krea2-presets — Krea-2 Turbo with block weights edited (experiment, 2026-10-15)

What this is

krea2_turbo_bf16_picks_half.safetensors is the official Krea-2 Turbo bf16 model with 19 of its 28 transformer blocks rescaled. No training was done. The goal was more realistic photos (skin, lighting, colour) from the base model, so a character LoRA could later be trained on top of it.

• Turned down ×0.90: blocks 3, 6, 10, 12, 13, 14, 15, 16, 17, 18, 25
• Turned up ×1.10: blocks 1, 7, 9, 11, 19, 20, 23, 24
• Every tensor in those blocks is scaled (attention, MLP, norms) except mod.lin.
• krea2_picks_half_preset.json is the recipe. bake_krea2_preset.py rebuilds the file from the original model.

⚠️ Verdict: do not use this file

In a blind test it ranked last (average rank 2.50, against 2.00 for the unedited model).

How we got here

Inspired by the Reddit post "Taking Krea-2 apart, one piece at a time" and the Arthemy Krea-2 Tuner ComfyUI node. All renders used Krea-2 Turbo, 8 steps, euler/simple, CFG 1, 1024×1280, with fixed seeds, and were judged by eye.

1. Character LoRA block strengths (Emmy LoRA): almost no effect. The LoRA locks the composition.
2. Base-model group edits (±10%): they move the pose a little. Blocks 24–27 are a texture dial: down gives mush, up gives grain or noise.
3. "Chaos" random nudges: one pattern (seed 1 at 0.2) gave great skin and ground texture on one prompt, but made a second prompt worse. It isn't reliable.
4. Prompt wording beat every weight edit: swapping "sharp saturated high-contrast" for "natural muted colours, flat even daylight, plain unedited phone photo" fixed the colours. Background blur couldn't be fixed by the prompt, the weights or a depth-of-field slider, which suggests it's a Krea-2 limit.
5. Block map: each block was changed alone by ±20% and judged for "most real" (2 seeds). The picks are the lists above.
6. Blind rank, 12 fresh seeds: the picks applied live through the Arthemy node at ±10% won (average 2.33 against the original's 3.42). They won again in a 3-way rerun (1.50 against 2.00).
7. Export: this file bakes that recipe. Arthemy's own ModelSaver node produced a file that renders pixel-identical to this one. Both differ from the live node render, and the baked version ranked last.

Takeaway

The winning look only exists while the node patches the model live. Once baked into weights, it isn't the same, and it's worse. The cause is unknown. Our guess is that the live scalar patch is only partly applied in ComfyUI.