obsxrver/ComfyUI-Native-INT8_ConvRot

ComfyUI-Native-Int8-ConvRot

INT8 ConvRot models, converted to the native quantization format ComfyUI expects.

INT8 ConvRot currently offers one of the best quality-to-performance ratios of any quantization method. In my personal experience, INT8 ConvRot models provide quality close to BF16 at generation speeds matching or beating FP8_Scaled.

"INT8 ConvRot is row-wise INT8 with parameters and activations rotated before quantization via ConvRot." — ComfyUI-INT8-Fast Metrics.md

Models

If there is a model that you would like that is not in this list, make a request

File	Model	Original Creator's Repo
checkpoints/hidream_o1_image_int8-convrot.safetensors	HiDream-O1-Image	HiDream-ai/HiDream-O1-Image
checkpoints/hidream_o1_image_dev_int8-convrot.safetensors	HiDream-O1-Image-Dev	HiDream-ai/HiDream-O1-Image-Dev
checkpoints/ltx-2.3-22b-dev-int8-ConvRot.safetensors	LTX-2.3 22B Dev	Lightricks/LTX-2.3
checkpoints/ltx-2.3-22b-distilled-1.1-int8-ConvRot.safetensors	LTX-2.3 22B Distilled v1.1	Lightricks/LTX-2.3
checkpoints/sulphur_dev_INT8_ConvRot.safetensors	Sulphur 2 Dev	SulphurAI/Sulphur-2-base
checkpoints/sulphur_distill_INT8_ConvRot.safetensors	Sulphur 2 Distill	SulphurAI/Sulphur-2-base
diffusion_models/anima-preview3-base-int8-ConvRot.safetensors	Anima Preview 3 (Base)	circlestone-labs/Anima
diffusion_models/flux-2-klein-9b_int8_convrot.safetensors	FLUX.2 [klein] 9B	black-forest-labs/FLUX.2-klein-9B
diffusion_models/Krea2-Turbo-int8-ConvRot.safetensors	Krea 2 Turbo	krea/Krea-2-Turbo
diffusion_models/qwen-image-2512-int8-ConvRot.safetensors	Qwen-Image-2512	Qwen/Qwen-Image-2512
diffusion_models/wan2.2_i2v_high_int8_convrot.safetensors	Wan2.2 I2V A14B (High Noise)	Wan-AI/Wan2.2-I2V-A14B
diffusion_models/wan2.2_i2v_low_int8_convrot.safetensors	Wan2.2 I2V A14B (Low Noise)	Wan-AI/Wan2.2-I2V-A14B
diffusion_models/wan2.2_t2v_high_int8_ConvRot.safetensors	Wan2.2 T2V A14B (High Noise)	Wan-AI/Wan2.2-T2V-A14B
diffusion_models/wan2.2_t2v_low_int8_ConvRot.safetensors	Wan2.2 T2V A14B (Low Noise)	Wan-AI/Wan2.2-T2V-A14B
diffusion_models/z_image_turbo_int8_convrot.safetensors	Z-Image Turbo	Tongyi-MAI/Z-Image-Turbo

Quality Ranking

Per the latent-divergence benchmarks in Metrics.md:

GGUF Q8 > INT8 ConvRot > MXFP8 > FP8 >= INT8 Row > INT8 Tensorwise

Note: this is the general takeaway across all tested models. In several individual benchmarks (e.g. Anima, Flux2 Klein 9B, Qwen Image 2512), INT8 ConvRot actually scored better than GGUF Q8.

Requirements

• A ComfyUI version that includes native INT8 support (Comfy-Org/ComfyUI#14636, merged June 2026). Update if your loader reports an invalid quantization type.
• Models load with the standard Load Diffusion Model node — no custom node needed.

How to Quantize a Model to INT8 ConvRot

1. Install silveroxides' convert_to_quant:

   pip install -U convert-to-quant
   

   INT8 kernels require Triton (native on Linux; use triton-windows on Windows). PyTorch must be installed separately with the correct CUDA version.

2. Convert the model:

   ctq -i source_model_bf16.safetensors -o converted_model_int8_convrot.safetensors \
     --int8 --scaling_mode row --simple --convrot --convrot-group-size [64,256,1024] \
     --comfy_quant --save-quant-metadata --<model-arch-flag>
   

Notes

• --convrot-group-size accepts 64, 256, or 1024. It is recommended to choose a value that divides evenly into all of the model's layer dimensions.
• --<model-arch-flag> selects the layer-exclusion preset for your model architecture (e.g. --wan, --flux2, --zimage). Run ctq --help-filters (or ctq -hf) for the full list.

References

1. Reddit: "So is INT8-ConvRot the new hot thing?"
2. ComfyUI-INT8-Fast — Metrics.md (benchmark methodology & full tables)
3. Comfy-Org/Boogu-Image discussion #10
4. ComfyUI PR #14636 — Support int8 models
5. bertbobson/ComfyUI-INT8_ConvRot
6. silveroxides/convert_to_quant

Licensing and Commercial Use

All models are distributed strictly under their upstream license with no added restrictions. Verify permissibility with the original creator before commercial use.