ComfyUI Tutorial Series: Ep19 - SDXL & Flux Inpainting Tips with ComfyUI

[@pixaroma]

Join the conversation on Discord discord.gg/gggpkVgBf3 You can now support the channel and unlock exclusive perks by becoming a member: pixaroma kzbin.info/door/mMbwA-s3GZDKVzGZ-kPwaQjoin Check my other channels: www.youtube.com/@altflux www.youtube.com/@AI2Play

[@pablodara8720]

This video is Top tier quality, I love you

[@Uday_अK]

Thank you. I have been waiting for this in-painting and out-painting video for a long time. There are many tutorials available on KZbin, but understanding the method and implementing it is very important. You have explained it in a very simple way.

[@pixaroma]

thanks uday 🙂

[@Patricia_Liu]

Thanks for all your hard work! I really appreciate the effort you put into your tutorials.

[@pixaroma]

Thank you so much for your support 🙂

[@sergeysaulit]

Simple, accessible and understandable! You are the best one who delivers complex information! Thank you and don't stop!

[@pixaroma]

Thank you 😊

[@Kozitaju]

Perfectly explained ! Thanks. I can at last move from Forge to Comfyui for Inpainting thanks to your tutorial and this stich and crop tool 🙂

[@hyperdeloutz2444]

Perfect tutorial. Like the way you explained the workflow.

[@pollapatjampeeklang7493]

It was worth the wait; you never truly disappoint.

[@pixaroma]

Thank you

[@SumoBundle]

Thank you for another fantastic episode!

[@59Marcel]

Great tutorial! You made everything so clear and easy to follow - thanks for breaking Inpainting down so well!

[@pixaroma]

Thank you 😊

[@emergentcomics4926]

I've been waiting on this one. Thanks so much! I can't wait to give it a try

[@baheth3elmy16]

Great episode! Thank you very much!!!!

[@SebAnt]

Wonderful tutorial once again. I like that you explained about denoise and also demonstrated photoshop techniques at the end! I will try out once I get some free time…

[@pixaroma]

thank you 🙂

[@cives]

Me, on the other side, I prefer no PS here. Or, if there is any, not using the AI into it, so that we can use as well GIMP alternatively. The magic of this channel is its focus on ComfyUI and its freedom to let us do what we want, under our control. I moved away from Leonardo and the like looking for that. Besides, PS is the exact opposite of free (in both meanings). If this channel gears into more PS, it will lose me. There are other fantastic channels out there, and this one will be losing its edge by losing its focus.

[@staicpaic4928]

Inpainting is an awesome feature, this video was very helpful, thank you very much.

[@pixaroma]

Glad it was helpful :)

[@ob3ythee.t.128]

Very well done tutorial series, very easy to follow, I also recommend if anyone does not want to setup ComfyUI in the same way instead use Stabilitymatrix

[@Jinjinyajin]

Love the tutorial details and it's easy to understand

[@RealNazrax]

I was hoping you'd cover inpainting! Now, please do IP Adapter :)

[@aminshobeiri8901]

Life Saver, thanks for sharing.

[@Filokalee999]

Thanks! Very clear and valuable Inpaint tutorial. While trying a similar workflow, using Differential Diffusion node in combination with InpaintModelConditioning seems to have better integration results than just InpaintModelConditioning alone. Additionally if this node refuses to inpaint the shirt, try using another Inpaint model, (such as BrushNet).

[@pixaroma]

thank you

[@JefHarrisnation]

Love your tutorials.

[@UmarandSaqib]

awesome!

[@alexandrapadureanu4192]

New interesting stuff ❤

[@maxmad62tube]

very impressive and informative thank you very much

[@arep7424]

For the color problem and fine tuning denoise, just add a controlnet ( depth are good ) with the cropped image as reference. this allow higher denoise ( even 1) on sampler without deform and transfigure the subject

[@pixaroma]

thank you, I will give it a try :)

[@ling6701]

Thanks. And now DIY photoshop's Generative Fill in our arsenal, another christmas :)

[@hannibal911007]

Hi. Thanks for this valuable serie on comfyui. I would suggest you an episode on How to train a lora on comfyui, which workflow depending on check point used, best ckpnt you May suggest us to try... Keep going and thnx

[@pixaroma]

thanks, yes people keep asking me, just didnt find an easy way to do it locally, only online and that is not free. The method I tried are harder to install and gives errors, that I am not able to explain the community how to fix them, since I am a designer not a coder. So I have been waiting for an easy solution, like fluxgym can work sometimes but still give some errors sometimes

[@K-A_Z_A-K_S_URALA]

You explain everything very coolly, well done! Question: where can I get monitoring like yours, next to the "Manager" button, I updated it, but I don't have it... And the question: will you record a video of "Laura Inpаint" with new accelerators?

[@pixaroma]

Is the crystools node that shows those, you can move the topbar to bottom and again on top if doesn't appear right away. As for new accelerator i need to do first research to see what is about

[@K-A_Z_A-K_S_URALA]

@@pixaroma Thank you very much, have a nice day time! we have a night out) in Russia)

[@K-A_Z_A-K_S_URALA]

Очень Крутой человек Вы и Умный !

