DarkBeast Krea 2 Turbo Merge

All credit goes to the original authors of Krea and the wizards at @AiMetatron and their already perfect Dark Beast | 黑兽 INT8 Convrot 2倍速 | Krea2 Aggressive Edition

I wanted to try a model merge so I took the OG Krea 2 Turbo and blended in the Dark Beast. I am happy with the results so I will share here, but I encourage you to download the original Dark Beast and Krea 2 Turbo models for your creations as my merge may not be to everyone's liking.

The results?

A custom Krea 2 Turbo merge combining the speed, prompt adherence, and low-step performance of Krea 2 Turbo v1.0 with the darker, more dramatic visual character of DarkBeast Krea 2.

The goal of this merge was not to create a simple 50/50 blend. Instead, the model preserves most of Krea 2 Turbo’s core conditioning and early transformer structure, while gradually introducing more DarkBeast influence through the middle and later transformer blocks.

The result retains excellent performance at 8 steps, while adding richer contrast, moodier lighting, more detailed textures, and a stronger cinematic or dark-fantasy aesthetic.

Merge Structure

Model 1: Krea 2 Turbo v1.0
Model 2: DarkBeast Krea 2

The merge uses Krea 2 Turbo as the primary structural base.

Key timestep and projection layers remain approximately 90% Krea 2 Turbo to preserve distilled low-step behaviour. Text-conditioning layers remain mostly Turbo-biased for reliable prompt adherence.

DarkBeast influence increases gradually across the transformer stack:

• Blocks 0–6: 25% DarkBeast

• Blocks 7–13: 35% DarkBeast

• Blocks 14–20: 40% DarkBeast

• Blocks 21–27: 45% DarkBeast

The final output section returns to a stronger Krea 2 Turbo weighting for stability.

Merge Ratios

first.                                  0.80
tmlp.                                   0.90
txtmlp.                                 0.70
tproj.                                  0.90

txtfusion.layerwise_blocks.0.           0.70
txtfusion.layerwise_blocks.1.           0.70
txtfusion.projector.                    0.80
txtfusion.refiner_blocks.0.             0.65
txtfusion.refiner_blocks.1.             0.65

blocks.0 through blocks.6               0.75
blocks.7 through blocks.13              0.65
blocks.14 through blocks.20             0.60
blocks.21 through blocks.27             0.55

last.                                   0.70

Ratios represent the amount of Model 1, Krea 2 Turbo v1.0. The remaining percentage comes from DarkBeast Krea 2.

Recommended Settings

The model performs well using standard Krea 2 Turbo workflows.

Suggested starting point:

Steps: 8
CFG: 1.0
VAE: qwen_image_vae.safetensors
Text encoder: qwen3vl_4b_fp8_scaled.safetensors
Sampler: er_sde
Scheduler: sigmoid_offset
You can also use euler/simple

The model can also be tested at 9 or 10 steps for additional refinement, although 8 steps already produces strong results. Too high add unwanted noise.

Strengths

• Strong performance at low step counts

• Good prompt adherence

• Cinematic and dramatic lighting

• Rich contrast and atmospheric rendering

• Detailed faces, clothing, textures, and environments

• Excellent skin texture

• Works well for NSFW or SFW

• Particularly suitable for dark fantasy, gothic, cinematic, fashion, portrait, and dramatic photographic imagery

Notes

This is an experimental model merge and may behave differently from either parent model depending on the prompt and workflow.

It was created with ComfyUI’s native ModelMergeKrea2 node and saved as a standalone Krea 2 diffusion model.

Please do not share or re-merge this model without giving credit to @AiMetatron