ControlNet Deep Dive - Depth - Preprocessors, Weight and Guidance, and Generating at Max Resolution

This video will teach you everything you ever wanted to know about using the Depth model in Stable Diffusion ControlNet. First, I will teach you the strengths and weaknesses of the two Depth Pre-processors, MiDas and LeRes, as well as how their settings impact the generated depth maps.
Weight and guidance are the most important settings for using ControlNet. I will teach you what they do, how they impact the output, how they are impacted by other variables, and provide recommended settings for each.
Although ControlNet is much better at generating high resolution images compared to baseline Stable Diffusion, there are still a couple challenges, which I will teach you how to mitigate or avoid. Finally, I will provide you with my personal method for using ControlNet Depth to generate images at the maximum resolution your GPU is capable of.
Intro 00:00
Pre-processors 01:49
Pre-processor Settings - 03:52
Weight and Guidance Intro - 05:58
Guidance - 06:49
Weight - 08:31
Generating at High Resolution - 09:37
Method for Generating at Maximum Resolution - 14:23
Outro 15:48
CFG Deep Dive: • Stable Diffusion Deep ...

[@ShawnFumo]

Thanks as always for such detailed videos. One thing I'd just like to remind viewers of is that you aren't limited to depth and depth_lres for generating depth maps. It is just an image, so anything from alternative depth detectors, blender, even painted from scratch photoshop can be used. Or you could use both the default depth for closeup and lres for background and combine the two in an image editor to get details at all distance levels.

[@siliconthaumaturgy7593]

That is a good point.
I try to limit myself to features within A111 or SD for simplicity, but this is definitely an option if you find the built-in preprocessors lacking

[@Netsuko]

Fantastic video! I have only been messing around in controlnet without really understanding the differences. This all makes so much more sense now. Thank you!

[@smire2591]

These detailed explanations is what is often missing in all kinds of tutorials - please keep up your good work!

[@fr0zen1isshadowbanned99]

Great Video!
Guides like these are very welcome and needed by new Creators.
For me, they are a little late, but I appreciate having my theories confirmed and expanded on :P

[@107cdb]

Awesome work!

[@ghostsquadme]

This is great. Thanks!

[@moriavarda2566]

Thank you! Nice, short and intelligible

[@hrmpk26]

Textbook quality material. Thank you.

[@lukas5220]

i never clicked a video so fast

[@alecubudulecu]

Shout out to Sekiro! Glad you had that

[@dobaovuongtamdamkt4197]

for super hi res with detail I use tiled MultiDiffusion, not many people know about it so far, but it will esentially split the generated image into multiple tiles, so you can generate as much as you want, as long as you wait for all the tiles to be completed

[@gorkskoal9315]

Multidiffusion, can also handle having a combination of items in a prompt better...(like a road with traffic)

[@moon47usaco]

How about guidance start and end. I would love to see your test results from that. Packed full of information. Thank you. =]

[@siliconthaumaturgy7593]

Guidance end is going to behave exactly the same as guidance in the old version of A1111
Haven't done much testing with guidance start, but you really want those early steps to use controlnet because they have so much impact on the final image.

[@user-cw3nb8rc9e]

Combine both? Midas and Leres?