Introduction

Need a two-step peek at Krea 2 Turbo before you commit to a full eight-step pass? lvladikov just dropped an experimental 2-step distillation LoRA that takes Turbo from its usual 8 steps down to 2 — half the cost of the author’s 4-step adapter, a quarter of the teacher’s denoising. Think golden-hour drafts: fast, usable, honest about what two steps can and can’t hold.

Workflow download:

krea2_turbo_2step_lora_t2i.jsonJSON · 10.56 KB

Cards to keep open: the 2-step LoRA on Hugging Face and the base krea/Krea-2-Turbo.

What it is

This is a rank-64 LoRA on the same 228 modules as the author’s 4-step adapter. Sampling stays drop-in: stock Euler, Turbo’s own two sigmas [1.0, 0.5128], guidance 0.0 in diffusers / cfg 1.0 in ComfyUI. No custom sampler, no policy head — if it needed a special sampler, it wouldn’t be this project.

The published checkpoint is alpha / experimental (around 13663 training samples in the 2-step stages). Files update in place when a better checkpoint lands; the 4-step LoRA remains the author’s quality recommendation.

Why creators should care

  • ~3.8× faster denoising vs 8-step on the author’s Apple Silicon / MLX bench at 1024² — about 79.9 s → 20.9 s (author’s numbers; end-to-end on your box will sit below that as prompt encode / VAE stay fixed)
  • Half the cost of the 4-step LoRA, quarter of the 8-step teacher — built for previews and drafts
  • Diffusers + ComfyUI key variants, plus a ready stock-node ComfyUI workflow JSON
  • Trained across 12 multi-aspect resolutions from 512² to 1440²
  • Method pivot: trajectory distillation stalled at 2 steps (averaging → blur / doubled edges); distribution matching (DMD2-family) got pictures moving again

Files on the card

Grab weights from Hugging Face (external links stay on HF):

Hosted here for one-click load into ComfyUI:

krea2_turbo_2step_lora_t2i.jsonJSON · 10.56 KB

Usage cheat sheet

diffusers: steps=2, guidance_scale=0.0, LoRA scale 1.0ComfyUI: steps 2, cfg 1.0, euler / simple, LoRA strength 1.0sigmas: Turbo’s [1.0, 0.5128] — pipeline / workflow apply them

In ComfyUI, cfg 1.0 means “no CFG” (one forward per step). Diffusers says the same thing as guidance_scale=0.0. Don’t set cfg to 0.0 in Comfy — that isn’t Turbo’s intended path.

Pro Tip

Reach for steps before strength. This adapter is trained at 2; when quality matters more than speed, the author’s 4-step LoRA is still the better tool. Keep strength at 1.0 unless you’re deliberately chasing a denser look (1.0–1.5 is extrapolation).

Known issues (honest)

Two steps give up more than four. Expect, in rough order of how often they show:

  • Soft fine structure (feathers, hair, signage) — most at 1280² and up
  • Faint doubled contours on faces and limbs
  • Small / distant faces and crowd scenes can lose features or pick up duplicates
  • Composition can differ from the teacher at the same seed — shorter path, different settle, not just a blurrier twin

Treat the 8-step teacher as a quality reference, not a pixel target for the 2-step path.

Conclusion

If your loop is “prompt → peek → commit,” this 2-step distill LoRA is a sunny little turbo for Krea 2 drafts — drop-in Euler, hosted workflow below, weights on Hugging Face. Keep the 4-step adapter for the keepers; let this one burn through the maybe pile.

Workflow download:

krea2_turbo_2step_lora_t2i.jsonJSON · 10.56 KB

—Aurelia ♡