base_model: krea/Krea-2-Turbo base_model_relation: adapter license: other license_name: krea-2-community-license license_link: https://huggingface.co/lvladikov/Krea2-Turbo-Distill-3step-LoRA/blob/main/LICENSE.pdf library_name: diffusers tags:

lora
text-to-image
distillation
step-distillation
distribution-matching
krea-2 pipeline_tag: text-to-image

Krea 2 Turbo — 3-Step Distillation LoRA

The third of my step-distillation LoRAs for Krea 2 Turbo, picking up where the other two left off. The 4-step LoRA took Turbo from 8 steps to 4 and is still the closest to Turbo's own 8-step pictures; the 2-step 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/H200 with everything the 2-step project taught me.

🧪 A project still in training. This checkpoint is shared now because it is already good enough to use; training goes on, and further updates and LoRAs will likely be published.

Three steps instead of eight · 2.6× faster denoising (text encoding and VAE decode excluded) · fine detail at 1.23× the 8-step teacher's on the 22 test prompts, above the 4-step LoRA's 1.04× (stock Turbo at 3 steps: 0.58×) · colour at the teacher's level · 42,000 training samples on recorded teacher trajectories · one H100/H200, about 115 hours · released from settled weights.

A LoRA for 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 1.23× and the 16- and 8-pixel bands 1.29× / 1.24× 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 below runs it through my LVNodes)
🧬 Same shape as the 4-step and 2-step 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 seven anneals is one the adapter had not seen
🔢 42,000 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

The 22 test prompts, rendered by Krea 2 Turbo with this LoRA at 3 steps

Files

file	what it is
`krea2_turbo_3step_rank_64_lora.safetensors`	LoRA in diffusers key format — see diffusers
`krea2_turbo_3step_rank_64_lora_comfyui.safetensors`	Same weights under ComfyUI key names — see ComfyUI
`krea2_turbo_3step_lora_t2i.json`	Ready ComfyUI workflow, stock nodes only
`krea2_turbo_3step_lora_t2i_MLX_LVNodes.json`	The same workflow on Apple Silicon over MLX, through the LVNodes Krea 2 MLX custom node — see On Apple Silicon, over MLX
`krea2_turbo_3step_rank_64_lora_checkpoint_info.md`	Which checkpoint the two weight files are — `chk00042000`
`LICENSE.pdf`	Krea 2 Community License Agreement
`NOTICE.txt`	Required attribution notice

Both weight files are one adapter — only key names differ. Both carry training details in safetensors metadata. The file names carry no checkpoint number: both hold chk00042000, which also sits in _archive/checkpoints/ under its number, beside the earlier chk00030000, chk00018000, chk00012000 and chk00006000.

Quick Start

diffusers

pip install git+https://github.com/huggingface/diffusers.git

import torch
from diffusers import Krea2Pipeline
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Turbo", torch_dtype=torch.bfloat16).to("cuda")

lora = hf_hub_download("lvladikov/Krea2-Turbo-Distill-3step-LoRA", "krea2_turbo_3step_rank_64_lora.safetensors")
state = load_file(lora)
state = {f"transformer.{k}": v for k, v in state.items() if not k.endswith(".alpha")}
pipe.load_lora_weights(state, adapter_name="3step")

# guidance_scale=0.0 is CFG 1 in this pipeline: no guidance
image = pipe("a fox in the snow", num_inference_steps=3, guidance_scale=0.0).images[0]
image.save("krea2_3step.png")

ComfyUI

file	put it in
`krea2_turbo_3step_rank_64_lora_comfyui.safetensors`	`ComfyUI/models/loras/`
`krea2_turbo_bf16.safetensors` (from Comfy-Org/Krea-2)	`ComfyUI/models/diffusion_models/`
`qwen3vl_4b_bf16.safetensors` (same repo)	`ComfyUI/models/text_encoders/`
`qwen_image_vae.safetensors` (same repo)	`ComfyUI/models/vae/`

Load krea2_turbo_3step_lora_t2i.json. Full bf16, no quantisation, runs on CUDA / Apple Silicon / CPU unchanged. Smaller Turbo builds load the LoRA too — ComfyUI's loader applies it to any build — except fp8_scaled on Apple Silicon (MPS has no Float8_e4m3fn). A quantised base trades some of the detail: on Comfy Cloud's 4-bit (nvfp4) Turbo this adapter's fine texture measured about a fifth under the full-bf16 render's.

