ComfyUI Tutorial+Workflow: Inpainting only on masked area, fast outpainting, and seamless blending

This tutorial presents novel nodes and a workflow that allow fast seamless inpainting, outpainting, and inpainting only on a masked area in ComfyUI, similar to the AUTOMATIC1111 feature but with extra flexibility.
The main advantages of inpainting only in a masked area with these nodes are:
- It's much faster than sampling the whole image.
- It enables setting the right amount of context from the image for the prompt to be more accurately represented in the generated picture.
- It enables upscaling before sampling in order to generate more detail, then stitching back in the original picture.
- It enables downscaling before sampling if the area is too large, in order to avoid artifacts such as double heads or double bodies.
- It enables forcing a specific resolution (e.g. 1024x1024 for SDXL models)..
- It doesn't modify the unmasked part of the image, not even passing it through VAE encode and decode.
- The nodes take care of good blending.
The Inpaint Crop and Stitch nodes can be downloaded using ComfyUI-Manager, just look for "Inpaint-CropAndStitch".
The example workflow featured in this video can be downloaded from github.com/lqu...
I hope you like them!! Subscribe for more :)
The song playing in the background is "The last dial-up handshake", from the album "Last Transmission" of my music project "Elezeta". Check it here: • The last dial-up hands...
#comfyui #stablediffusion #genai #inpainting #tutorial #workflow

[@marcinchaciej]

This should be default in Comfy, you made a great great contribution and we all really appreciate it

[@elezetamusic]

@@marcinchaciej thank you!!

[@NywlMac]

THANK YOU!! I was looking for a tutorial so I could understand how the nodes for a inpaint workflow works, and all the ones I found is already those HUUGE workflows. This one is simple and works, and now, as I could understand how it works, I can improve it as I please. Great tutorial!

[@subashchandra9557]

Now I just need to know how to do the Fill/Original/Latent Noise/ Latent Nothing options in ComfyUI!

[@EmeranceLN13]

subbed! Thank you for such a straight forward explanation !

[@derrickpang4304]

This is exactly what I need. Thank you so much!!!!!

[@elezetamusic]

@@derrickpang4304 thank you!!

[@matthallett4126]

This is a great solution, I love it. Thank you so much!

[@李云-f1b]

The work is great, it hasn't been updated for a long time

[@Vigilence]

I have a photo with a woman. I created a mask for the ear and also masked some of the below portion so I can add an earring. However when in-painting no earring is added. Any tips?

[@elezetamusic]

@@Vigilence try setting a context mask with the whole face and write a prompt such as "a woman with earrings". That probably works

[@Vigilence]

@@elezetamusic I will try! Ty

[@Milo_Estobar]

Thank you for your contribution 👍🏽 (App extension creator + Tutorial)...

[@MrPer4illo]

Thank you for the video. Can this approach be implemented for animatediff (animate only certain areas of an image)?

[@elezetamusic]

totally, I am aware of some users using it for that purpose.

[@mahilkr]

Hi @Elezeta, excellent work! How can I generate multiple variations of a stitch? Currently, it only works with the repeater set to a value of 1.

[@elezetamusic]

Hey, please provide more details on what you want to do. If you want to generate multiple images, you could enqueue the job multiple times or in comfyui advanced options, set a higher batch number. If you want to use a repeater node, you'd have to repeat both the image and the masks (not sure if the repeater node can do masks). Not sure if I answered your question. If not, please let me know what you mean by "it only works...". Does it give an error?

[@李云-f1b]

Very good, thank you very much

[@PanKazik]

Very nice tutorial. The only problem I have is that despite the model I use (tried a few inpainting models) the result is always black square in place of mask. Any idea why it happens? (i am using comfy on mac)

[@elezetamusic]

@@PanKazik do inpainting models work for you without my nodes? If not, then the issue is unrelated to the nodes and I don't think I can help. If it works without my nodes, I'd suggest to load the workflow from github and try it with different models without changing anything. That should work. I also use Mac. There's an issue with some Mac updates that the GPU only generates black images, but it would affect all models, not only inpainting. Check if that's the case.

[@PanKazik]

@@elezetamusic I tried simple inpaining workflow and it worked. However I found the reason why it worked. The problem was number of steps. Anything over 16 produced black square. With that in mind everything works flawlessly. Thanks for your response :)

[@hphector6]

i have an issue where the mask is still very visible, like a gray mask over the inpainted area

[@elezetamusic]

@@hphector6 hi, I'd bet that you're using vae encode for inpainting and a denoise lower than 1. If you want to use a denoise lower than 1, use InpaintModelConditioning instead of vae encode for inpainting.

