SDXL models seem to reload on every generation

[begin2]

Expected Behavior

The second generation should be fast (around 12 seconds) because the model should already be loaded in memory (hot start).

Actual Behavior

The generation takes noticeably longer (around 22 seconds). It feels like the model is loading from the disk every single time I click the button.

Steps to Reproduce

1. Run SwarmUI v0.9.8 Beta on default settings.
2. Generate an image (takes about 12 seconds as usual).
3. Immediately click Generate again without changing any settings.

Debug Logs

In the logs, I see the model loading every time, even for the second generation:
[ComfyUI-0/STDERR] Requested to load SDXLClipModel
[ComfyUI-0/STDERR] loaded completely; ... full load: True
[ComfyUI-0/STDERR] Requested to load SDXL
[ComfyUI-0/STDERR] loaded completely; ... full load: True

Other

my setup: RTX 5090, Windows 11

Any help would be appreciated, thanks!

[mcmonkey4eva]

(1) that's definitely not the full debug logs
(2) That's unusual behavior. It shouldn't be doing that, and in my testing it isn't doing that.
(3) this comfy PR Comfy-Org/ComfyUI#13802 made significantly recent changes to loading behavior, so it's possible it is misbehaving on your system. Try reverting comfy to a commit older than that (make sure you don't have auto update enabled in server->backends) to see if it changes behavior

[begin2]

thanx for superfast checking :) (as always :) )

[begin2]

Thank you again for the quick and accurate response!

I reverted the ComfyUI backend to a commit from May 20, 2026 (prior to PR #13802), and the issue is completely resolved. Models now stay loaded in VRAM between generations, and generation times are back to normal (~12s instead of 22s). The Requested to load SDXL message no longer appears on subsequent requests.

Thanks again for your fast support and for pointing me in the right direction!