Once paired with the ComfyUI-Cosmos-Reference node, this LoRA can be used to uncensor 2D/anime-style images.
Simply enter “uncensored” in the prompt. In addition to the Anima base model, it also works with turbo.
For the anima-edit workflow, refer to the Anima Edit LoRA page:
https://civarchive.com/models/2650553/anima-edit
___
For details on the editing workflow, refer to the explanation below:
The current Anima base-v1.0 model does not support editing, but the Anima-edit mechanism has already been demonstrated by others.
For example, the following two LoRAs are both excellent:
Anima Edit-experimental
https://civarchive.com/models/2652469/anima-edit-experimental
nude filter | simple edit
https://civarchive.com/models/2587631/nude-filter-or-simple-edit
When editing, you need to install the following node in ComfyUI:
ComfyUI-Cosmos-Reference
https://github.com/Mirumo0u0/ComfyUI-Cosmos-Reference
Non-ComfyUI users can refer to the explanation in the following article:
How to Use Anima-Edit (Cosmos Reference + Edit LoRA) in Forge Neo
Description
This version enhances the training dataset.
FAQ
Comments (5)
First of all, thanks for this lora.
As always there is too many edge cases when training such loras. I have mostly 3 problems:
1 When I try to uncensor compic page, not all panels uncensored. It tends to uncensor most of bar censor (black ones), but not all.
2 Even worse quality on high resolution (for example 2MP).
3 Generated pixels usually poorly follow the content of image: sligtly changed color or completely missing composition/lines.
For test I generated square 1.5 MP 4 panel monochrome comic with dot shading with close up of pussy and black bar censor. Used your lora and result is not all bars got removed.
Please consider expanding database to manga/comics, adding more styles in dataset and train higher resolution.
Also surprisingly your lora works with anima turbo1.1
Here are my responses to your questions:
1. Yes, it is a known issue that uncensoring black bars can sometimes fail. I am still working on overcoming this issue.
2. Due to the limitations of my hardware, the training dataset was trained at a resolution of 1024, so it may not be able to handle images above 2MP very well.
3. The slight color shifts may be related to the cleanliness and quality of my training data, so this is something that may be difficult to resolve immediately. As for the original composition or lines being missing after generation, that is also possible. After all, image generation relies on the model making predictions and reconstructing what it thinks should be there. Ultimately, I believe that having more training data will be necessary to address these issues effectively.
Thank you for your suggestions and feedback!
aahh this is too much info for a simple person like me can you make a video on how to use this lora and also can you share a workflow you are using to uncensor? thank you
Sorry, I don't usually make videos. Instead, here are the steps for using it with ComfyUI:
1. Download the sample workflow created by mattehe:
https://civitai.red/api/download/models/2976234?fileId=2856079
2. Use ComfyUI Manager to install the ComfyUI-Cosmos-Reference node required by the workflow (author: Mirumo0u0).
3. Replace the LoraLoaderModelOnly node in the workflow with the LoRA provided here.
If you're not using ComfyUI, I'm afraid I can only refer you to the instructions on the following website:
https://note.com/hirorohi03/n/na72233a8d6a4
Thank you!
@avans really thank you for the response really thank you i did it and thank for the lora and if you can please can we chat in a dm ?