JorgeR81

When I use Q8, I don't have RAM spikes while loading. It never goes above 32 GB. But I never tried to use it after FP8.

I've just updated. For me Flux fp8 is still slower than Flux fp16. I'm on Windows 10 GTX 1070 ( 8GB VRAM ) 32 GB RAM I'm using the UNET...

I updated all via manager. --- > Don't know if it's related to the new UI change, the main difference in the log seems to be "[rgthree] NOTE: Will NOT...

> And some older generation GPUs have half speed fp16 operations, which these weights are likely being cast to and/or calculated at. I have a Pascal GPU, so this commit...

The Unet loader has several benefits: - The Unet loader allows to use the ( larger ) fp16 versions of the Flux models and the T5 text encoder, which give...

Oh, right ! Perhaps the Comfy UI team wants to wait and see if @cubiq is going to release it's custom nodes for this, like on previous models. I'm OK...

The [ComfyUI-PhotoMaker-Plus](https://github.com/shiimizu/ComfyUI-PhotoMaker-Plus) node, mentioned above, also supports, PhotoMaker-V2.

> If you're running it using `run_nvidia_gpu.bat`, you shouldn't use pip directly. > You have to execute `pip` via `python_embeded\python.exe -m pip` Exactly ! That should be in the readme.

Can you use other big loras, like this one ( 1.28 GB )  https://civitai.com/models/641309/formcorrector-anatomic?modelVersionId=717317 I can use it with Flux FP8 and  `--reserve-vram 1.2` Or with GGUF models, without any...

Just tested: - GTX 1070 ( 8GB ) - 32 GB VRAM - Windows 10 `Hyper-FLUX.1-dev-8steps-lora.safetensors` works for me in Flux FP8 and `--reserve-vram 1.2`