← sources

Krea 2 Turbo Distill - 3 step LoRA

source: reddit (reddit:StableDiffusion)published 2026-10-29 15:06ingested 2026-10-30 02:38

retained text

# Krea 2 Turbo Distill - 3 step LoRA (work in progress)

The **third** of my step-distillation LoRAs for Krea 2 Turbo, picking up where the other two left off. The [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) took Turbo from 8 steps to 4 and the [2-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA) went down to 2 for fast previews but, despite best efforts, never quite reached the 4-step's quality.

**At one step fewer than the 4-step and one more than the 2-step, this one renders more fine detail than either, trained on a single cloud H100 with everything the 2-step project taught me.**

**Three steps instead of eight · 2.6× faster denoising (text encoding and VAE decode excluded) · fine detail at 1.15× the 8-step teacher's on the 22 test prompts, above the** [**4-step LoRA**](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)**'s 1.04× (stock Turbo at 3 steps: 0.58×) · colour at the teacher's level · 12,000 training samples on recorded teacher trajectories · one H100, about 41 hours · released from settled weights.**

A LoRA for [**Krea 2 Turbo**](https://huggingface.co/krea/Krea-2-Turbo) that takes the model from **8 steps down to 3** — Turbo's own weights and its own 3-step sigmas, CFG 1 (no guidance; diffusers writes it as `guidance_scale=0.0`) — aiming at the best quality three steps can give. It keeps — and passes — the 4-step LoRA's fine detail at three quarters of its denoising time, with the picture a little further from the teacher's own.

* ⚡ **8 → 3 steps** — on Turbo's own 3-step sigmas `[1.0, 0.8633, 0.6123]`
* ⏱️ **2.6× faster denoising** — 65.4 s → 24.9 s at 1024×1024 (float16 compute), text encoding and VAE decode excluded, as they take the same time at any step count; the adapter's own cost per call is within about 1%
* 🎯 **Fine detail above the teacher's** — fine texture **1.15×** and the 16- and 8-pixel bands **1.17× / 1.16×** the teacher's on the 22 test prompts (stock Turbo at 3 steps: **0.58×**; the 4-step adapter **1.04×**), colour at the teacher's level
* 📊 **Distribution matching, not imitation** — matches what the teacher would plausibly produce rather than its exact trajectory, so the student commits instead of averaging into blur and doubled edges
* 🔎 **Eight critics** — artefacts, real photographs, the first call's layout, content, prompt agreement, lettering, large faces and the whole three-call rollout, each a small head on the frozen base's own mid-network features
* 📐 **11 trained resolutions** — from 512×768 up to 1440×1440, and 1824×1024
* 🔌 **Drop-in, no exceptions** — plain LoRA, stock Euler, diffusers / ComfyUI / MLX. No custom nodes, no custom sampler (the LoRA needs none; the optional [MLX workflow for Apple Silicon](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#on-apple-silicon-over-mlx) below runs it through my [LVNodes](https://huggingface.co/lvladikov/ComfyUI-Nodes-and-Workflows/blob/main/LVNodes/README.md))
* 🧬 **Same shape as the** [**4-step**](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) **and** [**2-step**](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA) **adapters** — rank 64 on the same 228 modules
* 🎲 **About 30,000 recorded teacher trajectories** — Turbo's own 8-step paths at the 11 training sizes; every sample outside the two anneals is one the adapter had not seen
* 🔢 **12,000 training samples**, starting from the released 4-step adapter's weights
* 🧘 **Settled weights, not an average** — the last 600 samples take the learning rate to zero over already-trained data, so the published weights are the training weights at rest

ComfyUI workflows provided including an [MLX](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#on-apple-silicon-over-mlx) one in addition to the default Torch one. Details for ComfyUI here: [https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#comfyui](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#comfyui)

***Using it on Raw***

Trained on Turbo, for Turbo — but it also loads on **Krea 2 Raw**, which shares the architecture. On Raw at 13 steps with CFG 3.5, it keeps Raw's own composition and adds the detail Raw lacks at that step count: skin and freckles, fur, falling snow, a lit and signed night street. At 20 steps the pictures stay the same, with sharper lettering and fine edges. Keep Raw's CFG on: at cfg 1 or 0, Raw falls apart. It is neither trained nor tuned for Raw, so treat it as a bonus rather than a supported mode. Details on RAW (included sample images) here: [https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#using-it-on-raw](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA#using-it-on-raw)

So yes, this is another comeback to Krea 2, and a good one I think. You should be able to get more (quality) with less (steps, and effectively time). It is still work in progress so let me know your feedback. It is one very expensive project since I trained all *(including a lot of early lengthy experimentations and approaches I had cancelled, until I settled on working recipe)* on a **cloud based H100** (all critics full precision all loaded at the same time, something I couldn't do before on my RTX 3090 where I trained 2 and 4 step LoRAs). I hope you all appreciate it.

Full Details and to download, visit my 3 Step LoRA project at HF: [https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA)

document fetched 2026-10-30 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1ximik9/krea_2_turbo_distill_3_step_lora