dataautogpt3/Krea2-weights-experiments

Krea 2 Turbo — Hand-Edited Weight Experiments

[image]

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)

B3 — Attention-Only Partial Inversion

D — Gate Scaling (All Blocks)

F — Early/Late Block Inversion

G — Mid-Block Alpha Sweep

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

All target blocks 12-14, ALL layers.

H — Layer-Selective Mid-Block

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

I — Gradient Alpha

Excluded Variants (Broken)

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

Usage

ComfyUI

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

1. 1.10% inversion is the sweet spot. Alpha=0.10 (scale=0.80) on mid blocks 12-14 produces the most artistically interesting results.

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

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

1. 1.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