[@damnned]

Fooocus inpaint is the king I think

[@pixaroma]

i think I saw someone using a fooocus model in comfyui, but not sure.

[@radedr007]

great video

[@pixaroma]

thank you 🙂

[@kattamaran]

Add a depth controlnet to have the generation more similar. Specially in flux

Exciting as always! Could you share the Inpainting and Outpainting workflow so that I don't have to rush from the video?

[@pixaroma]

Is on discord on the pixaroma-workflows channel link to discord in the header of the channel or in video description

​@@pixaromaOh, I found it, thank you!

[@pixaroma]

this is the direct link, mention pixaroma there if you still cant find it discord.com/channels/1245221993746399232/1270589667359592470/1300813045127188591

[@MrOlivaz]

Thanks a lot man! you're amazing, what is the recommended imagen size for do inpaint in flux using Q8? i mean to get the best results, thanks in advance!

[@pixaroma]

the node only make a selection of what you paint in the mask editor so that is the actual size that can inpaint so i will say if that inpaint area is around 1024px is ok, with flux can go a little higher, but with sdxl can not go too high

[@silfrasjoeld]

The shortcut to increase brush size, rotating the mouse wheel zooms, what is the current shortcut to increase brush size?

[@pixaroma]

Should be scroll but try with one of the key pressed i dont remember shift, alt or control, or maybe use the brackets []keys

[@djwhispers3157]

Is there a workflow that helps clear up and add detail to low res low light images, from low quality to 4k UHD? I noticed in this workflow that refining with flux can add more detail.

[@pixaroma]

I only use upscaler like in episode 12

[@TimesNewRomanAI]

Thank you very much for all the information and how easy it is to follow. What pc configuration do you have and what time do you get in the image generation?

[@pixaroma]

Rtx4090 24 gb of vram, 128 gb of ram. Depends on the model between 3-15 seconds, sdxl is fast and flux takes like 14-15 seconds to generate an image

[@TimesNewRomanAI]

@@pixaroma Thanks for the info. 15 seconds is really fast

[@ArielTavori]

FWIW, unless I missed some critical changes to architecture or tooling, inpainting models should ONLY be used with a full denoise strength of 1.0. Standard inpainting workflows feed pure noise in to the masked area, and inpainting models generally do very poorly at img2img. If you want slight modifications, you will have much better luck and much more control using standard models with img2img workflows and a low denoising strength anywhere from 0.3-0.9 depending on the model and situation

[@bratan007]

Thank you for great tutorial! How do you do workflow tabs in ComfyUI?

[@pixaroma]

go to settings (that gear wheel) and search for workflow, look for Opened workflows position, and choose Topbar instead of sidebar

[@Gabbaki]

Can you tell me how I can change the color of a garment with such a small workflow? But without changing the garment! Only the color.

[@pixaroma]

Is hard to keep things intact, i will probably just change the color in Photoshop and use image to image with low denoise to better blend the colors.

[@ian2593]

After upgrading to the new interface, I can't tell where my workflows are being saved to. Is there some way to tell?

[@pixaroma]

you can click on export and put it in any folder you want, if you go to top left corners and workflow and choose export, and choose a folder. and with workflow open you can open it. If you save it with save as or save it will go into the folder workflows, the path is something like this ComfyUI_windows_portable\ComfyUI\user\default\workflows

[@bentontramell]

Flux chin sighted. 😅😊

[@pixaroma]

😁 can be fixed with sdxl Inpaint 😂

[@CharlesPrithviRaj]

how do we include the new flux canny and depth lora with the inpainting workflow. How to combine pix to pix instruct and the inpainting condition nodes

[@pixaroma]

you can connect the lora just like i did at the end of episode 24, but I remember I tried something and didnt work quite as expected, maybe because lora need 10 for flux guidance and inpainting didnt need so much or something, or maybe i didn put the right settings

[@CharlesPrithviRaj]

@@pixaroma Yes I was confused which latent would go to the ksampler. Is it the one from inpaint conditioning or the one from pix to pix. See if you can figure it out.

[@pixaroma]

@@CharlesPrithviRaj Ah I see what you mean, probably that why I didnt get the result I expected I think i skipped one of the inpaint nodes :)) I didnt test it but you can try with one of this nodes, LatentAdd or LatentBlend and that comes from both inpaint node and go to one of that latent node and from that node you connect to ksampler, so you combine both latent, in theory should work, but as I said I didnt test it, let me know if it works

[@emrekiratli2614]

is there a anyway to save the inpainted image only, for instance i inpaint a hat. I retrieve 2 images which are stiched and and unstiched. what I need is only generated hat.

[@pixaroma]

when it inpaint it doesnt just add a hat with transparent background on top, it used that cropped image the area you masked and then based on that it generate a new image that is used to be placed on top of the other one, so if is a bunny head, and you ask for a hat, the output will be a bunny head with the hat, that is merged with the original image with the full body bunny. You can see what that image look like if you drag a preview or save image node from the vae decode just right before it go to stitch node at the end

[@r3vdev]

When are you going to do an IMG2VIDEO tutorial???? :)

[@pixaroma]

when there is a good video model, so far the video models I saw are not really usable, compared with Kling ai for example that create decent video

[@yasinolgun7548]

In the introduction part of the video, the female picture appears in motion. How can I do that, what should I research?

