Krea 2 — INT8 ConvRot

INT8 ConvRot quantized weights for Krea 2 (K2), for fast, low-VRAM inference in ComfyUI via the ComfyUI-INT8-Fast node.

This is a modified (quantized) version of the Krea 2 model. It is not an official Krea release and is not endorsed by Krea. The original weights are © Krea, licensed under the Krea 2 Community License.

ConvRot is a near-lossless INT8 scheme (~GGUF-Q8 quality) that runs on the INT8 tensor cores of any NVIDIA GPU with sufficient INT8 TOPS (RTX 30-series and up). On a 3090, INT8 is meaningfully faster than FP8 (which has no tensor-core acceleration on Ampere) and roughly half the VRAM of BF16.

Models

File	Precision	Size	Use
`Krea2-Turbo-int8-ConvRot.safetensors`	INT8 ConvRot	14.1 GB	8-step distilled, fast text-to-image
`Krea2-Raw-int8-ConvRot.safetensors`	INT8 ConvRot	14.1 GB	Undistilled base — fine-tuning / LoRA training / research

Original BF16 checkpoints are ~26.6 GB each. Quantized with the K2 profile: the 28 main DiT blocks are INT8, while the sensitive layers (first, last, tmlp, tproj, txtfusion, txtmlp) are kept in high precision.

Requirements

ComfyUI ≥ 0.25.0 (native Krea 2 support).
ComfyUI-INT8-Fast custom node (with a krea2 model-type profile — see note below).
Text encoder: qwen3vl_4b_fp8_scaled.safetensors → ComfyUI/models/text_encoders/ (from Comfy-Org/Qwen3-VL).
VAE: qwen_image_vae.safetensors → ComfyUI/models/vae/ (from Comfy-Org/Qwen-Image_ComfyUI).
Put the INT8 .safetensors in ComfyUI/models/diffusion_models/.

Note on the krea2 profile: Krea 2 is newer than INT8-Fast's built-in model list. These weights were produced with a small added krea2 exclusion profile (keep first/last/tmlp/tproj/txtfusion/txtmlp in high precision). Loading the pre-quantized files here does not require that profile — ConvRot metadata is embedded per-layer — but reproducing the conversion does.

Usage

Load with Load Diffusion Model INT8 (W8A8) (OTUNetLoaderW8A8):

unet_name: the INT8 file
on_the_fly_quantization: false (already quantized)
enable_convrot: true
model_type: krea2

Then the standard K2 graph: CLIPLoader (type: krea2) → VAELoader → CLIPTextEncode → KSampler → VAEDecode.

Recommended sampler settings:

Turbo: 8 steps, CFG 1.0, euler / simple, shift 1.15 (model default).
Raw: ~52 steps, CFG ~3.5, euler / simple (undistilled base; mainly for training).

Quantization

Method: INT8 ConvRot (row-wise INT8 with convolutional rotation), via ComfyUI-INT8-Fast.
Compute: torch._int_mm on INT8 tensor cores; Triton kernels.
Quality: near-lossless vs BF16/FP8 in testing (verified by generation).

License & attribution

These files are a quantized derivative of Krea 2; the Krea 2 model name is retained per the license. Commercial use is permitted only for entities under $1M USD trailing-twelve-month revenue; above that an Enterprise License from Krea is required. Deployers must implement reasonable content-filtering (e.g. NudeNet, Falconsai/nsfw_image_detection, Hive, or human review) to prevent prohibited content, and disclose AI-generated outputs where required.

Credits: Krea (base model) · BobJohnson24/ComfyUI-INT8-Fast (INT8 ConvRot method).

Downloads last month: -; Downloads are not tracked for this model. How to track