Krea2-Turbo-Distill-4step-LoRA

Krea 2 Turbo — 4-Step Distillation LoRA

🆕 New: a 3-step LoRA takes Krea 2 Turbo from 8 steps down to 3, at three quarters of this LoRA's denoising time, with more fine detail (fine texture 1.26× the 8-step teacher's on its 22 test prompts). This 4-step LoRA stays the closest to Turbo's own 8-step pictures.

Half the steps · ~1.6× faster end to end · cleaner than the first release: grid patterns, ghosting and flat-area grain all closer to the 8-step teacher · 13 trained resolutions up to 1824×1024 · 133,000 training samples, 1.7× the first release.

🆕 This version (chk00133000) replaces the first 'final' release (chk00078000), from two months ago. I decided to train a 'few more' samples (55K more to be precise), the newer part with five critic heads where the first release had one. Measured against the 8-step teacher on the 15 test prompts at 12 sizes:

• cleaner fine structure — the 8- and 16-pixel grid patterns at 0.96–0.98× the teacher's, where the first release was at 1.09–1.11×
• less ghosting and fewer doubled edges — closer to the teacher at 10 of 12 sizes on both measures
• cleaner flat areas — grain at 0.99× the teacher's, down from 1.08×
• no over-sharpening at large sizes — fine texture at 1.01–1.04× the teacher's at 1440 px, down from 1.18–1.20×
• saturation matching the teacher's

Fine texture now sits at the teacher's level, a touch softer than the first release at mid sizes. The first release stays in _archive/checkpoints/.

A LoRA for Krea 2 Turbo that reduces the minimum usable step count from 8 to 4 — Turbo's own model and sigmas, half the denoising passes, and the fine texture that 4-step Turbo loses put back.

• ⚡ Half the steps — 8 → 4, on Turbo's own deployment sigmas
• ⏱️ ~1.6× faster end to end — 54.5 s vs 88.7 s at 1024×1024 for the whole run, model load, text encoding and VAE decode included (1.8× on denoise alone, without them)
• 🎯 Fine texture at the teacher's level — 0.96× the teacher's fine-texture energy on average over 12 sizes, 1.01–1.04× at 1440 px; the 8- and 16-pixel grid bands 0.96–0.98×, flat-area grain 0.99×, saturation 1.00×
• 🗣️ Prompt-aware training — critic scores images against their prompts during training
• 📐 13 trained resolutions — multi-aspect from 512×512 up to 1440×1440, and 1824×1024 since the first release
• 🔌 Drop-in — plain LoRA weights for diffusers and ComfyUI. No custom nodes, no patched sampler, no code
• 🎲 13,750 prompts from Lakonik's 3M-prompt dataset, each recorded as a full 8-step teacher trajectory; since the first release, 91,044 training chords on Turbo's own 8-step paths
• 📷 43,044 real-photo crops — 25,560 from LSDIR + 17,484 from Flickr2K, captioned, in critic's real set; since the first release, 5,234 crops cut from 4K photos
• 🔢 133,000 training samples in shipped weights — 1.7× the first release's 78,000
• 📅 21 days on a single RTX 3090 to the first release; since then about a day on an H200 and two more on the RTX 3090
• 🔁 16 recipe adjustments — each made on measurement of the one before; since the first release, five critic heads instead of one

💡 Too strong on a prompt? Turn it down. The LoRA restores fine texture, and on some subjects — stylised art, high-contrast splash pieces, very large renders — that can read as too much at strength 1.0. Effect scales smoothly: 0.75 is a good second setting, 0.6–0.9 is fair game.

---

Files

Both weight files are one adapter — only key names differ. Both carry training details in safetensors metadata.

---

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-4step-LoRA", "krea2_turbo_4step_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="4step")

image = pipe("a fox in the snow", num_inference_steps=4, guidance_scale=0.0).images[0]
image.save("krea2_4step.png")

• num_inference_steps=4 is the whole config. Pipeline applies Turbo's fixed timestep shift (μ = 1.15) and evaluates at σ = 1.0, 0.905, 0.760, 0.513 — exactly the four points the LoRA was trained on. Keep guidance_scale=0.0.
• Strength: pipe.set_adapters(["4step"], adapter_weights=[0.75]). Stock Turbo for comparison: pipe.unload_lora_weights() + num_inference_steps=8.

