Krea-R-Turbo

Link to Model:
https://huggingface.co/realrebelai/Krea-R-Turbo/tree/main

Test the model on HF Spaces for free:
https://huggingface.co/spaces/realrebelai/Krea-R-Turbo

Full Weight and GGUF quants of custom merge of Krea-2-Turbo and 2 of my style LoRAs at specific strength values. Displays a heavy focus on photorealistic portraits to achieve high grade aesthetics while also retaining Krea-2s sharp detail!

This is a derivative of krea/Krea-2-Turbo: LoRA-merged, then quantized. It is not the original checkpoint.

Quants

INT8 Covrot now available!

Pick by how much VRAM you can spare for the diffusion model (leave room for the text encoder + VAE). Sizes are approximate — see the Files tab for exact numbers.

FileType~SizeNotes

Krea-R-Turbo-Q8_0.gguf Q8_0~14 GB Near-lossless. For 16 GB+ cards or offload.

Krea-R-Turbo-Q6_K.ggufQ6_K~11 GBExcellent quality, hard to tell from Q8.

Krea-R-Turbo-Q5_K_M.ggufQ5_K_M~9.4 GBGreat quality/size balance.

Krea-R-Turbo-Q4_K_M.ggufQ4_K_M~8.0 GB Recommended all-rounder. 12 GB cards, or 8 GB with offload.

Krea-R-Turbo-Q3_K_M.ggufQ3_K_M~6.4 GBBest pick for 8 GB VRAM. Small quality trade.

Krea-R-Turbo-Q2_K.ggufQ2_K~5.5 GBTightest fit. Visible degradation but usable.

On an 8 GB card (e.g. RTX 3070), Q3_K_M or Q2_K fit most comfortably alongside the VAE and text encoder. Larger quants still run via ComfyUI's --lowvram / block-swap offload — just slower.

Usage in ComfyUI

Place the files:

FileFolderKrea-R-Turbo-QX.gguf (this repo)ComfyUI/models/unet/ (or diffusion_models/)qwen3vl_4b text encoderComfyUI/models/text_encoders/qwen_image_vae.safetensorsComfyUI/models/vae/

The text encoder (Qwen3-VL 4B) and VAE (Qwen-image VAE) are the standard Krea 2 companions and can be pulled from the ComfyUI repack: Comfy-Org/Krea-2.

Wire it up: load the GGUF with the Unet Loader (GGUF) node, feed it the Qwen3-VL text encoder and Qwen-image VAE, and sample as normal Krea 2 Turbo.

Recommended sampling

Krea 2 Turbo is a distilled, few-step model:

• Steps: ~8

• Guidance: CFG-free — set CFG to 1.0 (no negative prompt), or guidance 0

• Sampler / scheduler: er_sde + simple works well (or euler + simple)

• Resolution: up to 2048², 1024–1280 sweet spot on 8 GB

No positive/negative CFG split is needed; this checkpoint was distilled without it.

Requirements

• ComfyUI (recent build with Krea 2 support)

• RealRebelAI/ComfyUI-GGUF_KREA-2 custom node

• Qwen3-VL 4B text encoder + Qwen-image VAE (from Comfy-Org/Krea-2)

This repo contains the diffusion model (transformer) only — it will not generate anything on its own without the VAE and text encoder above.

License & attribution

These weights are a derivative of Krea 2 Turbo © Krea.ai, Inc., and remain governed by the Krea 2 Community License Agreement and the Krea Acceptable Use Policy — exactly as the original model. See the license: LICENSE.pdf. In the event of any conflict, the Krea Acceptable Use Policy and Krea 2 Community License control.

Deployer obligation: if you serve or redistribute this model, the license requires you to implement content-filtering or equivalent review to prevent generation or distribution of unlawful or policy-violating content. Report harmful or illegal outputs to [email protected].

Upstream:

• Base model: krea/Krea-2-Turbo (distilled from krea/Krea-2-Raw)

• ComfyUI repack (VAE / text encoder): Comfy-Org/Krea-2

• GGUF tooling based on city96's ComfyUI-GGUF

The merge and quantization are RealRebelAI's work; the underlying model rights belong to Krea.ai, Inc.

Credits

Merged, quantized, and released by RealRebelAI.

• GitHub: github.com/RealRebelAI

• Hugging Face: huggingface.co/realrebelai

If these are useful, a like on the repo helps others find them. Issues with loading go on the ComfyUI-GGUF_KREA-2 tracker.