The ComfyUI workflow

Settings: steps 3, cfg 1.0, sampler euler / simple, LoRA strength 1.0.

⚙️ cfg 1.0, not 0.0. cfg 1.0 is "no guidance" in ComfyUI: one model call per step, the setting Turbo is made for. cfg 0.0 is not off — it keeps only the negative prompt's prediction and drops your prompt, which is why an image at 0.0 comes out broken. The diffusers example's guidance_scale=0.0 is this same cfg 1.0, counted the diffusers way.

On Apple Silicon, over MLX

krea2_turbo_3step_lora_t2i_MLX_LVNodes.json runs the same LoRA file on MLX instead of PyTorch / MPS, through the LVNodes Krea 2 MLX custom node: Krea 2 Model (MLX) loads the same bf16 files, Krea 2 Sampler (MLX) encodes, samples and decodes. To install the custom node, see ComfyUI-Krea2-MLX and LVNodes; its first run sets up its own MLX environment and leaves ComfyUI's packages alone.

Settings: steps 3, cfg 1.0, scheduler euler; the LoRA in the model node's step-LoRA slot for 2-3 steps (lora_2step), enable_step_loras on, strength 1.0. About 64 GB of memory suits bf16; the node's quantize 8 or 4 needs less. MLX reads the bf16 files only (ComfyUI's fp8 and int8 builds are for PyTorch).

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.

Krea 2 Raw, 13 steps, CFG 3.5, this LoRA at 1.0 — the five renders in raw-13step-cfg3.5-lora/

Krea 2 Raw at 13 steps, CFG 3.5, with this LoRA

Krea 2 Raw, 20 steps, CFG 3.5, this LoRA at 1.0 — the five renders in raw-20step-cfg3.5-lora/

Krea 2 Raw at 20 steps, CFG 3.5, with this LoRA

Krea 2 Raw, 6 steps, CFG 3.5, this LoRA at 1.0 — all 22 prompts in raw-6step-cfg3.5-lora/

Krea 2 Raw at 6 steps, CFG 3.5, with this LoRA

Six steps is an extreme, short-stepped use of Raw, and it still works for most of the 22 prompts: faces, animals, crowds and illustrations come out whole. Low light suffers most (the pianist is nearly black, the night street and the dinner table come out dark), the clay chef shrinks into an empty set, and the splash-art sorceress gains a title banner. Take it as a starting point for experimenting with Raw at fewer steps.

Rendered on Comfy Cloud with the full bf16 Raw at 1024×1024, seed 4242, euler / simple, an empty negative prompt.

LoRA Strength

strength	what happens
below 1.0	Correction only partly applied — softer skin and hair, lighter freckles, a step back toward 3 steps without the adapter (0.5 below)
1.0	Trained point, recommended
1.0–1.5	Extrapolation — texture denser than the subject warrants, freckles heavy, hair wiry, contrast up (1.5 below). Usable per prompt
above 1.5	Not recommended, not measured

At three steps the dial scales the adapter's whole job — turning three coarse calls into a finished image — so the useful range is narrow around 1.0. Reach for steps before strength: when closeness to the teacher's own picture matters more than speed, the 4-step adapter is the better tool.

LoRA strength comparison

Known Issues

In order of how often they show:

Further from the teacher's own picture than the 4-step adapter — distance 0.379 against 0.320, detail agreement 0.27 against 0.39. Three steps is a shorter path from the same noise, so framing, pose or a detail can differ from the 8-step render at the same seed: the signs along a neon-lit street, a drone's view over a night city, the dishes on a family's dinner table. Treat the teacher's render as a quality reference, not the picture three steps will reproduce
Texture stronger than the teacher's — fine texture 1.23×, grain in flat areas 1.74× (the earlier release 1.57×): watercolour paper most of all, then skin pores and film grain read more pronounced than in the 8-step render, freckles denser. A strength under 1.0 softens it (see LoRA Strength)
Doubled edges — the blur-invariant ghosting index is 1.24× the teacher's (the 4-step adapter's 1.02×): a faint doubled contour can show on fine edges — hair, rain streaks, thin limbs — most at the middle sizes, 768×1024 and 1024×768
Small faces in crowds — still the weakest spot of any few-step adapter: faces far back in a crowd can come out soft or smeared, where a portrait-sized face holds up

These are the published checkpoint's limits, stated as they are.

