Krea-2-Turbo-Distill4-mflux-q8

Krea-2-Turbo-Distill4 · MLX · 8-bit (mflux)

Krea 2 Turbo, 8-bit MLX, with the 4-step distillation LoRA already fused. Four steps, no adapter to ship, no bake at load.

This is a derivative of Krea 2 Turbo, not an official Krea product.

What it is

mflux-community/krea-2-turbo-mflux-q8 with lvladikov/Krea2-Turbo-Distill-4step-LoRA applied at scale 1.0 and baked in, then saved as an mflux checkpoint.

It exists for one reason: mflux pays to bake the adapter every time a process starts. Measured on an M5 Pro, loading q8 and baking the LoRA takes 11.118 s; loading this checkpoint takes 2.751 s. That is 8.37 s saved per process start, and one file instead of two.

It is not faster per step and not higher quality than load-and-bake. It is the same thing, pre-baked.

Measured quality with a persona LoRA on top (updated 2026-09-20)

Read this before choosing a sampler. The card originally recommended er_sde, which is mflux's own stepper on mflux's own sigma schedule, and on a 4-step render with a persona LoRA stacked on top that combination produced ghosting and crushed blacks — 0 wins in 35 blind pairwise comparisons against an 8-step reference.

The cause was not this checkpoint. It was two things in the renderer around it, found on 2026-09-20 and measured on an M5 Pro:

1. mflux's Krea 2 LoRA mapping has no pattern for ComfyUI's .diff_b keys. An adapter that carries bias deltas is loaded with those tensors silently dropped. The distill adapter fused here carries none, so this checkpoint is faithful — but a persona LoRA applied on top of it may not be.
2. mflux's sigma schedule is not the one Krea 2 is sampled on in ComfyUI. ComfyUI uses a beta(0.6, 0.6) schedule over a shift of 1.15 with euler_ancestral; mflux uses a linspace with a dynamic shift (mu 1.0146 at 1024x1280). At 8 steps mflux arrives at its final step with 76% more noise left to remove.

With the bias deltas applied and the ComfyUI schedule reproduced, the 4-step render with a persona stack is clean — identity intact, no doubling, blacks not crushed — and 5 and 6 steps add only small monotone detail (mean abs diff 5.70 and 7.20 out of 255). Warm median 42.41 s at 1024x1280 (n=4, 41.83–43.98) against 84.15 s for 8 steps on the same box.

So: 4 steps is the right recipe, and er_sde is not. If your runtime can reproduce ComfyUI's beta + euler_ancestral over shift 1.15, use it. If it cannot, this checkpoint still works at 4 steps on its own — the failure above needed a persona LoRA stacked on top to show up.

Verification

Output is pixel-identical to load-and-bake. mflux renders are deterministic at a fixed seed, so this is an exact check rather than a similarity score: the SHA-256 of the decoded RGB buffer is 517500d4492908ea5c1a5c90d63aa51477a2b0 1004ba25b1cbe4d392ae34e6e9 for both paths (prompt fixed, seed 42, 4 steps, 832x1216, guidance 1.0, er_sde), and the same hash appears again from an independent third run that baked the adapter in a different process.

Measured on: Apple M5 Pro, 20-core GPU, 48 GiB unified, macOS 26.5.1, mflux 0.19.2, mlx 0.32.2. Raw record: fuse-distill.json in the build folder.

Use

hf download 00cyre/Krea-2-Turbo-Distill4-mflux-q8 --local-dir ./Krea-2-Turbo-Distill4-mflux-q8

mflux-generate-krea2 \
  -m ./Krea-2-Turbo-Distill4-mflux-q8 \
  --prompt "..." --seed 42 --width 832 --height 1216 --steps 4

Recipe: 4 steps, guidance 1.0 (CFG off). Running it at 8 steps is not an upgrade — the distillation targets a 4-step schedule. On the sampler, see "Measured quality with a persona LoRA on top" above: er_sde on mflux's own sigma schedule is what the 2026-09-18 numbers below were taken with, and it is not the schedule Krea 2 is sampled on in ComfyUI.

from mflux.models.krea2 import Krea2
model = Krea2(quantize=8, model_path="./Krea-2-Turbo-Distill4-mflux-q8")
model.generate_image(seed=42, prompt="...", num_inference_steps=4,
                     width=832, height=1216, guidance=1.0).save("out.png")

Limitations

• The benefit is 8.4 seconds per process start. If you already keep a model resident, this saves you that once. It is a convenience artifact, not a performance result.
• The adapter strength is frozen at 1.0. The upstream LoRA card recommends 0.75 when the texture reads too strong on stylised or very large renders. This checkpoint cannot be dialled back; use the base plus the adapter if you need that knob.
• q8 only. Fusing onto q4 is the interesting case — mflux's runtime bake re-quantizes the touched layers up to q8 and gives back q4's memory advantage (measured again on 2026-09-20 with a persona stack: 263 layers re-quantized, and the q4 server then peaked at 35.22 GB against q8's 35.59 GB) — but doing it properly needs the dense base weights to fuse before quantizing, which were not available on the build machine.
• Inherits everything from both parents, including the base model's behaviour and the adapter's trained resolutions (512–1440).
• Stacking a persona or style LoRA on top is where the sampler starts to matter, and where an adapter carrying .diff_b keys will be silently partially applied by mflux 0.19.2. See the quality section above.
• Not tested on any chip other than an M5 Pro.

License and attribution

Licensed under the Krea 2 Community License Agreement, the same license as both parents, as that license requires of derivatives. The agreement itself is LICENSE.pdf in this repository, and the attribution notice section 3.1(c) requires is NOTICE.txt:

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.

Use is also subject to the Krea Acceptable Use Policy (krea.ai/krea-2-use-policy). See krea.ai/krea-2-licensing.

This is a modified, derived work:

• Base: krea/Krea-2-Turbo by Krea (via the mflux 8-bit MLX conversion mflux-community/krea-2-turbo-mflux-q8)
• Adapter: lvladikov/Krea2-Turbo-Distill-4step-LoRA

The name begins with "Krea" as the license requires of derivative model names. This is not an official Krea product and is not endorsed by Krea. Downstream terms may not be more restrictive than the original agreement. Commercial use is limited by the agreement's revenue threshold, and anyone deploying this behind a generation service takes on the agreement's content-filtering obligation — that is an active duty, not boilerplate.

Contains no private data and no persona adapters.

What is in this folder

The weights are committed here (22.19 GB): the seven *.safetensors shards and model.safetensors.index.json, plus text_encoder/, tokenizer/ and vae/. Nothing else is needed to run it.

build.sh in this directory is the recipe that produced them, and reproduces the checkpoint from the two public parents on any Apple Silicon machine with mflux 0.19.2+. Its fusing step calls fuse_distill.py, which is not published here, so the script documents the build rather than running end to end unchanged.

Companion code

mflux-krea2-fidelity/ in this repository holds four small modules for mflux users working with Krea 2 LoRAs: a tool that audits a LoRA file for the bias-delta (.diff_b) keys mflux 0.19.2 silently drops, the patch that applies them the way ComfyUI does, and a schedule/stepper pair that reproduces ComfyUI's beta(0.6, 0.6) schedule over a shift of 1.15 with euler_ancestral for Krea 2 — see the "Measured quality with a persona LoRA on top" section above for why both matter. That folder has its own README and its own, separate license note; it is extracted from a private render server and provided as-is.