Excellent lecture! More of these, please! It's always fascinating to learn how things actually work under the hood with examples. You're approach is great. You don't just tell people "do this and that and you're done". You actually explain things like a good teacher. You and Matteo (Latent Vision) are my favorite Stable Diffusion/ComfyUI youtubers. I never miss a video from you, guys. 👍
Wow! Comprehensive, very useful, interesting and visual! What software did you use to present Unet in 3d ? Also, it would be interesting to know if the same t2i adapters differ in any significant way from standard controlnet. Plus there is also the coadapter (also for 1.5 models) I think they were trained a bit differently, but there are usable models. For example there is colors, it is much better at rendering colors compared to tile and inpaint models
Not sure about the other controlnets, but from my experience, the SD 1.5 tile controlnet is still unbeatable to this day. I've tested pretty much all tile controlnets, including 2-3 tile controlnets for SDXL and the latest one for Flux. SD 1.5 still seems to produce the best creative image with the richest detail. If anyone cares to explain why, I'd be interested to hear. I would expect the Flux or SDXL tile controlnets to be the best, but that's not really the case in my experience.
Please test ttplanetsdxlcontrolnettilerealisticV2 with advance Controlnet extension and soft weight node+ Perturbed-Attention Guidance The result will be amazing 🤩 civitai.com/models/330313/ttplanetsdxlcontrolnettilerealistic
@@ArchAi3D I've tried the TTplanet tile controlnet V2 before, but maybe my settings were not correct, because I was not getting good results when doing img2img creative upscale of 3D renderings. I will need to try again the way you explained. Thank you!
nice, very interesting! I'm trying to understand why we would keep usind SD1.5 controlnets vs SDXL? also, isn't the "soft weights" for 'my prompt is more important' the same thing as lowering the strength of the controlnet? what's the difference? Cheers!
some SD1.5 ControlNet trained with higher step so they can work better, soft weights change the ControlNet strength inside the Unet layer and then it can make same affect as: my prompt is more important
@@ArchAi3D thank you, I didn't know this feature "my prompt is more important", and I still don't grasp exactly how it differs from modifying the strength of controlnet. Thanks
Is there a difference is speed if multi controlnet node for multi controlent in your workflow compared to using single controlnet and combining them one by them. Doesn't that increase the dimensions of the vector in latent space?
How can we combine controlnet with inpaint - since that usually becomes too constrained and the somehow the model is unable to follow the prompt better
Excellent lecture! More of these, please! It's always fascinating to learn how things actually work under the hood with examples. You're approach is great. You don't just tell people "do this and that and you're done". You actually explain things like a good teacher. You and Matteo (Latent Vision) are my favorite Stable Diffusion/ComfyUI youtubers. I never miss a video from you, guys. 👍
Thank you
A very clear explanation and some useful tips that other people tend to miss out. Thanks!
Great! Thanks for providing such clear details!
Kudos to your patience ❤❤
Thank you! 😊
Thank you for the lessons, you give a deeper understanding of working with neural networks and help solve applied problems
thanks for this excellent tutorial ❤
You're welcome 😊
Such an amazing video
Thank you for this video!!! Very useful
Wow! Comprehensive, very useful, interesting and visual!
What software did you use to present Unet in 3d ?
Also, it would be interesting to know if the same t2i adapters differ in any significant way from standard controlnet.
Plus there is also the coadapter (also for 1.5 models) I think they were trained a bit differently, but there are usable models. For example there is colors, it is much better at rendering colors compared to tile and inpaint models
Not sure about the other controlnets, but from my experience, the SD 1.5 tile controlnet is still unbeatable to this day. I've tested pretty much all tile controlnets, including 2-3 tile controlnets for SDXL and the latest one for Flux. SD 1.5 still seems to produce the best creative image with the richest detail. If anyone cares to explain why, I'd be interested to hear. I would expect the Flux or SDXL tile controlnets to be the best, but that's not really the case in my experience.
Please test ttplanetsdxlcontrolnettilerealisticV2 with advance Controlnet extension and soft weight node+ Perturbed-Attention Guidance
The result will be amazing 🤩
civitai.com/models/330313/ttplanetsdxlcontrolnettilerealistic
@@ArchAi3D I've tried the TTplanet tile controlnet V2 before, but maybe my settings were not correct, because I was not getting good results when doing img2img creative upscale of 3D renderings. I will need to try again the way you explained. Thank you!
nice, very interesting! I'm trying to understand why we would keep usind SD1.5 controlnets vs SDXL?
also, isn't the "soft weights" for 'my prompt is more important' the same thing as lowering the strength of the controlnet? what's the difference?
Cheers!
some SD1.5 ControlNet trained with higher step so they can work better, soft weights change the ControlNet strength inside the Unet layer and then it can make same affect as: my prompt is more important
@@ArchAi3D thank you, I didn't know this feature "my prompt is more important", and I still don't grasp exactly how it differs from modifying the strength of controlnet. Thanks
Is there a difference is speed if multi controlnet node for multi controlent in your workflow compared to using single controlnet and combining them one by them. Doesn't that increase the dimensions of the vector in latent space?
How can we combine controlnet with inpaint - since that usually becomes too constrained and the somehow the model is unable to follow the prompt better
Are the kolors controlnets better than using the xinsir controlnet with the adapter for kolors?
Yes I think it’s better, I will test and share the results