[@hphector6]

@@elezetamusic I was using inpaint model conditioning as per the default workflow. I think it might have been the model im using which is a pony variant, no issues with epic real xl

[@CemilAL]

good job 👍

[@elezetamusic]

Thanks 👍

[@fromloveandlifestory]

THank you!!!, Please create a note that can import a high-resolution image and split it into specified sections. Then, it should run the img2img (inpaint) process on each section and finally combine all the split sections back into a complete image. The goal is to be able to process high-resolution images without having to manually split them in Photoshop and edit each section individually

[@elezetamusic]

@@fromloveandlifestory there is tiled sampler for that!

[@fromloveandlifestory]

Could you please share a similar workflow with me? Thank you very much!

[@InaKilometrosX1TUBO]

Hi, thanks for your videos, I have an error, see if you could help me. - Prompt outputs failed validation UnetLoaderGGUF: - Value not in list: unet_name: 'None' not in [ ]

[@elezetamusic]

@@InaKilometrosX1TUBO this doesn't seem related to my nodes but to other nodes that I don't know. Sorry, can't help

[@ARAI96969]

Thank you so much for this node! Before this I've never found a way to do inpaint one area without sending the whole image to Encode and spoiling the overall quality.. This Is a god send!! But I do have 1 query on which setting should I do if I wish to inpaint a large area (example 1/4 of the whole image to mask a whole character) and change a character entirely with another lora character, thus creating 2 unique characters interacting Cause if I mask a large area, the output is usually very bad, lacks details and distorted, do I upscale and area, Downscale it or increase size of mask perhaps? Thanks if u have any tips, cause I wish to keep my workflow simple and avoid using segment and auto detect to mask and repaint characters to my lora. I prefer to choose and mask them myself for more control

[@ARAI96969]

Hi, do u have any tips for inpainting larger areas? Cause I used your settings but it will generate distortions, any advise on good settings would be greatly appreciated

[@elezetamusic]

@@ARAI96969 well, for larger areas I'd suggest to inpaint the whole area first, then detail the key areas in it with several passes. You could also consider a tiled sampler. There's no magic solution to sample on higher resolutions with high detail in a single go and fast.

[@davoodice]

Thank you

[@Vigilence]

I used this node successfully with a regular mask. However, if I use the invert mask option with the precious mask and the same base image, the image that is inpainted is much lower in quality. Same settings are used from the successful in paint, so I’m not sure what the issue is. Can you confirm by testing with the lasted comfyui that the invert mask option is working properly?

[@elezetamusic]

@@Vigilence Hi, this is because the inpainting area is so much larger if you do the whole image, so you get it downscaled by the node to fit the target resolution, and therefore, you lose resolution. A good solution is to apply a detailer afterwards. I'm figuring out if I could implement some node that does that detailing for very large images on a way that's easy to set up with these nodes

[@olekXDDDD]

it does not work with the eff ksampler.advanced. ill get an error. in case someone is having the same problem: "Input and output sizes should be greater than 0, but got input (H: 1, W: 1) output (H: 0, W: 0)"

[@elezetamusic]

@@olekXDDDD it should, though, as basically you'd still sample on an image and mask. You must be plugging things incorrectly

[@AInfectados]

First, thx for this wonderful WF, was searching for this for a long time to stop using AUTO's to only do the Inpainting part. Can you adapt this for *FLUX* and also add the node LORA POWER LOADER, please? PD: We have a now an Inpaint Model FLUX (Controlnet): *FLUX.1 dev Controlnet Inpainting Alpha*

[@elezetamusic]

@@AInfectados these nodes should be completely independent from the model you're using! Just plug in your samplers etc on the cropped image

[@Krashl]

Thanks for the work! I use these nodes all the time to inpainting, but I was missing some features like custom sizing and blurring the mask (because this node creates hard mask borders by default), and the abrupt transition was visible on generations. I am very glad that the recent update has improved all of this

[@elezetamusic]

Awesome!! I'm glad that you found it useful. Thanks a lot for commenting.

[@jonjoni518]

fantastic!!!

[@henryphillips6167]

How do we modify this workflow to work with premade masks? I would also like to use another image as the fill for the masked areas. Could you detail how I would go about this?

