Finally a channel that is actually teaching things intermediate users need to know. I'm so sick of every video showing how to install SD/Controlnet at the start for padding and then not much else of substance.
Thanks! That is why I started the channel. When I was starting out, I watched other TH-cam tutorials and found them a bit less informative than I personally prefer.
@@siliconthaumaturgy7593 indeed, i still hit my head against a wall when it comes to dreambooth training and Lora for some parameters. many written guide helped me understand what i was missing but as shown in your video. it comes with time and effort.
As Always thanks for providing this knowledge for free to the small "deep divers" audience. You work save us precious Time on question we didn't dare to research. I hope you keep on experimenting and stay motivated.
re: your section on excessive brightness or darkness. My experience is that this tends to happen when controlnet is "fighting" against what the model/prompt would usually want to produce. i.e. both sides are pulling in different directions. I've tended to solve this by improving the prompt so that the model's "standard" output more closely resembles the controlnet image, or by simplifying the prompt to give more flexibility and remove the conflict of model vs controlnet. Your solution also works of course; weakening the "strength" of either side of the fight lets one of the sides win more easily.
I've been trying to do Tiled Diffusion with the multi diffusion upscaler extension. And all I ever get is multiple images forced together ungracefully. A video on this extension and the tile controlnet model would be awesome. And your awesome too thanks for the videos 😊
Tiled Diffusion is on my radar and near the top of my list. It's probably going to end up being multiple parts since there is so much in that extension
Thanks for the tutorial. Short, but every second of the video is knowledge. I took a lot of notes from this 1 video. I would like to ask you why you think that highres fix is necessary for text to img? I haven't used SD enough to make a call. But from a professional designer's point of view, also using my not too cheap PC to try, adding this step does significantly increase the render time. So I incline to think that finetuning can be moved to img to img. But again, I am just a baby in SD. What do I know. Please explain a bit if possible. thank you. liked and subbed. xyz plot is amazing. 1 shot and we get to see all of them. Perhaps this is why you suggest using Hires fix everytime. Only issue is, I can totally feel the difference in terms of the wait.
10:00 for this part we can fix stuff with Adjusting prompt like you said and maybe adjusting other stuff like VAE, Lora for darken or lighten the image without destroying the composition and some CFG rescaling like "Neutral prompt" extension, There is some setting in aswell similar to this but from the webUI dev they said this one is a bit better (I'm forgot where in the setting tab as well. xd) 18:47 maybe we can use ctrlnet segmentation here. idk great content as always. ty and have a nice day. (I'm done rewatch this and maybe ill watch again if I forgot some stuff as most likely would happen in the future. xd)
Thanks. ControlNet Reference is definitely the most challenging to use feature I've covered yet, so thanks for your patience. Unlike most other ControlNets, almost everything seems to matter and have a chance to mess it up. It's like riding a raging bull. So I really appreciate the suggestions and I'll try them out soon.
do you know if theres a way to use reference-only without using a prompt? i have an image and i want it to be replicated, but i don't want to prompt it
best episode yet, very useful info, thanks ! Darksushi is a borked model thats overtrained on anime, you cant do photographs with it which is probably the reason for overbaking
cool content but hard to follow, the editing jumping around as always and it's kinda hard to focus on what you talking about and what I actually seeing atm. idk how to explain it. i think I need to rewatch this later. ty and have a nice day.
I agree as well. It would be helpful when showing a big grid of images to actually mouse over or highlight the specific ones that you’re talking about. There’s just too much to process all at once when you show the entire grid and it moves very quickly from scene to scene.
Finally a channel that is actually teaching things intermediate users need to know. I'm so sick of every video showing how to install SD/Controlnet at the start for padding and then not much else of substance.
Thanks! That is why I started the channel. When I was starting out, I watched other TH-cam tutorials and found them a bit less informative than I personally prefer.
@@siliconthaumaturgy7593 indeed, i still hit my head against a wall when it comes to dreambooth training and Lora for some parameters.
many written guide helped me understand what i was missing but as shown in your video. it comes with time and effort.
As Always thanks for providing this knowledge for free to the small "deep divers" audience.
