Q: Why do this?
A: Because the midjourney-generated characters have a kind of oil painting texture, and they need to be fixed with sd every time. So I developed a workflow to fix them directly.
Q: Why isn't finger repair good?
A: Because this process only focuses on realism, it doesn't handle making unreasonable things reasonable. In other words, it can't fix a distorted finger back to normal, it can only add blood vessels to the original hand.
-
Recommended big model: majic V7
-
Recommended Vae: Realistic difConsistency Negative (Pack)
-
Recommended high-definition repair node. If you want the image to be more realistic, even including pore details, please use this scaling node. Link
My nodes use G-DinoSAM to recognize objects, not just redraw faces and such. You can freely copy and paste, freely combine them.
-
If the whole head is recognized when identifying hair, please adjust the threshold parameters in G-DinoSAM.
-
After all, it's automation. Please don't expect too much. I can only try to replicate manual operations, and post-processing in Photoshop is unavoidable.
-
If there is anything that is not user-friendly or any suggestions, please leave a comment actively.
- wd-tagger on face.
- Optimization of update parameters.
- Optimization of update parameters.
- Change: Updated Yoloworld ESAM to G-DinoSAM.
- Update: Brand new layout.
- Update: Brand new parameters for smoother and softer results.
- Addition: Added mask softening feature.
- Addition: Added DA deep preprocessor, anyline preprocessor, and openpose preprocessor.
- Fix: Previous approach was incorrect; cutting through masks did not improve results and slowed down performance. Changed method to expand mask range.
- Fix: It seems that tile is not necessary for HD restoration, and using tile can cause skin darkening issues.
- Fix: Adjusted cropping size and mask blur, removed mask overlay for facial restoration.
- Addition: Quick switch to directly close unnecessary node groups.
- Fix: After modifying parameters, characters turned black but had more details, not sure why.
- Known issues: Cannot correctly identify arms and thighs.
- Fix: Previous nodes failed to detect the neck effectively, causing neck errors. Mask inter-cutting now meets neck scene requirements, and all 30 tested images worked normally.
- Fix: Previous nodes were ineffective for clothing, possibly detecting pants, etc. Mask cutting only meets clothing detection scenes, and 30 tested images worked normally.
- Fix: Changed keywords and settings for more accurate pants recognition.
- Deprecated: High-frequency and low-frequency nodes officially deprecated.
- Fix: Seemingly fixed the issue of face darkening.
- Fix: High-frequency and low-frequency modules are apparently unnecessary, so images can be obtained in one step.
- Bug: I was wrong; component transfer should not be used as it reduces image quality.
- Known issues: With increased amplitude of repainting and facial detail, the face darkens.
- Fix: Third node connection error.
- Addition: Added a type of image node to avoid restarting the demo due to subsequent step errors.
- Addition: Multi-person judgment: false for multiple people, true for a single person.
- Addition: High-frequency and low-frequency facial restoration nodes. If you don't want to adjust Sharpen's parameters, you can place two images in PS and use the mask to erase.
- Bug fix: Image transfer did not proceed sequentially but all at once.
- Modified facial parameters to make the face look less like a mask.
Author Pages:
liblib
civitai
小红书
抖音
openart
Please adhere to the open source license, redistribution is allowed, but please credit the author and commercial sales are prohibited.