25/06/2026 UPDATE - Replaced the models with versions that work with the new native implementation of INT8 in comfyui, for use with the native Load Diffusion Model node, Make sure you update you ComfyUI (Still works with the custom node but using native loads the model a bit faster for me)
Files gathered from https://huggingface.co/bertbobson/Ideogram-4-INT8-ConvRot
This is an unavoidable double quantization due to the release state of Ideogram4.
The FP8 weights were cast to FP32 with the FP8 scales, then downcast to BF16 before being converted to INT8.
For use in ComfyUI with https://github.com/BobJohnson24/ComfyUI-INT8-Fast
Speed is 1.78x faster(2.03s/it) than FP8(3.62s/it) on a 3090, without compile.
4.4-6.2s/it on my 3060 12gb
Description
Unconditional model for use as the model negative
FAQ
Comments (5)
noob question but i am on a 3070ti, should i be using INT8?
yes preferably, only if you need full precision go for full model. INT8 is for 30`s
I got a pixelated image instead. even though I'm already using the INT8 node. What should I do?
Working ComfyKitchen (needs latest comfy and possibly pytorch with cu130)
Triton for linux or triton-windows https://github.com/woct0rdho/triton-windows/releases/tag/v3.6.0-windows.post24
@tsolful Thank you so much