Hardware

Trained on a single NVIDIA H100 (initially) then switched to H200 (for quicker results). The frozen base runs in bf16 (the 4-step and 2-step adapters trained on a single RTX 3090 with a weight-only int8 base). Batches of 3 at 512×768 / 768×512, 2 at 768×1024 / 1024×768 and 1 from 1024×1024 up keep every term inside about 45–50 GB; a step with every term live takes 4–24 s by size. Throughput averaged about 365 samples an hour (42,000 in about 115 hours) — each sample costs a dozen or so passes of the 12.8B base (the fake score's three updates alone are about 40% of a step). The trajectories were recorded once and are read from disk.

Released LoRA is bf16.

Usage Notes

🎯 Made for Krea 2 Turbo — trained against Turbo's weights and schedule; on Raw it is a bonus at 6–20 steps with CFG on, not a supported mode (see Using it on Raw)
🚫 No guidance: CFG 1 — cfg 1.0 in ComfyUI (0.0 breaks the image), guidance_scale=0.0 in diffusers (its way of writing the same setting)
📐 Keep μ = 1.15 — the three training sigmas are anchored to that grid
🔬 Training used a bf16 base; released LoRA is bf16

Examples

Every sheet below, at 1024×1024: base model (8 steps), base at 3 steps without LoRA, base at 3 steps with LoRA — same seed and prompt. Panels 2 and 3 are ComfyUI renders (full bf16; panel 3 is the published resolution sweep's own) from the same noise, so comparing them isolates the LoRA's effect; panel 1 is the quality bar, drawn from its own noise, not a pixel target. NFE = steps (Turbo is CFG-free).

Portrait of a young woman with freckles and windswept auburn hair...

portrait comparison

Turbo 8 steps	Turbo 3 steps, no LoRA	Turbo 3 steps + LoRA

Kingfisher bursting out of water...

kingfisher comparison

Turbo 8 steps	Turbo 3 steps, no LoRA	Turbo 3 steps + LoRA

Rainy night city street with glowing neon signs...

neonstreet comparison

Turbo 8 steps	Turbo 3 steps, no LoRA	Turbo 3 steps + LoRA

(19 more examples, and a two-panel sheet against the teacher for every prompt, in the full version — see assets/ for all 22 prompts)

Resolution Sweeps

_archive/resolution_sweeps/chk00042000/ — this LoRA at its 11 trained sizes for all 22 test prompts; 3 steps, strength 1.0, rendered on Comfy Cloud with the full bf16 Krea 2 Turbo through ComfyUI's stock LoRA loader (the _comfyui file as it ships). The three raw- folders hold the prompts on Krea 2 Raw with this LoRA at CFG 3.5 (see Using it on Raw). The earlier chk00030000, chk00018000, chk00012000 and chk00006000 keep their own sweeps beside it.

_archive/resolution_sweeps/chk00042000/
├── 512x768/                  <slug>.jpg for the 22 prompts
├── 768x512/
├── ...                       (the 11 trained sizes)
├── 1824x1024/
├── raw-6step-cfg3.5-lora/    the 22 prompts on Krea 2 Raw, 6 steps
├── raw-13step-cfg3.5-lora/   5 prompts on Krea 2 Raw, 13 steps
└── raw-20step-cfg3.5-lora/   the same 5 at 20 steps

The LoRA panels in the examples above are this sweep's 1024×1024 renders.

Detailed Model Card

The full write-up is in the Detailed Model Card:

Where it stands — the release at a glance, and its known issues
How I got here — from the 4-step adapter to this checkpoint
chk00042000 vs chk00030000 — what this release changed against the earlier one
Measured against the teacher — the 22 prompts at 768×1024, five at 1024×1024, every sweep size
Inference with diffusers — the settings explained, strength, stock Turbo for comparison
Performance — the time per model call and per render
File format and compatibility
Method — the schedule and the teacher grid, the losses, the critics, settled weights
What the LoRA touches — the 228 modules
Training data — the recorded trajectories and the photo crops
Resolutions and batches — the share of training at each size
Hardware
Examples — all 22 prompts against the teacher

License

This adapter is a Derivative of Krea 2 Turbo under the Krea 2 Community License Agreement. Everything the agreement says about Krea 2 Turbo applies to this LoRA: Acceptable Use Policy, revenue threshold for commercial use, content-filtering duty for deployments.

Copy of agreement: LICENSE.pdf, required notice: NOTICE.txt. See krea.ai/krea-2-licensing.

Not an official Krea product, not endorsed by Krea. Base model is Krea's; adapter weights and everything else in this repo are mine.

Krea2-Turbo-Distill-3step-LoRA