krea 2 int4 convrot

Krea 2 Raw & Turbo — ConvRot INT4 Quantized (ComfyUI)

Overview

This is a ConvRot INT4 quantized version of both Krea 2 Raw and Krea 2 Turbo , converted using the Starnodes Ultimate Model Converter (built on comfy-kitchen).

Base models: Krea 2 Raw (krea/Krea-2-Raw) and Krea 2 Turbo (krea/Krea-2-Turbo)

Quantization format: INT4 ConvRot (4-bit integer weights + group-wise Hadamard rotation, rotation group size 256 / quantization group size 64)

Quantization tool: comfyui-starnodes-modelconverter (comfy-kitchen based)

Target hardware: SM 8.0+ (Ampere/Ada/Blackwell). Tested on an RTX 3060 12GB.

ComfyUI loader: Load Diffusion Model node (or an equivalent ConvRot INT4-compatible loader)

Which one to use:

Raw — the undistilled base checkpoint. Best for LoRA training and fine-tuning; more diverse/malleable outputs.

Turbo — the 8-step distilled checkpoint. Best for fast inference; LoRAs trained on Raw transfer well to Turbo for generation.

This is a Derivative of the Krea 2 Model distributed by Krea.ai . It is not endorsed, verified, or supported by Krea.ai .

License (please read before downloading)

This model is distributed under the Krea 2 Community License Agreement . Quantizing the weights into ConvRot format constitutes a modification of the original model, so this release carries forward the same license terms.

Key points

Attribution and notices must be preserved You may not remove, alter, or obscure any copyright notices, license terms, attribution notices, or other proprietary markings contained in the Krea Model. This page and the accompanying files retain those notices.

Commercial use conditions The Krea 2 Community License restricts commercial use above certain organizational thresholds, above which an Enterprise License from Krea is required. Check the current terms directly at krea.ai/krea-2-licensing before any commercial use — this page does not restate the specific thresholds since they are subject to change.

Content filtering obligation Any deployment of this model, or outputs generated from it, must implement reasonable Content Filter measures to prevent the generation of illegal content, non-consensual intimate imagery (NCII), CSAM, or defamatory content. This can include open-source classifiers, commercial moderation APIs, or manual human review. Failure to do so is a breach of the license and gives Krea the right to revoke access or require model updates.

AI-generated content disclosure Where applicable, you must disclose that outputs were generated by AI, per the underlying license terms.

This is an unofficial, community release Not produced, reviewed, or endorsed by Krea.ai . Please direct questions about this specific quantized file to the uploader (me), not Krea.

Technical details

Item Value Source models Krea 2 Raw and Krea 2 Turbo (BF16) Quantization algorithm ConvRot (group-wise Hadamard rotation) + INT4 per-group scaling Rotation group size 256 Quantization group size 64 Weight packing Signed INT8, 2 nibbles per byte Scales FP32, per-group Quantization tool comfyui-starnodes-modelconverter Tested on RTX 3060 12GB / ComfyUI

Usage

Place krea2_raw_convrot_int4.safetensors and/or krea2_turbo_convrot_int4.safetensors in ComfyUI/models/diffusion_models/

Load it with the Load Diffusion Model node (or another ConvRot INT4-compatible loader)

Use the standard Krea 2 Qwen3-VL text encoder as usual

Note: Raw is undistilled and typically needs more sampling steps / higher CFG than Turbo to converge to a clean image — it isn't a drop-in speed replacement for Turbo, it targets training/fine-tuning workflows.

Disclaimer

Provided as-is, with no warranty. The uploader and the quantization tool authors are not responsible for output quality or safety. Users are responsible for complying with the Krea 2 Community License Agreement and all applicable laws.

This page is created and distributed by a third party (nununu568) unaffiliated with Krea.ai .