[Bug]: 1.9.3 vs 1.8.0 Inpainting Differences?

[onewhoflies]

Describe the bug

When using the exact same settings (seed, positive/negative prompt, sampler, etc.) on A111 1.8.0 and 1.9.3, inpainting a person using person_yolov8n-seg.pt in ADetailer 1st then inpainting a face using face_yolov8n.pt in ADetailer 2nd gives different results. I have already confirmed that 1.8.0 and 1.9.3 generate the same images with ADetailer disabled to try to isolate the issue.

Steps to reproduce

(detailing what I did, everything may not be necessary to see the issue)

1. In txt2img, generate an image and use Hires Fix (2x, latent nearest-exact upscaler, denoising strength 0.55) in A111 1.8.0 with the DPM++ 3M SDE Karras sampler (on A111 1.8.0 - https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.8.0), with ADetailer 1st and 2nd enabled using the person_yolov8n-seg and face_yolov8n.pt models, leaving the ADetailer 1st prompt empty and a basic prompt describing the face in ADetailer 2nd.
2. Repeat step 1, but use A111 1.9.3 - https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.9.3

If any more info would be useful, I can provide it. Thanks for making such a useful extension.

Screenshots

No response

Console logs, from start to end.

venv "I:\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.8.0
Commit hash: bef51aed032c0aaa5cfd80445bc4cf0d85b408b5
Launching Web UI with arguments:
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
[-] ADetailer initialized. version: 24.4.2, num models: 20
ControlNet preprocessor location: I:\stable-diffusion-webui\extensions\sd-webui-controlnet\annotator\downloads
2024-05-01 22:31:46,791 - ControlNet - INFO - ControlNet v1.1.446
2024-05-01 22:31:46,844 - ControlNet - INFO - ControlNet v1.1.446
[sd-webui-freeu] Controlnet support: *enabled*
Loading weights [62bb78983a] from I:\stable-diffusion-webui\models\Stable-diffusion\epicphotogasm_lastUnicorn.safetensors
Creating model from config: I:\stable-diffusion-webui\configs\v1-inference.yaml
2024-05-01 22:31:47,762 - ControlNet - INFO - ControlNet UI callback registered.
Running on local URL:  http://127.0.0.1:7860

To create a public link, set `share=True` in `launch()`.
Startup time: 11.8s (prepare environment: 5.7s, import torch: 1.8s, import gradio: 0.6s, setup paths: 0.4s, initialize shared: 0.2s, other imports: 0.3s, load scripts: 1.7s, create ui: 1.0s, gradio launch: 0.1s).
Loading VAE weights specified in settings: I:\stable-diffusion-webui\models\VAE\realisticVisionV20_v20.safetensors
Applying attention optimization: sdp-no-mem... done.
Model loaded in 1.6s (load weights from disk: 0.2s, create model: 0.5s, apply weights to model: 0.7s).
100%|██████████████████████████████████████████████████████████████████████████████████| 30/30 [00:03<00:00,  9.35it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 10/10 [00:04<00:00,  2.12it/s]
postprocess_batch

0: 480x640 1 person, 67.0ms
Speed: 1.0ms preprocess, 67.0ms inference, 9.0ms postprocess per image at shape (1, 3, 480, 640)
100%|██████████████████████████████████████████████████████████████████████████████████| 13/13 [00:01<00:00,  7.61it/s]

0: 480x640 1 face, 4.0ms
Speed: 1.0ms preprocess, 4.0ms inference, 1.0ms postprocess per image at shape (1, 3, 480, 640)
100%|██████████████████████████████████████████████████████████████████████████████████| 14/14 [00:02<00:00,  6.34it/s]


(^ looks the same for 1.9.3)

List of installed extensions

No response

[dosubot]

The differences you're observing between versions 1.8.0 and 1.9.3 of the software when using person_yolov8n-seg.pt and face_yolov8n.pt for inpainting could be attributed to several factors, including changes in the inpainting algorithms, model handling, or related dependencies across these versions. To address this issue:

1. Review Change Logs: Check the release notes for both versions to identify any modifications in the inpainting process or model handling that might explain the differences in results.

2. Ensure Consistent Configuration: Double-check that all settings, including model parameters and inpainting configurations, are identical across both versions.

3. Dependencies and Environment: Make sure that the dependencies are consistent across versions, as changes in library versions or the environment can lead to variations in results.

4. Model Version Compatibility: Confirm that the models are compatible with both versions of the software. Updates to the models or their interpretation by the software could result in different outcomes.

5. Consider Software Updates: Be aware that updates might include bug fixes or improvements that could alter inpainting results, even with identical settings.

Focusing on these areas should help in troubleshooting and achieving more consistent inpainting results across different software versions.

Details

Sources

• CHANGELOG.md
• scripts/!adetailer.py
• controlnet_ext/controlnet_ext.py

About Dosu
This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

To continue the conversation, mention @dosu.

[Bing-su]

Probably an undocumented seed breaking change, or maybe an effect of the sampler change in 1.9.0.

Adetailer doesn't do anything about it.