Stable Diffusion 3 on ComfyUI: Tutorial & My Unexpected Disappointment

Подписаться 19 тыс.

Просмотров 12 тыс.

50% 1

🌟 Subscribe in the upcoming course get 40% off: aiconomist.gumroad.com/
Learn how to install and run Stable Diffusion 3 on ComfyUI! This tutorial shows you step-by-step how to set up the latest AI image generation model, but be warned - it might not live up to the hype. Find out why Stable Diffusion 3 might leave you feeling disappointed, even with its impressive capabilities, and discover why I'm sticking with SD 1.5 and SDXL for now. Get ready for a revealing look at the exciting world of AI art, even when it doesn't go as planned!
🌟 SD3: huggingface.co/stabilityai/st...
0:00 - INTRO
0:29 - Understanding SD3
1:39 - SD3 Weights
3:06 - CLIP Models
4:15 - Update ComfyUI
4:26 - Understanding SD3
5:25 - Examples
𝕏 Follow me : X.com/Aiconomist1
💼 For Business: aiconomistbusiness@gmail.com

Наука

Опубликовано:

27 июн 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 48

@baheth3elmy16 15 дней назад

Its extremely bad with humans. Very disappointed with it.

@bytesundbuechse 13 дней назад

I think its useless for humens. Super disappointing

@17th_Colossus 14 дней назад

SD3 shows why censorship is always a bad thing.

@tails_the_god 9 дней назад

EXACTLY i been saying how censorship is bad especially for AI

@Bicyclesidewalk 14 дней назад

Back to playing around with Stable Cascade - which can generate some really great images.

@bladechild2449 14 дней назад

All been downhill since 1.5 SDXL was ok when the custom trained models peaked, but theres been far less support for it LORA wise etc. We really need to just get our hands on the LLM supported prompting models now. I'd love an offline version of whatever ideogram uses but with trained models and customisation

@MajomHus 14 дней назад

You can always do your own lora if you want a specific thing.

@bigdaddy5303 13 дней назад

Some of the IP adapter face models and instant ID give better results than any Lora I ever created. With a xl turbo model and a good comfy UI workflow you can pump out amazingly realistic "Lora's" in 10 steps and a single photo

@MajomHus 13 дней назад

@@bigdaddy5303 lora or a well trained checkpoint is always better, but takes days to make

@WillOfPower321 13 дней назад

@bigdaddy5303 I would like to see your workflow, I'm very interested. I was trying with IP Adapter recently and was trying to determine whether it is superior to Loras.

@MajomHus 13 дней назад

@@WillOfPower321 a well trained lora will always beat ip adapter for now.

@AI.Absurdity 15 дней назад

So far in my testing, it does text a lot better.

@taucalm 14 дней назад

😁

@ciphermkiii 12 дней назад

Doing text better is literally one of the advertised improvements.

@amitjangra6454 15 дней назад

I was downloading it with high expectations

@Aiconomist 15 дней назад

Sorry for ruining it! 😅

@WillOfPower321 15 дней назад

Looks like PixArt is the future.

@GumOnTheWall 14 дней назад

I'm thinking that Stable Cascade might take over

@soucravn 14 дней назад

mate you sure you got a 3060? It took 11 secs for me on the same sample workflow

@gu9838 15 дней назад

yeah its so so......at least the base model i think xl lighting is best for now

@mjkht 14 дней назад

what is about HunyuanDiT?

@memerhuwhite 14 дней назад

Thanks for the video, Brother! Good to meet another IDM user too 🤣😎👊🏻

@alessandrorusso583 11 дней назад

Sd3 It sucks enough. So far I haven't seen any videos, among the hundreds, on RU-vid, that show a generation with a serious prompt, all using the example prompt, by the way very general, that they give with SD3.

@styrke9272 14 дней назад

SD 2.0 all over again ?

@duskairable 14 дней назад

yup DOA

@AltimaNEO 14 дней назад

Is say worse. At least 2,0 could do some anatomy and coherent scenes .This one fails so bad at everything

@styrke9272 14 дней назад

Sadge

@blacksage81 14 дней назад

Oof, good thing I'm set with my XL Models

@wonder111 15 дней назад

Not impressed either, though once they are properly plugged into existing quality SDXL workflows, I expect my opinion will change. As raw testing models, I found it rather literal.

@paoloadam 14 дней назад

I am still hopeful -- SDXL also failed to impress at first and took a while to mature, and this is just a preview not the final product. I am almost finished downloading, I'm very curious to see whst are supposed to he its main advantages, text rendering & most importantly a significantly greater capacity to understand prompts and thus deliver the right image.

@WNTrombone 14 дней назад

That’s unfortunate especially after all the hype they built around it…

@taucalm 15 дней назад

I would give it time for year to end, it was released like today so it cant be finetuned yet neither community driven.. sdxl has been around at least 6mo and I guess it was not as good as it is today when it was released..

@Cara.314 15 дней назад

Except nobody can fine tune it cause of the licensing.

@squirrelhallowino29 15 дней назад

XL was released a year ago. However this SD3 is not the full model, but it's not looking promising. Plus no increase on resolution which is a big limitation right now of any model.

@taucalm 14 дней назад

@@Cara.314 well that makes it pretty much unusable, it should be opensource to achieve same results as sdxl.

@GustavBoye-cs9vz 8 дней назад

Something feels weird about a company that trained in millions on artists work without consent talk about their license and terms of use. - Don't get me wrong I'm not some sort of anti-ai warrior that writes "AI sucks" on every AI image. - But to see them try to enforce copyright on users when they themselves cared little about it is a bit funny. - I think at some point there will be a complete open source version of some of these models.

@GustavBoye-cs9vz 8 дней назад

Also impressive work they did, but at the same time not. The newest discoveries in the field is thanks to AI scientists. The dataset is thanks to unappreciated artists. All they really did was just throw a ton of money and compute at it. Which basically anyone could do in theory. And all that compute will be a waste once a smarter architecture or training method is discovered.

@Jeremy-sk9qh 15 дней назад

you downloaded the text encoders not the model

@Cara.314 15 дней назад

Incorrect. They downloaded and explained both clip and unet models.

@nelathan 14 дней назад

your tts is so wrong

@StefanReich 13 дней назад

Are you using an AI voice? I really prefer actual humans

@user-ld8sy9xu2v 11 дней назад

Yeah but what if a guy cant speak a good english,TTS voice with clear pronouncing is always better then person with strong accent and when you have hard time to listen what he/she actually said. And TTS is not that bad at all.