🌟 Ultimate Flux Speed and Inpainting Tools (with free secrets!)

Рет қаралды 6,668

Күн бұрын

Пікірлер: 58

@Bikini_Beats 27 күн бұрын

Amazing tutorial, thanks. Subscribed. I managed to install all nodes, but i have 1 error i can't figure it out. Advanced inpainting workflow, those 2 little nodes on the right: Mask crop region coming from the bridge preview,are in red. Any idea what do I need to install or do to get them to work? Thanks

@GrocksterRox 27 күн бұрын

I really appreciate it! Did you try updating Comfy recently? There may be some new logic (if you haven't updated in awhile) that's used for some of these nodes to work. It's a bit tough to diagnose over YT chat but happy to take a quick look if you wanted to jump on the discord? discord.gg/SCK4MEZ3ph

@steveg9380 4 ай бұрын

wow your tutorials are the best Ive seen to date. Your explanation of how things function is well done my dude, hats off!

@GrocksterRox 3 ай бұрын

Thank you so much, the more we can help each other, the better the art and media we can create! Feel free to share the knowledge wealth (and videos) 😁

@runebinder 2 ай бұрын

The Advanced Inpaint workflow has been my go to for that task, so with the Flux Fill Model I wanted to see if I could intergrate it. You might already have done this, but in case not I swapped the basic guider for a CFG guider for the negative input. Connected the positive conditioning to InpaintModelConditioning node and an empty Clip Text Encoding for negative and hooked uo the rest of the inputs, have the positive and negative outputs from the InpaintModelConditioning feeding into the CFG Guider and the latent going to the SamplerCustomAdvanced and removed the Set Noise Mask and VAE Encode nodes and it's working great so far :)

@GrocksterRox 2 ай бұрын

Yup, just started playing with it a couple days ago and I agree it's pretty game-changing. I'm going to be introducing a new video on it pretty soon.

@Uday_अK 4 ай бұрын

Really appreciate your efforts!

@GrocksterRox 4 ай бұрын

Thank you so much!

@runebinder 4 ай бұрын

I was talking to someone last night about wanting to find a way to crop a masked area and upscale it for Inpainting and then shrink it back down to put it back and I'd installed some nodes to look at to try this, then you save me the effort with this video :) Awesome stuff, downloading to try it out now, thank you.

@GrocksterRox 4 ай бұрын

Great minds think alike! 😂 Good luck!

@runebinder 4 ай бұрын

@@GrocksterRox, I take it this could be retrofitted to other model types like SDXL, Kolors etc. by replacing the loaders and Custom Sampler? The reason I was thinking along these lines is because as far as I'm aware the crop/upscale/Inpaint/shrink/put back is what Fooocus does for it's Improve Details option in Inpaint, and as that's been the best tool for Inpaint it seemed like the way to try and replicate those kind of results in Comfy.

@runebinder 4 ай бұрын

@@GrocksterRox was so excited about the Inpaint I didn't notice the bonus at the end. I've got the beta toolbar enabled at the top and I still have Crystools showing in it, so you don't necessarily have to sacrifice it. It just appeared automatically for me when I enabled it, so not sure why it doesn't show for you.

@GrocksterRox 4 ай бұрын

@runebinder oh, that's awesome, I'm going to have to tinker with it then. Resource monitor is a must have

@runebinder 4 ай бұрын

@@GrocksterRox, definitely. The node had an import failed error for a couple of days once after installing some other nodes and it drove me up the wall not having. Just used your Inpaint workflow to mask a mangled hand holding a sword for a D&D character portrait I generated earlier and it's one of the best results I've seen on an Inpaint. It's exactly what I've been after :)

@yu-gg 4 ай бұрын

I learned a lot watching this, thanks :)

@GrocksterRox 4 ай бұрын

That makes me so happy, my goal is to help as many people in the community as possible!

@ShubzGhuman 4 ай бұрын

great guide bro

@GrocksterRox 4 ай бұрын

Thank you so much!

@sven1858 4 ай бұрын

nice video, learnt about a few new nodes with this.

