Comfy INT8 Acceleration
This node speeds up Flux2, Ideogram4, Chroma, Z-Image, Ernie Image in ComfyUI by using INT8 quantization, delivering between 1.5~2x faster inference on my 3090 depending on the model. It should work on any NVIDIA GPU with enough INT8 TOPS. It appears to be faster than FP8 on 40-Series and above as well. Works with lora, torch compile.
Further Reading:
Quality Metrics comparing against MXFP8, FP8, GGUF, etc.
List of Prequantized Checkpoints
Updates:
2026-06-06:
Fixes for 20-series GPUs
Ensuring proper handling of static weights when dynamic is deactivated
2026-24-05:
RAM usage for lora loading is fixed and on par with base comfy.
RAM usage for model loading is fixed.
Only thing that remains is on the fly quantization will create an extra int8 copy in memory, but it is too much of a hassle to work around. Please rely on swap or pre converted models if this is an issue.
Fixed an issue with loading loras on models that include .bias layers (WAN, LTX2.X) which would cause a OOM error.
2026-15-05:
Bringing back stochastic lora. Some loras appear to need it, others don't, try it if your lora is not working and you don't like pre-lora. TLDR is "sometimes it really helps, sometimes its a little worse". See our measurements here.
Attempt at reducing RAM usage
Fixed an issue with Pre-Lora crashing on windows
2026-10-05:
Overhauled the entire lora system. Normal lora loader node works now, no need for specialized lora loaders.
Converted QuaRot to ConvRot, which is a small but free quality gain.
Added Pre-Lora node, which you can connect to the INT8 Model loader to merge loras before utilizing on the fly quantization.
For more info on quality of convrot, lora approaches see the Metrics
Common GPU related issues:
RTX 20-Series will require you to either use Triton-Windows on windows, triton==3.2.0 or compile triton yourself with SM75 support which was dropped in 3.3.0.
A100 has no possible INT8 Speed-up https://github.com/BobJohnson24/ComfyUI-INT8-Fast/issues/71
FAQ:
Q: How do I quantize myself?
A: It is not recommended to quantize the human existence. If you would like to quantize a model, see example_workflows/int8_save_convrot_model.json
Q: What is ConvRot?
A: ConvRot is a variant of QuaRot. It basically rotates model weights and activations to eliminate outliers before quantization. This has some inference overhead, but is generally a large quality boost.
Q: What is Pre-Lora?
A: Pre-Lora is a way to merge the lora weights to a BF16 checkpoint within ComfyUI before you quantize the model. This requires an unquantized base model, and enabling on-the-fly quantization. It is generally a higher quality way to apply a lora.
Q: Torch compile takes forever and I hate it
A: Use the torch compile node from KJ Nodes and ensure you set the disable dynamic VRAM toggle.
Requirements:
Working ComfyKitchen (needs latest comfy and possibly pytorch with cu130)
Triton
Windows untested, but I hear triton-windows exists.
Credits:
dxqb for the entirety of the INT8 code during the very early versions of this node, it would have been impossible without them:
https://github.com/Nerogar/OneTrainer/pull/1034
If you have a 30-Series GPU, OneTrainer is also the fastest current lora trainer thanks to this. Please go check them out!!
newgrit1004 for the base ConvRot code we modified into proper ConvRot
https://github.com/newgrit1004/ComfyUI-ZImage-Triton
silveroxides for providing a base to hack the INT8 conversion code onto.
https://github.com/silveroxides/convert_to_quant
Also silveroxides for showing how to properly register new data types to comfy
https://github.com/silveroxides/ComfyUI-QuantOps