Closed zaxwashere closed 3 days ago
would be great if you could reproduce with env variable SD_LORA_DEBUG=true
set
no lora lora on lora removed
added the env variable, ran it again with --safe to disable all extensions. Let me know if you need anything else
i'm surprised this was not reported long time ago since this logic bug in the code existed for a while. basically, if there were no loras required, deactivate function disabled patched methods so they would not actually execute to restore weights from backup. anyhow, fixed.
Issue Description
Style loras for ponyxl models seem to persist until a model reload is applied. I've attached a log on a fresh venv with most addons disabled. I have also attached 3 pngs to show the issue visually.
This occurs at any lora strength. For ease of confirmation, I set a minimalist artstyle lora at 1.5 strength. This still occurs at 0.4.
Steps: Load model input base prompt, generate add lora to prompt, generate 2nd image delete lora from prompt, generate 3rd image.
The 3rd image will still show the lora's effects. I've tried disabling prompt cache, disabling timesteps, different models, different loras. The issue persists until i change models over.
attached are 3 images, and my config.json in case that's helpful. Thanks! copied_config.json
No lora, 1st generation Lora on, 2nd generation Lora removed, 3rd generation
Version Platform Description
Windows 11, 23H2 amd 5700x3d 64 gb drr4, 3600mhz rtx 3060 12 gb, driver 565.90 python 3.10.6 SDNEXT dev branch, db3e03cb System packages: {'torch': '2.4.1+cu124', 'diffusers': '0.31.0.dev0', 'gradio': '3.43.2', 'transformers': '4.45.2', 'accelerate': '1.0.0'}
Relevant log output
Backend
Diffusers
UI
Standard
Branch
Dev
Model
StableDiffusion XL
Acknowledgements