Krea 2 Turbo 2-Step Distill LoRA: Drafts in Two Steps
A community DMD2 distillation LoRA runs Krea 2 Turbo in two steps at cfg 1.0 for fast previews and drafts, with a ready ComfyUI workflow of stock nodes only.
The author of the Krea 2 Turbo 4-step distillation LoRA is back with an even more aggressive cut: Krea2-Turbo-Distill-2step-LoRA takes Krea 2 Turbo from 8 steps down to 2, about 4.2x faster denoising (81.4s down to 19.5s at 1024x1024). It is positioned as a preview and draft tool, with the 4-step LoRA still recommended for final renders.
 |  |  |
|---|---|---|
| Turbo 8 steps | Turbo 2 steps: hollow and soft | 2 steps with the LoRA: detail restored |
Distribution matching, not imitation
Like its 4-step sibling, the adapter is a plain rank-64 LoRA trained with distribution matching (DMD2): the student learns to match what the teacher would plausibly produce rather than copying its exact trajectory, which keeps two steps from averaging into blur and doubled edges. The student makes two calls at sigmas 1.0 and 0.7595, the first and fifth points of the teacher's 8-step grid, with stock Euler in between.
The current checkpoint (chk00017464, still marked as training in progress) targets what earlier versions left behind: grain and grid patterns at large sizes, small faces, and dense detail. Fine-texture energy measures at 1.01x to 1.17x the 8-step teacher across the 12 trained resolutions, and a blind prompt-adherence rubric scores 239 of 240 points. Training runs on a single RTX 3090.
ComfyUI usage
The repo ships krea2_turbo_2step_rank_64_lora_comfyui.safetensors under ComfyUI key names plus krea2_turbo_2step_lora_t2i.json, a ready workflow built from stock nodes only. File names never change: newer checkpoints replace both files in place.
The bundled workflow: Krea 2 Turbo plus the 2-step LoRA, stock nodes only.
Settings: steps 2, cfg 1.0, sampler euler / simple, LoRA strength 1.0. One trap worth repeating: ComfyUI expresses "no CFG" as 1.0, not 0.0 as in diffusers. Setting 0.0 in ComfyUI is a different thing and will not give the trained behavior.
Availability
- Hugging Face: lvladikov/Krea2-Turbo-Distill-2step-LoRA, with the full training card in DETAILED-README.md
- The 4-step LoRA remains the quality recommendation; the 2-step adapter is for previews and drafts
Comments
Sign in with GitHub to join the discussion.