ComfyUI

Load krea2_turbo_4step_lora_t2i.json. Full bf16, no quantisation, runs on CUDA / Apple Silicon / CPU unchanged.

Settings: steps 4, cfg 1.0, sampler euler / simple, LoRA strength 1.0.

⚙️ cfg 1.0, not 0.0. ComfyUI expresses "no CFG" as 1.0 (one forward pass); diffusers uses 0.0. Setting 0.0 in ComfyUI is not the same thing.

---

LoRA Strength

Reach for steps before strength: 1.0 at more steps is the dependable way to get more. Strength and step count trade against each other.

---

Hardware

Trained on single RTX 3090 (24 GB) to the first release; since then about a day on a single H200 and two more on the RTX 3090. Frozen base quantized weight-only to int8 (blockwise-64) — ~0.007 relative error, ~3% step time cost. Teacher trajectories rolled at same precision (targets permanent, error baked in).

Every bucket trained under full int8 up to 1440×1440. Large buckets fit via:

• Checkpointed block inputs staged to pinned host memory (numerically exact)
• Discriminator pass after generator backward (peaks don't overlap)
• Trunk stopped at feature tap it actually reads

Released LoRA is bf16 applied to unquantized base.

---

Usage Notes

• 🎯 Krea 2 Turbo only — trained against Turbo's weights and schedule
• 🚫 Keep guidance at 0.0 (ComfyUI: cfg 1.0, not 0.0)
• 📐 Keep μ = 1.15 — training targets anchored to that grid
• 🔬 Training used int8 base; released LoRA is bf16 on unquantized base

---

Examples

Every sheet below: base model (8 steps), base at 4 steps without LoRA, base at 4 steps with LoRA — same seed throughout. Compare panels 2 vs 3 to isolate LoRA effect. NFE = steps (Turbo is CFG-free).

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

---

Kingfisher bursting out of water...

---

Rainy night city street with glowing neon signs...

---

(12 more examples in full version — see assets/ for all 15 prompts)

---

Resolution Sweeps

assets/resolution_sweeps/ — this LoRA at every trained resolution for all 15 test prompts (same prompts, seed, 4 steps, strength 1.0), 1824×1024 included. Nothing cherry-picked. This version's sweep was rendered in ComfyUI (seed 4242); the teacher's renders came from another renderer, so pose and framing can differ between the two.

_teacher-8step/ — official Krea 2 Turbo 8-step reference renders for same prompts/seeds/resolutions.

Layout:

assets/resolution_sweeps/
├── _teacher-8step/         8-step stock Turbo reference
│   ├── 512x512/
│   ├── 1024x1024/
│   └── ... (all 12 buckets)
└── 4step-LoRA/             Same tree, this LoRA at 4 steps (+ 1824x1024)

Two ways to read: down the sweep (does it hold across resolutions?) or against teacher (open same file in both folders).

---

Archive

Every published checkpoint and its resolution sweep under _archive/ — checkpoints/ and resolution_sweeps/, each under its number: chk00133000, this release, the first 'final' release chk00078000 (its sweep in chk78000-final/), and the earlier chk00060000, chk00042000, chk00026000, chk00019000, chk00014000, chk00010000, chk00006000, chk00005000 and chk00004000.

---

Detailed Model Card

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

• How I got here — from stock Turbo to the first release, and its timeline
• chk00133000 vs chk00078000 — what this release changed against the first one
• Measured against the teacher — the first release's measurements: texture, the 4-step gap, the judged renders
• This is not a Raw→Turbo diff — and using it on Raw
• Inference with diffusers and ComfyUI — the settings explained
• Performance — the time per model call and per render
• LoRA strength — where the useful range is
• File format and compatibility — the keys, runtime by runtime
• Resolution sweeps — every trained size, every test prompt
• Method — the progressive-distillation targets and the critic
• What the LoRA touches — the 228 modules
• Training data — the recorded trajectories and the photo crops
• Resolutions — the share of training at each size
• Hardware
• Examples — all 15 prompts against the teacher, and why these prompts
• What's next — the 2-step and 3-step LoRAs that followed

---

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.