You work save us precious Time on question we didn't dare to research.
I hope you keep on experimenting and stay motivated.
Yay! A channel providing high quality SD content. Subscribing!
The community misses you bro, your vids are still in a league of their own!
I love this channel. I use comfy UI but everything here I can translate to that. Thank you for these very informative videos.
Your style reminds me a lot of the "Battle(non)sense" yt channel, which is great. Comprehensive and to the point. I love your content
an incredible amount of work i should say ! pure madness to pull out in only two weeks !
Thanks. This one was definitely rough to research.
Thank you so much for all the contents you have provided! I never thought about using hires fix without upscaling!
re: your section on excessive brightness or darkness. My experience is that this tends to happen when controlnet is "fighting" against what the model/prompt would usually want to produce. i.e. both sides are pulling in different directions.
I've tended to solve this by improving the prompt so that the model's "standard" output more closely resembles the controlnet image, or by simplifying the prompt to give more flexibility and remove the conflict of model vs controlnet. Your solution also works of course; weakening the "strength" of either side of the fight lets one of the sides win more easily.
thank youu! I've been struggling with this tool and needed in-depth overview
Great job. Helpful, clear. Thanks
Just found your channel - Great stuff!
Great video as always!
you are the boss! (I'd love to see your take on outpainting:1.5)
Thank you.
I've been trying to do Tiled Diffusion with the multi diffusion upscaler extension. And all I ever get is multiple images forced together ungracefully. A video on this extension and the tile controlnet model would be awesome. And your awesome too thanks for the videos 😊
Tiled Diffusion is on my radar and near the top of my list. It's probably going to end up being multiple parts since there is so much in that extension
very helpful！！
Thanks
Thanks for the tutorial. Short, but every second of the video is knowledge. I took a lot of notes from this 1 video.
I would like to ask you why you think that highres fix is necessary for text to img? I haven't used SD enough to make a call. But from a professional designer's point of view, also using my not too cheap PC to try, adding this step does significantly increase the render time. So I incline to think that finetuning can be moved to img to img. But again, I am just a baby in SD. What do I know. Please explain a bit if possible. thank you. liked and subbed.
xyz plot is amazing. 1 shot and we get to see all of them. Perhaps this is why you suggest using Hires fix everytime. Only issue is, I can totally feel the difference in terms of the wait.
10:00 for this part we can fix stuff with Adjusting prompt like you said and maybe adjusting other stuff like VAE, Lora for darken or lighten the image without destroying the composition and some CFG rescaling like "Neutral prompt" extension, There is some setting in aswell similar to this but from the webUI dev they said this one is a bit better (I'm forgot where in the setting tab as well. xd)
18:47 maybe we can use ctrlnet segmentation here. idk
great content as always. ty and have a nice day. (I'm done rewatch this and maybe ill watch again if I forgot some stuff as most likely would happen in the future. xd)
Thanks. ControlNet Reference is definitely the most challenging to use feature I've covered yet, so thanks for your patience.
Unlike most other ControlNets, almost everything seems to matter and have a chance to mess it up. It's like riding a raging bull. So I really appreciate the suggestions and I'll try them out soon.
do you know if theres a way to use reference-only without using a prompt? i have an image and i want it to be replicated, but i don't want to prompt it
best episode yet, very useful info, thanks ! Darksushi is a borked model thats overtrained on anime, you cant do photographs with it which is probably the reason for overbaking
how can i made that in comfyui
cool content but hard to follow, the editing jumping around as always and it's kinda hard to focus on what you talking about and what I actually seeing atm. idk how to explain it. i think I need to rewatch this later. ty and have a nice day.
Yea, I felt the same, but simply pausing often to intake / review the grids helped a bunch.
I agree as well. It would be helpful when showing a big grid of images to actually mouse over or highlight the specific ones that you’re talking about. There’s just too much to process all at once when you show the entire grid and it moves very quickly from scene to scene.
Change the speed to 50% and turn on subtitles; hit space bar to pause/resume. Learn how to use YT.
Joke's on you - I _want_ human-animal abominations.