Very high VRAM + RAM usage with XL models during highres fix or adetailer processing #13187
Replies: 1 comment
-
I have noticed something similar when running ADetailer (face_yolov8s), even though I'm using SD.next. I have 16Gb VRAM, generating a batch of 4 or even 6 images (768x1152) works fast and fits neatly into the VRAM, but the moment the ADetailer starts, iteration speed drops and task manager shows it's using "shared GPU memory" (i.e. RAM).
in any case, it seems an ADetailer issue since I'm using a different UI and having the same problem. note I'm not using HiRes/refine pass. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Looking for tips about this one, anything appreciated.
Even on single image batches, when basic generation steps are complete and highres or adetailer starts, vram raises from about 10-11 GB with an XL model to full VRAM (16 GB) + full shared memory (16 GB in ram - at this point total ram used is close to 80%) for a minute or more, slowing down generation a lot. Then when progress bar of each appears in the console after a good minute, the extra memory use is cleared.
For highres fix: usually going with 4xUltrasharp, 832x1216 or the opposite, 1.5x upscale (=1248x1824)
For adetailer: face_yolov8n or mediapipe_face_full (the latter seem a bit better)
Note it does that on SD 1.5 models, just much less noticeable.
Startup options:
--xformers --no-half-vae --gradio-img2img-tool color-sketch --api --update-check
Example generation settings:
Typical generation speeds:
A1111 1.6.0 on a RT 4060 Ti (16 GB VRAM) / driver version 537.13 / 32 GB RAM
Are these issues related ? #6307 #6725 #12273
I've read somewhere mediapipe is cpu-based no matter what unless you recompile it and i don't think i'm good enough to do that. That wouldn't explain highres though...
Thank you
Beta Was this translation helpful? Give feedback.
All reactions