[@elezetamusic]

@@henryphillips6167 i get what you're asking but this is out of the scope of this video and these nodes. I'd suggest you to continue to learn using comfyui and you'll eventually figure out how to do this. Sorry, I can't offer tailored support for comfyui

[@user-yg4qo9zg3u]

You are a wonderful person, thank you for sharing! This is what I've been looking for! Just perfect! 🔥

[@KevinScandinavia]

Yeah no kidding, and including the json file was just the icing on the cake

[@MarcSpctr]

seriously great work 👍🏻

[@elezetamusic]

thank you!! :)

[@Damian151614]

Do you know why I get unmatched colors? I set denoise to 0.00 on purpose because I wanted to get the same output as input image, but somehow I get faded colors on masked area.

[@elezetamusic]

@@Damian151614 that's the encoding and decoding process of the VAE. A different VAE may get more accurate colors for that specific image

[@Damian151614]

@@elezetamusic It was model problem. I tested it on another models (non VAE and VAE baked models) and only that one caused problems.

[@treksis]

thank you so much. saved my life. I tried to copy auto1111's Inpainting myself, this is much better.

[@dfhdgsdcrthfjghktygerte]

I want to erase something from a skin and flood fill it with just one color that maches the surroundings. Is this possible? When i try to use the "skin" or "color" prompt - it inserts faces or random stuff into the masked area.

[@elezetamusic]

hi! extend the context area enough so you can see where the skin is (e.g. an arm, a leg, whatever) and then type in "an arm", "a leg", or even extend the context area further to show there's a person, and then type a prompt like "a person". that will give the sampler enough context to fill in the gap seamlessly.

[@elezetamusic]

also, use an inpainting model, they work much better and don't add random stuff in the masked area

[@Pyugles]

Thank you so much! This is a fantastic set of nodes, and great tutorials on using them! You've definitely gotten a sub from me!

[@elezetamusic]

thank you!!

[@joelandresnavarro9841]

Hello 🙋🏻‍♂️, you could consider improving the compatibility of using these nodes with the comfyui-photoshop extension (NimaNzrii) When sending photoshop, he mentions that the image and the area are not the same size. I don’t know if it’s a problem with your node or rather with the extension for psd

[@elezetamusic]

Hi, I do not have photoshop and can't install that extension. Anything related to photoshop integration you should reach out to who developed that node. Cheers!

[@RemiStardust]

Excellent work! This is something I really wanted! Very simple workflow, easy to even stitch the updated area back in to the original image! Thank you!

[@elezetamusic]

@@RemiStardust thank you!

[@Daralima.]

These nodes make inpainting in Comfy super convenient and easy to adjust to your needs. Thank you!

[@elezetamusic]

Thank you!

[@skycladsquirrel]

Perfect. Thank you. Subscribed!

[@ronnykhalil]

so glad i came across this on a reddit thread. super helpful

[@Kikoking-y9b]

Hi ,very nice node thank you a lot. I have a question, How can I do upscaling before sampling ,what is the meaning of that or how can I do it? Iam not sure if you mean what I think. I think of cutting an area around a 'Face' and upscale only the face, and stitch the upscaled face back. Can you help me?

[@elezetamusic]

If you set mode to ranged size or forced size, the cropped image is automatically upscaled (or downscaled) to fit that resolution. Then you sample on it, and then during stitch it is returned to the original size. So you don't have to worry, the node takes care of it for you! You can check it by previewing the cropped image and checking its size

[@NeonSparks]

This is exactly what I have been looking for! Amazing, can't wait to try this. Thanks

[@elezetamusic]

@@NeonSparks thank you!! Enjoy!

[@imonutiy]

Looks like blend_pixels just adds more pixels to context so it is the same as 1st option.

[@elezetamusic]

@@imonutiy not really! Blend adds more pixels to have enough context to blend, but blend also does a gradual blending of the newly generated area into the original image, so that the transition is less abrupt. You see more context because it is required to blend.

[@imonutiy]

@@elezetamusic Thank you for the answer, by the way what are other options for inpainting sketching nodes in comfyui. It feels really wonky compared to web ui.

[@831digital]

Instead of manually painting the mask in, do you have an example of this working with a detector that generates masks? This would make it more useful for animation.

[@elezetamusic]

No, but you can easily put it together :) give it a go!

[@831digital]

@@elezetamusic I tried doing it with SEGS, but SEGS resizes the video and throws an error when trying to feed the mask back to the original. If it's super easy, please share an example.

[@natlrazfx]

brilliant, thank you so much

[@elezetamusic]

Thank you!