[@pixaroma]

I used image to video on the platform klingai.com/ so far i didn't find a free options that does that good like locally that why i used online platform

[@icedzinnia]

When I did this tutorial, it didn't work (5:52). I realized after much fiddling around, using different loaded images, different masks, that it won't work if the mask is not continuous. I couldn't make it work when i had a mask on the far right edge and a seperate one on the far left edge. WEIRD.

[@pixaroma]

I usually do one selection, there is on the inpaint crop an option called fill_mask_holes, maybe try to turn it off to see if helps, if not maybe is some bug

[@topy706]

is there a way to increase the batch size?

[@pixaroma]

From queue you have batch size 1, you increase that number

[@topy706]

@@pixaroma but in queue thats batch count not batch size. i tried to add a "repeat latent batch" node but the stitch node does not like it

[@pixaroma]

I don't know any, i usually just use that batch and check in a few minutes or use increment queue, so only if you find a node to do that

[@topy706]

@@pixaroma okay thank you anyways. 👍

[@eros6398]

friend i have an annoying problem i just cant find the florence2 imageprompt node

[@pixaroma]

there are two, i use this one ComfyUI-Florence2 by kijai yo can find it in manager. this is their page github.com/kijai/ComfyUI-Florence2 look how i use it in episode 11 kzbin.info/www/bejne/r6bXiohvbKedbacsi=EZ2yHl6-7tsbGapX

[@eros6398]

@@pixaroma thank you men

[@audiogus2651]

Is there a Comfy node that lets you paint on the image like in Forge?

[@pixaroma]

I dont know one, but there are so many nodes, I am sure there are some that let you do that

[@sollmasterdoodle]

Thx! But what’s the version of your comfy ui?

[@pixaroma]

I i go to manager and scroll down in the right I see this ComfyUI: 2797[770ab2](2024-10-29) Manager: V2.51.8 As for release is v0.2.5

[@sollmasterdoodle]

@@pixaroma thx but is the .exe ? the ComfyUI Desktop V1? I'm waiting for

[@pixaroma]

@@sollmasterdoodle no, that is still in beta from what I know and if you are not on the list dont have access to it yet

[@sollmasterdoodle]

@@pixaroma thx a lot too you!Perfect video!

[@bubuububu]

Excuse my amateur question, but how can I batch more images?

[@pixaroma]

Depends, you can use a load image from itools node from example that let you take images from a folder, or you can use batch the one next to queue to run the workflow multiple time, so depending on what you need to do. Chech episode 15 if you want to load a folder with images

[@bubuububu]

@@pixaroma I don't think i'm understanding. So for example, if I want to get 10 result for different glasses on the same face. Where to put the batch size?

[@pixaroma]

@@bubuububu is hard to explain in text, would have been much easier to show screenshot on discord. If you have the new interface with that floating QUEUE button, that has a number next to it like 1, you can increase that number to 10 and it will run that workflow 10 times, so if your seed is not fixed usually is set to randomize will run 10 times each time with a different seed so different results and it will stop. If you have the old interface that as the button QUEUE PROMPT then under it you have a check mark Extra Options, and there you have batch count that default is one, so you can put 10 there and run it.

[@bubuububu]

@@pixaroma thanks a lot. i knew i was missing something basic 😅

[@AInfectados]

And what about the *Controlnet Inpaint BETA* for FLUX?

[@pixaroma]

I didnt try it yet since it was flux + control net so it will take extra time to generate since is loading an extra model, but I will play with it to see if can get better results

[@jessedbrown1980]

I am trying to do a single transform inpainting with flux 1.1 dev trained loras and the same model to do inpainting The model is trained on materials and when I run it through with a mask, the whole picture chages instead of the mask, I might be missing a node, but I just wat the masked area to be changed and it always changes the whole picture and outs artifacts in them I already have lora strgh to 1 and I need to play with a lot, but I have not been able to control the mask area or the picturre, so weird! The setup you have does not have a lora in it, which would be effective, anyone have a setup with a lora?

[@pixaroma]

I didn't try it with lora yet

[@jessedbrown1980]

@@pixaroma Would you be interested in colaborating? I need to get it done! good experience too, we have like 700 loras trained waiting to be used!!

[@pixaroma]

@@jessedbrown1980 i am checking now the message on discord

[@OaklandSignalCollapse]

Thanks for not loading this up with a bunch of BS

[@raz0rstr]

can I replace an entire face with my face based on my trained lora? h

[@pixaroma]

Check the episode 24 at the end i did something like that with lora inpaint

[@mleitejunior]

Somebody has a invite to join matrix official channel of ComfyUI? Me and my friend are trying to join it for days...

[@pixaroma]

maybe ask on discord if someone has

[@devon9374]

Man, Adobe is cooked...

[@pixaroma]

😁