Update: Comfyui Workflow has been officially released!
This model is an inpaint model for realistic images.
-This model allows you to naturally modify parts of the generated images by prompting.Improve the quality of your inpainting with the rich knowledge of pony models and the flexibility of prompts to meet the needs of more creative purposes.
-V4 version
V4 was created by adding NoobAI to merge list and then remerging the illustrious models to restore realism.Non-remerging model also uploaded as ALT version.
The difference between the two is the internal ratio of NoobAI. The normal version is slightly better in skin tones, etc. In most cases, there's not much difference, but in N-situations or complex compositions, the ALT version will produce better output.
How to use:
Sampler
V3 and V4 = DPM++ 2M SDE (karras for comfyui) or DPM++ 2S a
V1 and V2 = Euler a
Soft inpainting check
Mask bluer 4~8
CFG Scale 4~7
Denoising strength 0.9~1
steps 20~50
-Soft inpainting settingchange "Transition contrast boost" to 1 (default 4)change "Difference threshold" to 0.1~0.25 (default 0.5)
Don't use "Apply color correction to img2img results to match original colors".
Refiner:
Using normal model as refiner (eg. pony realism) at 0.65~0.95 (recommend 0.8) can produce even more great results. I will pick out models that can be used as a refiner, so please try it out.
Good Refiner model list
Real Wives Pony v_1.1 [believe me, best skin texture ever. (need neg fat)]
https://civarchive.com/models/709934?modelVersionId=794078PornMaster-Pro Realism-IL-V3 [prompt variety, danbooru tag]
https://civarchive.com/models/1045588?modelVersionId=2099739RedCraft Illustrious3 Relustion [skin tone, prompt variety, liquid]
https://civarchive.com/models/958009?modelVersionId=1517097ULTRA v9 [soft, dark atmosphire]
https://civarchive.com/models/228525?modelVersionId=2098608Ponyvinculum Exotic_Mix [clear, soft, lighting, animal, nature]
https://civarchive.com/models/1424028?modelVersionId=1809768
Tips:
Only masked is best for quality
Start with Masked content "original", if not worked change to "latent noise" or "fill".
Use image size with height + width>2000~2500 (Recommended size for SDXL).
or use resized by scale around recommended size (aspect-ratio-helper).
If you mask a small part, it is a good idea to mask the top and bottom of the main mask so that the model can see wide area
Your image works as quality prompt, You don't need a lot of prompts, keep them to a minimum and let the model do work.
dog, in beach-Use this model only for AI-generated images.
-Prohibits uses that harm any person by using this model on real-world images.
Description
This version kept about 70% of the noobai on outer layer. It's better than the v4 normal version most case, but it can sometimes look like an 2D or illustration, so I made it ALT.
FAQ
Comments (14)
I saw you mentioned "PornMaster-Pro Realism-IL-V3", but your URL is wrong and people might be downloading the wrong checkpoint.
Fixed it.
Thank you for the great models<3
cool!how to use it in comfyui?please
I added my workflow as training data in model download section.
It's still under draft and hasn't been tested much, so I can't guarantee it will works well.
This is a truly excellent model. Thank you.
I have a question.
When using Refiner, the following log appears and it fails:
TypeError: ‘NoneType’ object is not iterable
(using reForge)
or
RuntimeError: Given groups=1, weight of size [320, 4, 3, 3], expected input[2, 9, 224, 149] to have 4 channels, but got 9 channels instead
(using Forge)
I am not using “Apply color correction to img2img results to match original colors”.
Is there a solution?
The Reforge series appears to be missing code to resolve the discrepancy in the number of channels between the inpainted model and the standard model when using a refiner.
Unfortunately there is no solution
@habibing Thanks to you, I figured out the cause.
The most advanced inpainting model. Bravo!
I didn't try inpaint but out paint worked great and i just refined the image after to clean it up. I use Flux Fill a lot to outpaint and it totally sucks. I think this works better and fast as hell in comparison. I used my own set up though. Good job for real
Comfyui cant find Sam3 after manuall download into sam3 folder
From RMBG requirment.txt
you need below
decord
ftfy
typing_extensions
#triton-windows
@habibing Hey. Thanks for your feedback. That's exactly the text in my file. I've tried installing Sam3 manually and via ComfyUI Manager. It just can't find the node.
# Base dependencies
huggingface-hub>=0.19.0
transparent-background>=1.1.2
segment-anything>=1.0
groundingdino-py>=0.4.0
opencv-python>=4.7.0
onnxruntime>=1.15.0
onnxruntime-gpu>=1.15.0
protobuf>=3.20.2,<6.0.0
transformers>=4.30.0
diffusers>=0.18.0
#SAM2
hydra-core>=1.3.0
omegaconf>=2.3.0
iopath>=0.1.9
#SAM3
decord
ftfy
typing_extensions
#triton-windows
First, check if RMBG is the latest version. In the folder ComfyUI/custom_nodes/Comfyui_RMBG, is there a file named AILab_SAM3Segment.py?
If it exists, please check the following:
decord
ftfy
typing_extensions
triton-windows(if you use windows)
Confirm whether these 4 packages are installed.
(Example) In the console/terminal, enter the following commands one by one:
pip show decord
If it’s not found / not installed, you can install it like this:
pip install decord
When entering pip install ～, please adjust the command according to your own virtual environment (venv, conda, etc.) if you’re using one.
@habibing thank you!!