole memory budget, not just the diffusion model.
6. License: weights we can download and run locally for our users.

A model that is great at one thing (only text to image, only upscaling, only faces) doesn't fit Kokon's default, however good it is.

Why Qwen-Image 2.1 today

• Strong on all four tasks above, with the official 8-step Turbo for speed.
• Its text encoder is Qwen3-VL, a full vision-language model. Kokon reuses it to write prompts (Prompt+) at no extra download and almost no extra memory. Most other models don't have that, and an extra LLM for the same job would cost several GB.
• Runs in Kokon from 12 GB GPUs (8 GB in beta) and on Apple Silicon.

How a model would replace it

• We run it on our own test set: text to image, single edits, multi-image edits, brush edits, on Windows (CUDA and Vulkan), Mac and Linux, and measure memory and time.
• If it beats Qwen-Image 2.1 on everything that matters overall, it becomes the new default in a release.
• You keep a way back: rollback to the previous model and a switch in Settings for anyone who prefers the old one.
• It also needs to run in stable-diffusion.cpp (Kokon's engine), or be close enough that we can add it.

Proposing a model

Reply here with:

• Name and link (weights + license)
• Which of the 4 tasks it's better at than Qwen-Image 2.1, ideally with the same prompts / images side by side
• Memory and time you measured, and on what hardware
• Does it run in sd.cpp today?

We'll post updates here when we test something (results, not just impressions). Models suggested so far: Flux 3, Moss, Fibo, MiniMax H3 / FastH3 Omni (#34), Krea 2 (Reddit).