MOSS-Image and Krea 2 in one picture: one model draws, the other repaints half

Partly worked2026-10-30 · MOSS-Image SFT, Krea 2 Turbo · RTX 5070 12GB

Six standard prompts, each made four ways: MOSS-Image alone, MOSS repainted by Krea 2, Krea 2 alone, Krea 2 repainted by MOSS — Seed 300, 1024×1024, one rented RTX 5070 12GB. Columns: MOSS alone, Mix A (MOSS, then Krea repaints half), Krea alone, Mix B (Krea, then MOSS repaints half). Rows: portrait, landscape, anime, text, busy scene, long four-paragraph prompt.

A ComfyUI workflow published today, KreaMossMix by NicoGJ, runs one prompt through two models we already have on the site, MOSS-Image and Krea 2, and lets one model finish what the other started. We tried a simpler version of the idea on our standard prompts, on one 12 GB card.

What we ran

Both models on one RTX 5070 12GB (48 GB of system memory), the same prompt and seed for both:

MOSS alone: MOSS-Image SFT (int8), 25 steps, er_sde, cfg 2.5. Our standard MOSS settings.
Krea alone: Krea 2 Turbo (fp8), 8 steps, cfg 1. Our standard Krea settings.
Mix A: the finished MOSS picture, repainted by Krea 2 Turbo at denoise 0.5.
Mix B: the finished Krea picture, repainted by MOSS at denoise 0.5.

This is not the original workflow. The original hands over a half-finished image in the middle of sampling and adds an upscale pass. We handed over the finished picture (image-to-image), because it is easy to repeat. The five standard prompts plus one long prompt of our own, written in the four-paragraph style the workflow's notes recommend. One seed.

What we saw (judged by eye)

Mix A keeps MOSS's layout and adds Krea's texture. The fisherman keeps the same pose and cap, but his skin gets fine wrinkles and the beard gets single hairs. The market and landscape get stronger light and contrast. This is the more useful direction.
Mix A also changes small things. The chalkboard text beside the bakery and some faces in the crowd were redrawn. Check text and faces after a mix.
Mix B changes very little. At denoise 0.5, MOSS mostly kept the Krea picture: skin a little smoother and warmer. In the flower-stall scene the ring on her hand disappeared.
Text survived in all four columns. "OPEN" is spelled right everywhere.

Close crops: the fisherman and the flower-stall woman, MOSS alone and Mix A on the top rows, Krea alone and Mix B on the bottom rows — Full-resolution crops. Left: portrait. Right: long prompt. In each, top row = MOSS alone, Mix A; bottom row = Krea alone, Mix B.

Time

Server time per picture on the same node, after the first picture of each kind: MOSS about 15.4 s, Krea about 11.7 s, the Krea repaint about 11.8 s and the MOSS repaint about 15.6 s. A repaint at denoise 0.5 costs as much as a full run, because ComfyUI still runs every step on the second half of the schedule. So Mix A is about 27 s of GPU time per picture, roughly twice a single model. One node, one run; these numbers are not in the Results table.

How much to trust this

One seed, one card, judged by eye. Denoise 0.5 only; we did not try other strengths or the original workflow's hand-over in the middle of sampling.
Our first two attempts failed before making pictures (a timeout on a model switch, then a node that did not start in 100 minutes). The third attempt made all 24.
All images passed an automatic nudity check (NudeNet) before publishing.