My HuggingFace: https://huggingface.co/jspsoli
The starting point of this model was a Anima realistic finetune merge.
I baked this lora into that finetune: https://civarchive.com/models/2650553/anima-edit
Then trained a lora exclusively for uncensoring on top of that.
Finally, I merged the resulting lora into the model it was trained on and this is the result.
(lora diff version: https://civarchive.com/models/2804197/anima-uncensor-v1-merge-finetune-lora-diff)
REQUIRES: ComfyUI-Cosmos-Reference custom node
BASE PROMPT (ALWAYS USE AS PREFIX): remove bar censor and mosaic censoring. nsfw, uncensored
Negative prompt: censored, bar censor, mosaic censoring
Basic Workflow with masking support included - check 'Optional Files'.
EDIT: forgot to mention in the workflow instructions - so you might miss it - but the 'base prompt' is automatically added as prefix to your prompt. the text prompt area of the Main Subgraph is simply a suffix to that
Training data:
Image Types:
hentai-like images (~70%)
doujinshi (~10%)
realistic content (~20%)
Censor Types:
gaussian blur (~5%)
bar censors (~70%)
mosaic censoring (~25%)
Censored Anatomy:
pussy (~40%)
anus (~5%)
penis (~50%)
nipples (~5%)
Its a very limited test model but I'm sharing in case someone finds it useful.
Description
v1
FAQ
Comments (11)
Is there any way to avoid using Nvidia_RTX_Nodes_ComfyUI in the workflow?
You can literally replace it with every other node for scaling images. RTX nodes are just really good and fast.
How to batch images in a folder in an easy way for dummy like me
Find a node that iterates through files in a directory and outputs a List of paths. Connect it to a node that loads image from path and connect that to the Main Subgraph. If that image loader node also outputs Mask -> connect that to Main Subgraph as well, otherwise just do a empty Mask with one of the native nodes (don't remember the name but its something like 'Create Solid Mask' - you need to set one of the values there to 0 for the mask to be completely black).
ComfyUI will iterate though each file in that directory but the problem is that it works as a single workflow run and the cache is cumulative. Avoid folders with hundreds of images unless you have TONS of RAM. If necessary split your folder into {Folder}\1\ ; {Folder}\2\ ; ... - each with ~20 to ~50 images depending on your RAM then execute the workflow for each subfolder.
@Winchester99 Appreciate it！
What's the use case for this? Not trying to be dismissive, genuinely curious why you would need this. I've personally found between "uncensored" in the positive and the various censor types in the negative I've never had a problem with censors for Anima, so I'm curious why you would need a Lora for this.
Its for removing censorship from censored images you did not generate/draw yourself. Its an Edit model - not meant to generate images from scratch
@Winchester99 Gotcha. I don't do much img2img stuff myself so I never considered that. Thanks for the response.
Thank you very much for your checkpoint and workflow reference. After trying it out, I would like to ask a few questions :
1 Does the seed have an impact on the output? At least after I used it, I found that it seemed to have no effect?
2. It seems that the effect of mosaics is very limited (whether it's full-color or monochrome images). I also use 30-steps and CFG3.
3. I use Sam2 to detect the bar censor area, but for some areas, even if I use a manual mask, it still doesn't work. Should this be normal? However, the complexity of the bar censor area I selected is far less than that of your sample image, but it is still ineffective. Is there something wrong with my Settings?
4.Are the Settings of the sampler and scheduler important for the output?
1 - Depending on the input image - seed may or may not change the output significantly. For instance, in cases where the censor is very light mosaic or bar censor with low opacity -> the model can extrapolate most of what is supposed to draw from the input image so changing the seed will have very little impact.
2 - That's a known limitation. Perhaps if I trained with higher gradient accumulation ; lower LR and for much longer - the capability of the model would increase but since this was kinda experimental I simply rushed through the training so it came out with plenty of limitations. Since Krea2 is also good at anime I'll make a better one - more carefully trained - when and if Krea2 Edit ever gets released. This Anima project was just for fun and very experimental
3 - If the censored area is very small - both in terms of total % of the input's total pixels as well as its width * height area -> theres a high change the model will not be able to detect it. You could enhance your workflow to make a mask around that area and expand it (to add some context to the model) -> upscale that crop if width * height is less than 768 (ideal input size is 1024 x 1024) -> run inference on it -> re-scale to match original dimensions -> paste on top of the original image.
4 - The settings I used may not be the best. I barely tested before uploading and I haven't had the time to fully figure it out yet. Sorry about that.
@Winchester99 Thank you for your answer. If the version of krea2-edit is released, please update it as soon as possible! I'm really looking forward to it!