Thanks my dude! Awesome work development looking forward to more practical examples

Absolutely amazing insight too seeing someone using it professionally. And the fact that so many youtubers fail on is the over complicated large scale graphs that add very little improvements but lots of process. This is professional, to the point with the aim of delivering. Very refreshing tutorial/information. ❤

@risunobushi_ai

Ай бұрын

Thank you for the kind words! Yeah, sometimes I find workflows that I can't properly understand even after inspecting them and testing them carefully. That's why I always try to put notes, or make them as linear as possible. ComfyUI is scary enough already without the need to overcomplicate things!

Love Barragán, loved this.

Hey, you should do a full tutorial series of Comfy UI and charge for it on Udemy! I damn sure would be one of your first students! Anyways, good stuff! You just earned a new subscriber!

@risunobushi_ai

2 ай бұрын

Thanks for the kind words, however I believe that open source knowledge should be free and accessible :) I’d rather monetize with consulting and by integrating my workflows in my actual job, since that would be applying the knowledge to a commercial use case

Amazing work and breakdown on a very practical professional usecase. Thanks for sharing your experience.

@risunobushi_ai

2 ай бұрын

Thank you. Yep, it’s not every day that one can disclose almost everything that happened during a production in the fashion industry. I think it’s great to be able to share it, apart from the actual results.

@paulmarique

2 ай бұрын

The discussion/brief with the client was really interesting and precious. I would have love to also see a few rejected proposals from your research (ex: the 3/4 view). Is it possible to see Retina Studio's work somewhere?

@risunobushi_ai

2 ай бұрын

Noted, I will integrate those kind of things in the future as well!

so cool to see your subscriber count growing so fast. you deserve it!!!

@risunobushi_ai

Ай бұрын

Thank you! I’m lucky it’s been going so well

One of the real use cases for Stable Diffusion and ComfyUI

@risunobushi_ai

2 ай бұрын

Building workflows in comfyUI for these kinds of things is an absolute pleasure

Liked and Subscribed! Great tutorial. I've been using something similar for work. It's been a dream!

@risunobushi_ai

2 ай бұрын

Happy that you found this video valuable!

Can we get an in-depth tutorial for beginners like me

@risunobushi_ai

Ай бұрын

I’ll check if there’s enough material to do a full video about it, since most of what I did is already covered here. What are you having troubles with? So I know what I’d need to focus on, and / or I can give you a hand over here as well.

Great stuff! Where can one get the boring LoRAs? Can't find them on civit :(

@risunobushi_ai

2 ай бұрын

There you go! civitai.com/models/310571/boring-reality

Can you link to the Model Sheet civit page? Would be interested in checking it out. Great video!

@risunobushi_ai

2 ай бұрын

We used this one, although the note beside it in the original workflow said “not sure if this is even doing anything”: civitai.com/models/285203/turnaround-model-sheet-character-sheet-xl But in the heat of the moment we couldn’t afford to spending time swapping it in and out of the workflow, so we kept it in for the character turnaround sheet.

Hey, I have one issue with IPAdapter. After installation & adding a plus face preset, "ClipVision model not found" error pops up. Thanks for the valuable case study!

@risunobushi_ai

Ай бұрын

Have you installed *all* the models for your checkpoint (1.5 or SDXL) and the clip vision models from the GitHub page? You might have missed one or two

I am a bit lost in my work, I just want to share it here, what was my idea to start with. I was trying to train my custom model who is my brother so I took his pictures as I learned from other youtube tutorials how to train like best custom model. But I think my problem is even at the beginning. I cant solve this issue. Training my brother did not end up good in many ways. İt shows some weird face impurities and secondly face of my brother does not exactly the model I wanna use. So I came up with this idea I am really not sure trying it out. creating humans in blender by using human generator add-on so that ı can customize my model or as a same way using unreal engine and creating metahuman. then making virtual photos as intended and training again. also okey lets say none of them works as intended, is there any way finding a custom model Lora, at least better than mine. your custom model is perfect. Also using foocus and making face-swap for wanted positions, that what I thought. I am very new but I am doing my project so I really wanna make a progress. Maybe I just stop thinking about my custom model, ı just need to progress on the Stable Diffusion so that I will understand how the work flow is.

@risunobushi_ai

2 ай бұрын

If I understand correctly, what you’re trying to do is training a LoRA, whereas the custom model here is the Checkpoint. The faces were transposed via IPAdapter Plus Face, rather than training a LoRA for each character. With 24 hours to do the whole workflow, training LoRAs for each character would have left no room for error, so that wouldn’t have been feasible. Training LoRAs is a whole different beast, and while there’s many tutorials on that, it’s still a matter of trial and error with a lot of different things: for example, number of images, number of epochs, the kind of captioning you use, etc. there’s so many possible points of failing it’s hard to pinpoint what goes wrong without knowing each passage.

@aysenkocakabak7703

2 ай бұрын

@@risunobushi_ai thank you for the commnet. I got it in a way that how whole process start. Thank you

interesting project , tho the people dont look very good , their proportions are elongated , the mans face has bizarre triple teeth , why were these selected ? the elongation is something notice can happen when rendering a portrait aspect ratio and or trying to synthesize outside of the aspect ratio of the training data , in those cases sometimes better to render square image at the trained size ( 512x512 or 768x768 ) then upscale , that way the proportions are less likely to stretch . good start to a setup however needs a lot more refinement or better selection , further inpainting or detailer fixes .

@risunobushi_ai

2 ай бұрын

The bizarre aspects of the characters is a matter of Art Direction. Barragán’s communication is often bizarre, quirky, and at times grotesque, so the characters had to keep the same vibe. I know this sounds like the usual “it’s like this because I wanted it to be like this”, but in this case the characters where provided as is by the CD because of that. Edit: to add to that, funnily enough my first tries were deemed too normal, and Guillaume, the CD, knowing the client well wanted the characters to be more imperfect and grotesque. It’s all a matter of embracing the glitches, as the interview segment of the video talks about.

stable diffusion will slowly die and become irrelevant

@risunobushi_ai

2 ай бұрын

Well, that’s the cool thing about open source software: it literally can’t die! It can become irrelevant, yeah, but as long as there’s a community of devs who wish to keep it alive, it won’t die. Although I’m not sure demonstrating an actual use case that neither MJ nor Dall-E can do qualifies as dying. But who knows!