vladmandic / automatic

SD.Next: Advanced Implementation of Stable Diffusion and other Diffusion-based generative image models
https://github.com/vladmandic/automatic
GNU Affero General Public License v3.0
5.32k stars 380 forks source link

[Issue]: Generations break when using refine/force hirez and variation strength slider on SDXL #3182

Closed devils-shadow closed 1 month ago

devils-shadow commented 1 month ago

Issue Description

When using the variation seed slider during any generation with any SDXL/Pony model, the resulting image is broken. Please note that using the variation seed slider on SD1.5 models works as expected. No errors or additional information appears in the console when running with --debug. image

Prerequisites:

Steps to reproduce:

  1. Run SDN
  2. Select any SDXL/Pony model
  3. Set resolution to 768x1280
  4. Set sampling method to euler a
  5. Untick full quality
  6. Generate 1 image
  7. Reuse seed from image generated in step 6
  8. Set variation strength to a value above 0 (tested with 0.3)
  9. Enable second pass
  10. Select RealESRGAN 4x+ Anime6B upscaler
  11. Tick force hirez
  12. Set hirez steps to 10
  13. Attempt to generate an image Notice that the resulting image is broken

Version Platform Description

Python 3.10.11 on Windows
Version: app=sd.next updated=2024-05-29 hash=7a7cea3b branch=master
Platform: arch=AMD64 cpu=AMD64 Family 25 Model 33 Stepping 2, AuthenticAMD system=Windows
release=Windows-10-10.0.22631-SP0 python=3.10.11
nVidia CUDA toolkit detected: nvidia-smi present
Extensions: disabled=['sdnext-modernui']
Extensions: enabled=['Lora', 'sd-extension-chainner', 'sd-extension-system-info', 'sd-webui-agent-scheduler', 'stable-diffusion-webui-rembg'] extensions-builtin
Extensions: enabled=['a1111-sd-webui-tagcomplete', 'adetailer', 'sd-civitai-browser-plus'] extensions
Startup: standard
Load packages: {'torch': '2.3.0+cu121', 'diffusers': '0.28.0', 'gradio': '3.43.2'}
VRAM: Detected=11.99 GB Optimization=none
Engine: backend=Backend.DIFFUSERS compute=cuda device=cuda attention="Scaled-Dot-Product" mode=no_grad
Device: device=NVIDIA GeForce RTX 4070 n=1 arch=sm_90 cap=(8, 9) cuda=12.1 cudnn=8801 driver=552.44

Relevant log output

No response

Backend

Diffusers

Branch

Master

Model

SD-XL

Acknowledgements

vladmandic commented 1 month ago

fixed.