I notices these things in ComfyUI Desktop (I use 3080 10GB with FLUX_dev_Q8 + t5xxl_fp16.safetensors + ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors + windows 10):
- Something works different regarding Dual Clip Loader (for FLUX). In standard ComfyUI it works great as it is loaded partialy and use only up to 9GB (important that it is under 10GB VRAM for me) but the same workload for Desktop Comfy uses like 10,7GB VRAM and it shows around 1GB of shared memory). And because this, it works very very slow before it even start to generate images. This is worst blocker for me to use Desktop Comfy, instead of standard one.
- Not every file from old comfy can be selected in new Comfy nodes. I mean I discovered in standard Comfy that there is a possibility to download some model/clip files from manager (yes I did it manually before and put them manually to clip and unet folders etc
). I also noticed in Manager default locations that some files are supposed to be placed in other less standard folders like “diffusion_models” instead of “Unet” and in “text_encoders” instead of “clip” folder. That is why I moved these FLUX files also to these 2 folders, to see them installed in manager (otherwise it shows as not installed). Then I installed desktop Comfy which used models and resources from standard Comfy (which is actually amazing). But unfortunately it does not show for some reason these files from “diffusion_models” and “text_encoders” and I had to move these files again to “UNET” and “Clip” folders. It would be great if the model/clip files was shared fully. - Unfortunately the performance of desktop Comfy compared to standard Comfy is worse, even not counting this clip loading problem from point 1). In the case of generating image (partially loaded model) the VRAM usage is more or less similar to standard Comfy but no matter if I use bundled version of PyTorch or nightly or if I use xformers (which is in my case always faster in both versions of comfy comparing to PyTorch by 5-15%) generation speed in standard comfy is faster then in Desktop Comfy.
- I have error during first run:
“:\Users\Geralt\AppData\Local\Programs@comfyorgcomfyui-electron\resources\ComfyUI\custom_nodes\ComfyUI-GGUF\nodes.py:79: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\utils\tensor_numpy.cpp:209.)
torch_tensor = torch.from_numpy(tensor.data) # mmap”
I checked file looks like not read only. - In one moment it showed update window and after this ComfyUI was broken and need to reinstall like several times to make it work, uninstaller also did not work I just needed to delete one folder to install it again.
- it would be nice to have possibility to copy text from “Logs”. It is impossible now and i need to find log file on HDD instead of just copy it.
But Desktop comfy has future potential as:
- It is easier to install and set up
- it looks nice
- it is separate from standard web browser (even if it is still like standalone browser ) and has some nice additional options in settings