KreaMossMix Workflow: One model composes, the other finishes it
retained text
I've been using MOSS-Image and Krea 2 a lot lately, and they're good at different things, so I built a ComfyUI workflow that runs both from the same prompt and lets them work on each other's images. Every run can give you four results: * **Full MOSS**: MOSS-Image on its own * **Full Krea**: Krea 2 on its own * **Mix A**: MOSS builds the composition, Krea finishes it * **Mix B**: Krea makes the image, MOSS repaints part of it The images in this post are one prompt and one seed through all four paths, with no LoRAs, no inpainting and no editing. **What's inside** * One slider (FINISH SHARE) controls how much of the image the second model repaints in both mixes * Krea's "Raw recipe": Krea 2 Raw + the official Turbo LoRA at 0.6 with er\_sde, which gives a more natural, less polished look than plain Turbo. One switch flips back to official Turbo * Optional HIRES pass on every image. Krea images get decoded at 2x by spacepxl's Wan 2.1 upscale VAE and refined; MOSS images get upscaled 1.4x and refined * Krea images are sharper even with HIRES off, thanks to the 2x upscale VAE * MOSS Base or SFT, plus a Turbo mode with the 8-step DMD LoRA for fast drafts * Toggles only compute what's new: turn on MIX or HIRES after a run and the existing images are reused * Empty LoRA loaders for each model, so you can plug in your own (a character LoRA trained on both models gives you the same person in all four images) * The engine is packed into one subgraph, so the main canvas is just the controls. A How to use note covers every model file with download links, every switch, and prompting tips **Settings for these samples:** 9:16 at 2 MP, FINISH SHARE 0.5, MOSS Base in quality mode (30 steps, CFG 3), Krea Raw recipe, HIRES on. **Needs:** ComfyUI 0.43+, rgthree-comfy and ComfyUI-VAE-Utils. Everything is local and free. [Download KreaMossMix Workflow](https://civitai.red/models/3027247/kreamossmix-workflow?modelVersionId=3438272) A tip on prompts: both models do well with longer prompts in plain sentences that describe the whole composition (where the subject is in the frame, what's around her, where the light comes from) instead of tag lists. The sample prompt is in the comments. This is my first published workflow, so feedback is welcome, especially if you find settings that work better than mine.
document fetched 2026-10-31 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xj1xxe/kreamossmix_workflow_one_model_composes_the_other