dataautogpt3/Krea2-weights-experiments

Krea 2 Turbo — Hand-Edited Weight Experiments

Overview

This repository contains weight-edited variants of the Krea 2 Turbo diffusion model. Each variant was created by surgically scaling specific transformer block weights in the 12.8B parameter single-stream MMDiT, producing artistic and functional model variations without any retraining.

These are research artifacts from hand-editing diffusion model weights using the methodology described below. The base models (Krea 2 Turbo and Krea 2 Raw) are NOT included — only the edited variants.

Method

All variants use the core formula:

theta_new = theta_original * (1 - 2 * alpha)

Where alpha controls the inversion strength:

• alpha=0.05 → scale 0.90 (subtle)
• alpha=0.10 → scale 0.80 (artistic sweet spot)
• alpha=0.15 → scale 0.70 (strong)
• alpha=0.20 → scale 0.60 (aggressive but functional)

Full negation (alpha=0.5, scale=-1.0) breaks the model and is excluded from this repository.

Architecture: Krea 2 Turbo

• Type: Single-stream MMDiT (Diffusion Transformer)
• Parameters: 12.8B
• File size: ~25GB per variant (BF16 + F32 tensors)
• Structure: 28 uniform transformer blocks
• Block sub-layers:
  • blocks.N.attn.* (7 tensors): gate, qknorm, wq, wk, wv, wo
  • blocks.N.mlp.* (3 tensors): gate, up, down (SwiGLU)
  • blocks.N.mod.lin (1 tensor): conditioning modulation
  • blocks.N.prenorm.scale / blocks.N.postnorm.scale

Variants

B1 — Partial Inversion (Most Artistic)

Property	Value
File	Krea_2_turbo_inv_B1_partial10.safetensors
Blocks	12-14 (mid)
Layers	ALL (39 tensors per block group)
Alpha	0.10 (scale=0.80)
Result	Most artistic variant — strong style/content shift while remaining coherent

B3 — Attention-Only Partial Inversion

Property	Value
File	Krea_2_turbo_inv_B3_attn_p10.safetensors
Blocks	12-14 (mid)
Layers	attn only (21 tensors)
Alpha	0.10 (scale=0.80)
Result	Functional, subtler than B1 — attention-specific perturbation

D — Gate Scaling (All Blocks)

Property	Value
File	Krea_2_turbo_inv_D_gate_p20.safetensors
Blocks	0-27 (all)
Layers	attn.gate only (28 tensors)
Alpha	0.20 (scale=0.60)
Result	Functional, moderate effect — gate weights are more tolerant of aggressive scaling

F — Early/Late Block Inversion

Property	Value
File	Krea_2_turbo_F_early_a10.safetensors
Blocks	0-2 (early)
Layers	ALL
Alpha	0.10 (scale=0.80)
Result	Affects structure, composition, spatial layout

Property	Value
File	Krea_2_turbo_F_late_a10.safetensors
Blocks	25-27 (late)
Layers	ALL
Alpha	0.10 (scale=0.80)
Result	Affects style, color, detail, texture refinement

G — Mid-Block Alpha Sweep

Three variants at different inversion strengths on the same block zone:

File	Alpha	Scale	Notes
Krea_2_turbo_G_mid_a05.safetensors	0.05	0.90	Subtle
Krea_2_turbo_G_mid_a15.safetensors	0.15	0.70	Strong
Krea_2_turbo_G_mid_a20.safetensors	0.20	0.60	Aggressive but functional

All target blocks 12-14, ALL layers.

H — Layer-Selective Mid-Block

File	Blocks	Layers	Alpha
Krea_2_turbo_H_mid_attn_a10.safetensors	12-14	attn only	0.10
Krea_2_turbo_H_mid_mlp_a10.safetensors	12-14	mlp only	0.10

Isolates the effect of attention vs MLP perturbation on the same block zone.

I — Gradient Alpha

Property	Value
File	Krea_2_turbo_I_gradient.safetensors
Blocks	0-27 (all)
Layers	ALL
Alpha	0.03 → 0.17 (gradient across blocks)
Scale	0.94 → 0.66
Result	Smooth global perturbation — early blocks barely touched, late blocks aggressively inverted

Excluded Variants (Broken)

The following variants were created but are broken (model produces noise/garbage) and are NOT included:

Variant	What was done	Why it broke
B2_attn_full	attn weights * -1.0	Full negation destroys attention computation
D_wv_all	wv weights * -1.0	Full negation of value projection
E_ties_mid	TIES-style sign flip on mid blocks	Full negation variant

Usage

ComfyUI

1. Place .safetensors files in ComfyUI/models/diffusion_models/
2. Load via UNETLoader node
3. Use the same VAE, CLIP, and text encoder as Krea 2 Turbo
4. Generate with your standard Krea 2 workflow

Diffusers

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "dataautogpt3/Krea2-weights-experiments",
    torch_dtype=torch.bfloat16,
    variant="bf16"
).to("cuda")

Note: These are diffusion model weights only. You need the corresponding VAE, text encoders, and tokenizer from the original Krea 2 Turbo release.

Key Findings

1. Scaling works, full negation breaks. Partial inversion (scale 0.60-0.90) produces functional, artistic variants. Full negation (scale=-1.0) breaks the model.

2. 10% inversion is the sweet spot. Alpha=0.10 (scale=0.80) on mid blocks 12-14 produces the most artistically interesting results.

3. Mid blocks are safest to modify. Blocks 12-14 are the most redundant and tolerate perturbation best.

4. Gate weights are most tolerant. Attention gate weights can be scaled to 0.60 across all blocks while remaining functional — other layers break sooner.

5. The artistic effects come from compensation. Partial perturbation triggers creative reorganization in unedited blocks — the compensatory masquerade effect.

Research Context

This work draws on findings from:

• Task Arithmetic (Ilharco et al., ICLR 2023) — formal basis for weight negation
• weights2weights (NeurIPS 2024) — diffusion weight space as meta-latent
• Unraveling MMDiT Blocks (2025) — per-block role mapping for MMDiT
• C3: Creative Concept Catalyst (CVPR 2025) — low-frequency amplification in shallow blocks
• ConceptPrune (ICLR 2025) — tiny weight changes shift semantic output

Credits

• Base model: Krea 2 Turbo (Krea AI)
• Weight editing: DataPlusEngine
• Methodology: Hand-editing diffusion weights via mmap-based surgical tensor scaling