calcuis

/

hunyuanimage-gguf

like 68

use gguf-node (pypi|repo|pack)

btw, it works actually, but the comfyui-gguf node side is taking numpy dequant for i quant gguf, you should expect to wait 5-8 hours or longer for the dequant process; please use the node in reference

use gguf-node (pypi|repo|pack)

I have everything identical to your lite workflow example with hunyuanimage-lite-mxfp4_moe.gguf (downloaded your workflow-hunyuanimage-lite.json). And i get much worse result. What am i doing wrong? 8steps cfg 1.0 and so on, everything identical to your screenshot.

how's you setting? just follow the model card's setting, try lite v2.2, use steps=10, cfg=1.5

it seems the size was set to 1024x1024; should use 2048x2048 for this model, otherwise you will get even worse output

how's you setting? just follow the model card's setting, try lite v2.2, use steps=10, cfg=1.5

I tried 2048 too, not really better... As i said, your lite example screenshot is what i have

tested; we don't have such problem; forget lite v2.0, 2.1, just lite and lite v2.2; try other models, i.e., distilled, standard v2, etc., see will you get the same results or not

tested; we don't have such problem; forget lite v2.0, 2.1, just lite and lite v2.2; try other models, i.e., distilled, standard v2, etc., see will you get the some results or not

first i tried hunyuanimage-lite-iq3_xxs.gguf, then hunyuanimage-lite-mxfp4_moe.gguf, no improvement at all

have you tried the distilled or standard one? or the original safetensors? if those still give you the same output then should be something wrong with your machine settings

I'm using Q3_KS gguf for qwen image, and it generates amazing images, so ... no idea what could be wrong with hunyuan. ComfyUI version: 0.3.59 ComfyUI Revision: 5088 [596049a8] | Released on '2025-09-13'
I guess comfyui must be up to date.
I don't think distilled or standard makes sense, because my pc is not high end. And according to your screenshots, lite version should be more than enough.

don't understand your problem if no more info; guess you just try the low tier quants then should do a rough comparison very easy and troubleshoot it yourself