Run times?
Is 3.5 hours per generation normal? I have 6gb vram with 32gb ram, is anyone else having this problem?
Is your GPU a turing one? Does it even support b16? If not, you should apply this patch fix for fp16 GPU only:
https://huggingface.co/ourtruth/Stuff/blob/main/anima_fp16_patch.py
Note that the patch is for Comfy webui, the instruction is inside the file (open patch file in notepad first).
Is your GPU a turing one? Does it even support b16? If not, you should apply this patch fix for fp16 GPU only:
https://huggingface.co/ourtruth/Stuff/blob/main/anima_fp16_patch.py
Note that the patch is for Comfy webui, the instruction is inside the file (open patch file in notepad first).
It's a navidia geoforce gtx 1060
Is your GPU a turing one? Does it even support b16? If not, you should apply this patch fix for fp16 GPU only:
https://huggingface.co/ourtruth/Stuff/blob/main/anima_fp16_patch.py
Note that the patch is for Comfy webui, the instruction is inside the file (open patch file in notepad first).
It's a navidia geoforce gtx 1060
Okay, that's a pretty old one, you can try that patch file I linked to you and test it whether the speed is increase or not. Edit: Apparently ComfyUI already fixed the patch, you can use the node "ModelComputeType" and set as "fp16".
Unless you mean minutes instead of hours, that's certainly not normal, maybe it's running on CPU instead of GPU?
I have a gtx 970 (4gb vram) and it takes about 30 or 60 seconds per step for a 1-megapixel image, depending on whether using CFG or not. With a slightly lower resolution and a low number of steps it only takes a few minutes per image. Q8 GGUF
Running ComfyUI with the command-line option --fp16-unet solved remaining Anima speed issues for me.
Now < 7.5s/it for 1MP image, so a 75% improvement in speed and faster than SDXL (which unfortunately is slightly slower with this option).
These fixed the issue. I'm now getting one depending steps about every 10-30 minutes, thank you for your help. Now I just need to figure out what combinations of samplers and schedules work. Does anyone know any good ones?