It's the model created by lodestones. I only converted it, using int8-convrot quantization.
This DiT seems to like euler - simple pair for sampler - scheduler. Or rather simple as a scheduler (denoiser for the appearing image) CFG can go up to 1.5.
The model is VERY responsive to the prompt.
All credits go to the creators of this DiT and to the Krea2 team!
Description
After the lodestones' huggingface page:
What OPD is: v0.3.1 was distilled on-policy. Instead of the classic offline recipe — imitating the teacher on a fixed sampling schedule, which slowly pulls the student off the original data manifold — the student generates its own trajectories and the teacher corrects it on those exact points. Training only ever happens on states the model actually visits, so the distilled model stays inside the original model's distribution: Turbo speed without the usual distillation tax (mode collapse, washed-out detail, prompts that suddenly stop working).
EDIT: I added the Cyberdelia's-node converted file as well, which seems bigger in size. The 'usual', smaller file was converted using an extension in Forge Neo UI.