@GrocksterRox 4 ай бұрын

Awesome, so glad it was helpful (every tip and trick helps)

@MilesBellas 4 ай бұрын

Interesting. Thanks ! 😊👍

@GrocksterRox 4 ай бұрын

Absolutely! Glad it intrigues and thanks so much for sharing with others to help their learning.

@b3arwithm3 3 ай бұрын

thanks so much for this excellent tutorial. Which method would you recommend for my use case? I want to generate several pictures that happens in the same place. Let say, I have a picture of a hotel lobby. I want 2 men talking in the lobby. Another image with a woman waiting for the elevator. Just need them to roughly. Any tips?

@GrocksterRox 3 ай бұрын

Thanks so much. Depending on if you want the woman and the two men to be in the same picture, you would probably want to do some compositing that line up to a final img2img to bake in the image, but I'd have to learn more about the situation to help you. Feel free to jump on the discord and we can chat (link in the vid description)

@thibaudherbert3144 4 ай бұрын

hey thanks for the tutorial. what is call the tool you are using for tha face animation in your videos ?

@GrocksterRox 4 ай бұрын

I vary between live portrait and Hedra, also looking at Fuze lately as well.

@faycaltech5922 2 ай бұрын

Thanks for this tuto, when i download the worklow (inpainting advanced), it's not completed like here

@GrocksterRox 2 ай бұрын

Hmm... It's the same process (but many of the nodes are collapsed for cleaner organization). You can expand the nodes to have more control over the process. Good luck!

@zRegicideTVz 4 ай бұрын

I learned a lot after watching this, thank you so much. Do you happen to have a WF that can change a piece of clothes or object to different color?

@GrocksterRox 4 ай бұрын

So glad! Yes, definitely have done clothes changing as well as batch changes of clothes - kzbin.info/www/bejne/infJfHhpf96sY9k

@zRegicideTVz 4 ай бұрын

@@GrocksterRox thank you so muchhhh

@GrocksterRox 4 ай бұрын

@@zRegicideTVz anytime

@TentationAI 4 ай бұрын

Nice thanks a lot, for the last workflow at the end ( firework ). Can we replace the node " load diffusion model " by " unet loader (GGUF) " for load as model flux dev Q8 ?

@GrocksterRox 4 ай бұрын

Definitely, you can use any model you want for inpainting (no flux specific models for inpainting, at least right now) and it works great!

@divye.ruhela 4 ай бұрын

Does the Q8_O GGUF model work with 12GB cards? Is it better than the FP8 model?

@GrocksterRox 4 ай бұрын

It may struggle a little, but I just spoke with someone earlier today that's used it successfully on a 12GB card, so you may squeak by using it successfully. Of course there are also online services (like Runpod) that you can rent hardware at a very reasonable rate.

@tuurblaffe 2 ай бұрын

gguf dlows down my generation speed i use it for clip/t5 but for models are very slow when i use gguf

@GrocksterRox 2 ай бұрын

Thanks again, I've noticed exactly the same thing in terms of rendering speed for these models.

@Novalis2009 4 ай бұрын

Sorry, where do I find the gguf model? Can't find the link ...

@GrocksterRox 4 ай бұрын

No worries, the Flux LORA Inventory is here (the models are on the 3rd tab) - docs.google.com/spreadsheets/d/1543rZ6hqXxtPwa2PufNVMhQzSxvMY55DMhQTH81P8iM/edit?gid=1074472502#gid=1074472502 The WoW GGUF model is here - civitai.com/models/147933?modelVersionId=756568

@Pauluz_The_Web_Gnome 4 ай бұрын

Thanks bro! Nice vid! Greetz...Pauluz RTX 😂

@GrocksterRox 4 ай бұрын

Always appreciated!

@AberrantArt 3 ай бұрын

Is flux required for this method? Flux always crashes on me so i have to use other model checkpoints.

@GrocksterRox 3 ай бұрын

No, you can definitely use this layering method in SDXL as well, but the LORA is specialized for flux

@PeterLunk 4 ай бұрын

Niceone :)

@GrocksterRox 4 ай бұрын

Thanks so much!

@831digital 4 ай бұрын

Instead of drawing the mask, can you do this with something like SAM2 and video?

@GrocksterRox 4 ай бұрын

Very possible, it's easy to adapt many of these methods to video generation, though the video models are still in their infancy (at least the local ones)