comfyui?
Hi, I noticed that you posted an fp8 scaled model of qwen image 2512 without step distillation and I was thrilled to try it out in comfyui but it came out with only noise. Could we perhaps see a comfy version of this one too? (like with the four steps one). The default fp8 released by comfy team and unsloth team was not scaled, this results in loras trained in bf16 original model creating grid artifacts when used with these fp8 models. While an fp8 scaled model might be the solution to said problem, one can only know if they're able to test it, so I look forward to any aid available from you 😄
And thank you for all your hard work, much appreciated!
qwen_image_2512_fp8_e4m3fn_scaled_8steps_v1.0.safetensors和qwen_image_edit_2511_fp8_e4m3fn_scaled_8steps_v1.0.safetensors都是存在同样的问题，使用现有comfyUI社区的工作流出来的都是雪花图，无法正常成像。无论怎么调整参数都一样！官方最近发布的了这两个模型，难道内部就没有测试过吗？就是为了发布而发布？？？？
comfyui无法使用的模型请注明，我已经不知道下错多少次了，qwen_image_2512_fp8_e4m3fn_scaled_8steps_v1.0.safetensors 这个模型comfyui只能出雪花图。。。。。
I converted fp8 scaled model for comfyui https://huggingface.co/ingtellect/qwen-image-2512-fp8-scaled-comfyui
I converted fp8 scaled model for comfyui https://huggingface.co/ingtellect/qwen-image-2512-fp8-scaled-comfyui
Thanks, it worked.
I converted fp8 scaled model for comfyui https://huggingface.co/ingtellect/qwen-image-2512-fp8-scaled-comfyui
This worked for me, too. The one on ComfyUI's 2512 workflow page doesn't work with a 3090 for some reason. Maybe something breaks in the upconvert to fp16 or something?
But your model fixes everything!