@Wenaka_ thanks. I wanted to know the base model for a Control LoRA extraction (it's quite a bit smaller than 4GB even at rank 256 and seems to work fine too)

@Wenaka_ I used the StabilityAI ComfyUI nodes to convert the ControlNet into a Control LoRA. They have a node to do that.

解决了这个问题，加一个噪波遮罩就行了。工作流可以参考我发在下面的图片。另外注意到，仍会引起一些遮罩外的轻微细节变化，应该是VAE编码的问题。

Solved this problem, just add a noise mask. Refering to the workflow in the image I posted below. Also note that it will still cause some slight detail changes outside the mask, which should be a VAE encoding issue.

@oblevdor Is it possible to solve the slight change by hooking your workflow up with a Crop N Stitch node? It should be very adaptable and not affect anything in this workflow.

@Pupper 图片经过VAE编解码后必定会发生变化，如果要保持遮罩外部分的100%一致必须使用裁剪拼接的后期处理，既然都做到这一步了，我建议搭配ComfyUI-Impact。

The image will definitely change after being encoded and decoded by VAE. If you want to keep the part outside the mask 100% consistent, you must use Crop N Stitch post-processing. Now that you have come to this step, I recommend using ComfyUI-Impact.

Hi, does Crop N Stitch work? If it works, can you share the workflow? I'm comfy noob.😭

@oblevdor 我测试了vae encode和decode无中间处理输出与原图经过遮罩组合，发现sdxl的vae保真度非常高，完全没有遮罩接缝的痕迹，所以这个锅vae不背

@Wenaka_ it was a simple change so glad I could help. I'm repeating myself but the model you posted is so amazing, Solves a lot of issues with using img2img inpaint and configuring masks etc.
Many thanks!

@LyloGummy Hi, im a bit late, but does this work for forge? i dont know if changing the same files on forge would do the same as it does on reforge

@orhay1 Hi, sorry for the late reply, I've scanned through the files on forge and it should work, you can just copy the green lines from here (https://github.com/Panchovix/stable-diffusion-webui-reForge/pull/328/files) to extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py and restart the UI - let me know if you encounter any issues

@Wenaka_ 谢谢

I've "fixed" the issue by adding a regular img2img inpaint mask, larger than the controlnet mask.

you could also use the Color Match node, might save you some hassle

I have the same problem in reforge (it's working with a special preprocessor) so I hope it will be corrected in the model

@wewewew Can you share your "fixed" workflow please🥺?

@wewewew Could you please share the workflow for this fix?

@EX_Lazy_Cat My workflow has way too much stuff in it, just look up how to make a normal inpaint, using the same image, and the same mask but passing it through a grow mask node of a few pixels.

@wewewew Thanks for the reply, but I don't quite understand how the node resizing the mask will affect the colors that are not affected by the mask.

@EX_Lazy_Cat Growing the mask is just to have the normal inpaint on a slightly larger area than the controlnet inpaint, if they're the same size it creates a noticeable edge. You can use the "Grow Mask With Blur" node from KJNodes for a better transition. You can also use the ImageCompositeMasked node with that mask to just paste the original image on top.

https://github.com/ATata-name/sd-webui-inpaint-noobaixl-preprocessor.git

解决了，有个地方连错了现在才发现

@socialdeath  Can you share the workflow if you are using comfyui. I am having the same problem.

@diffusional_reactor https://civitai.com/posts/15490059

@socialdeath Thank you so much!

可以自己用 torch 转换成 fp16

https://civit.ai/models/1809778

https://github.com/ATata-name/sd-webui-inpaint-noobaixl-preprocessor.git

@asmdz Is it yours? if it is, then cool dude :)
Well, if it's not, I still want to thanking you for helping~

I'll try it maybe this weekend, and let you know the result, if it's yours XD

I'm still confused on how to use this. please help

@Wenaka_ 了解，感謝

To make it work https://github.com/ATata-name/sd-webui-inpaint-noobaixl-preprocessor.git

@asmdz Just saw this, thank you!

@asmdz reforge already have this preprocessor but I still couldn't get the inpainting to work
EDIT: nevermind I got it figured out right after I posted this comment lol

补充一下，遮罩选的是黑的

运行的时候K采样器一开始就把遮罩的地方弄成纯白了，后面的迭代过程中也一直是白的

解决了，在Inpaint Preprocessor后加了个Images to RGB就好了，从别的地方抄的，不知道什么原理……

大佬,怎么解决的,能给个链接吗

yuzhengzhang198854 我刚又看了看作者的示例流程，是我这边不知道什么时候把Inpaint Preprocessor的black_pixel_for_xinsir_cn参数变成false了，改成true就好了

如果你是说forge的主分支——似乎没有合并。

我确实挺头痛的，我不想再捣鼓一个新的包了，但是forge主线本身又几乎处于一种停摆的状态。也许在这里选择comfyui更加合适。

或者说还可以用于给图片的部分区域补充细节

还是说这相当于是可以用于comfyui的简化版边缘较为柔和的重绘？

我发现这和重绘又有点不同，它会改变一点点图片中蒙版外的部分，而且这对于画面的补充还挺有效果的，比如我可以将一张本来是NSFW的图片把胸部加上蒙版，提示词写上bra，这样就会不怎么违和的给角色加上胸罩。

似乎好像还会把图片中没有打上蒙版的其他部分加上多余的噪点，无论是漫画还是真实，漫画的噪点会很少但是图片的细节会有丢失

@Wenaka_  here a exemple:
https://files.catbox.moe/w23a9d.png

@Wenaka_ same here

Use the VAE Encode (Inpaint) node before K-sampler

Load it with the inpaint_noobai preprocessor. No idea where you would get that on its own but in my experience it is built in to some a1111 forks. I've had it available in forge classic and forge neo.
When I used an incompatible inpainting preprocessor it would just turn white where I masked. Sometimes I still forget.

@Guinquisition its clear, but what is the difference between this ControlNet model and the standard InPaint feature?

@mifink94 It's noticeably smarter, in regards to the area around the mask and picking up the style.

@Guinquisition thenk you a lot)

@Guinquisition Hey I'm facing the same issue with the inpaint area turning white. Using SwarmUI, do you know which InpaintPreprocessor is compatible?

same，Has this problem been solved?

@supersuika Yep. That was my mistake. I decided to stop using vae encode and not take the latent from the source image, since I needed an upscale. In the end, I dropped that approach and the problem went away