Open INTstinkt opened 7 months ago
A responsive free edit UI probably requires a pure front-end solution instead of gradio solution we have now.
A responsive free edit UI probably requires a pure front-end solution instead of gradio solution we have now.
Hi guys! I found controlnet's preprocessor may take very long time(20s+) to preprocess。I can't figure out what cause it. Wound you mind to give some advice?
the input_image shape is (1152, 1152, 3) and threshold_a = -1/threshold_b = -1
A responsive free edit UI probably requires a pure front-end solution instead of gradio solution we have now.
Hi guys! I found controlnet's preprocessor may take very long time(20s+) to preprocess。I can't figure out what cause it. Wound you mind to give some advice?
the input_image shape is (1152, 1152, 3) and threshold_a = -1/threshold_b = -1
What preprocessor are you using?
A responsive free edit UI probably requires a pure front-end solution instead of gradio solution we have now.
Hi guys! I found controlnet's preprocessor may take very long time(20s+) to preprocess。I can't figure out what cause it. Wound you mind to give some advice? the input_image shape is (1152, 1152, 3) and threshold_a = -1/threshold_b = -1
What preprocessor are you using?
It's "ip-adapter".
SDXL ipadapter uses clip-g as preprocessor which is slower than clip-h. You can try switch to ip-adapter_sdxl_vit-h
which uses clip-h as preprocessor.
SDXL ipadapter uses clip-g as preprocessor which is slower than clip-h. You can try switch to
ip-adapter_sdxl_vit-h
which uses clip-h as preprocessor.
Besides I found clip would init everytime. why chang it (load/unload)? Can I keep it loaded in clipe_encoder(never unloaded)? whether can I keep it loaded in clip_encoder?
SDXL ipadapter uses clip-g as preprocessor which is slower than clip-h. You can try switch to
ip-adapter_sdxl_vit-h
which uses clip-h as preprocessor.Besides I found clip would init everytime. why chang it (load/unload)? Can I keep it loaded in clipe_encoder(never unloaded)? whether can I keep it loaded in clip_encoder?
The behaviour changed in https://github.com/Mikubill/sd-webui-controlnet/pull/2754 when we ported forge's preprocessor structure, as forge aggressively unloads the preprocessor torch module after use (To save vram?).
If you think this behaviour is undesired, I can try restore previous behaviour.
SDXL ipadapter uses clip-g as preprocessor which is slower than clip-h. You can try switch to
ip-adapter_sdxl_vit-h
which uses clip-h as preprocessor.Besides I found clip would init everytime. why chang it (load/unload)? Can I keep it loaded in clipe_encoder(never unloaded)? whether can I keep it loaded in clip_encoder?
The behaviour changed in #2754 when we ported forge's preprocessor structure, as forge aggressively unloads the preprocessor torch module after use (To save vram?).
If you think this behaviour is undesired, I can try restore previous behaviour.
Thank you very much for your prompt and kind response!I want to keep the the instance of ClipVisionDetector in clip_encoder as I found it will cost extra time due to Initialize ClipVisionDetector with config(clip_g, clip_h, clip_vitl) everytime. But I'm not sure if this will cause any problems. Therefore, I would like to seek your valuable advice. Never unload
Is there an existing issue for this?
What happened?
If you manually type in the Weights for the ne Weight Type in IP Adapter XL to fast Controlnet doesn't get the changes and the Chart disappears. If this happens it actually dont register the Inputs and the Output picture is the same as if you didnt manually change the weight. It only happens for me if i type in the manual weights to fast. Little Workaround i found is deleting one diget, wait 1-2 sec then replacing it with a new digit till the chart is loaded back up, at this point it registered the input
Steps to reproduce the problem
What should have happened?
Registered the weight change and apply them to generation
Commit where the problem happens
webui: version: v1.9.0 • python: 3.10.11 • torch: 2.1.2+cu121 • xformers: 0.0.23.post1 • gradio: 3.41.2 • checkpoint: c9e3e68f89 controlnet: V1.1.445 IP Adapter Preprocessor: auto Model : Every Model
What browsers do you use to access the UI ?
Google Chrome
Command Line Arguments
List of enabled extensions
Console logs
Additional information
Consol log kind of irrelevant for this one but the ones actually showing the weighting in the logs are the ones i waited between replacing the weights till the chart showed up again