I've watched a few controlnet for beginners videos and have found it quite confusing, but this is by far the best and clearest explanation of the tool I've found. Thanks for taking the time to actually explain all the options rather than just telling us to click here and click there. Good work.
I may have misunderstood what you were saying about Hires Fix here but I use it all the time. It doesn't make your image a smaller resolution (which is how I interpreted your interpretation of it). Rather, it allows you to start generating your image as usual at a given resolution (e.g. I often set my main generation width x height to be 512x683) and then, halfway through the generation, it appears to re-diffuse the image slightly (so that it's a partially re-diffused version of the 512x683 image it initially generated), upscales to whatever height x width you specified under Hires fix (I usually set mine to Upscale by 1.5), and then continues generating from there (for me, resulting in a final image resolution of 768x1024). This has a couple of benefits: (1) If you don't have the beefiest GPU (mine has 6GB VRAM), it allows you to create larger, higher res images without crashing your UI; and (2) By taking that slight "un-diffusing" step at 50% generation and then continuing generation from there, it will very often fix (or at least significantly smooth out) any weird bits of anatomy or other visual gibberish that might appear in the initial generation. A good percentage of the time, I'll see the initial 50% of an image's generation (while it's still working at 512x683) have an extra leg, a missing arm, too many fingers, or whatever but then, once Hires Fix kicks in for the second half of the generation, those weird glitchy bits of anatomy get fixed. Any time I try running without Hires Fix checked, I find my images will often (though not always) come out significantly worse.
I've done all of this in the past few months and I'm getting good results with the process outlined in the video. Just wanted to drop a comment and let people who have got Stable Diffusion set up and ready to add some mods, that this is legit. About those canny low and high thresholds, technically they define how much high and low detail is filtered out of the black and white data image. You can allow almost all or none of it, but the other options are still going to determine the final image more. Setting the high, up high, leaves less allowance for the high data, and setting the low leaves less allowance for the low data - whatever they are, in the base noise image. For what I'm doing, I go with: Low threshold: 85 High threshold: 100 As a place to start that won't ruin your day while you test the other settings. You want your waifu doing that swimsuit model pose, start with 85/100. Come back and tweak that when you've mastered what the other options do. If you're getting more or less of the background detail than you like, it's probably the high/low thresholds.
Some awesome tips! I'm loving the experiments! Definitely going to be messing around with the threshold settings you mention Jason! Appreciate you watching and taking the time to comment!
I'm going crazy , I don't have the control net tab, everyhting installed updated, and files put into folder, but no control net tab, can someone help me pls?
Yeah i hate the same issue. You likely have an old version of stable diffusion. If you set it up properly the first time you may be able to update it by adding a line into the .bat file to launch it.
nobody told me on internet how to run these models. I dont know why people hinding this info ?? God bless you buddy
I've watched a few controlnet for beginners videos and have found it quite confusing, but this is by far the best and clearest explanation of the tool I've found. Thanks for taking the time to actually explain all the options rather than just telling us to click here and click there. Good work.
Thank you for the kind words! I put a lot into my tutorials 🙃
I may have misunderstood what you were saying about Hires Fix here but I use it all the time. It doesn't make your image a smaller resolution (which is how I interpreted your interpretation of it). Rather, it allows you to start generating your image as usual at a given resolution (e.g. I often set my main generation width x height to be 512x683) and then, halfway through the generation, it appears to re-diffuse the image slightly (so that it's a partially re-diffused version of the 512x683 image it initially generated), upscales to whatever height x width you specified under Hires fix (I usually set mine to Upscale by 1.5), and then continues generating from there (for me, resulting in a final image resolution of 768x1024). This has a couple of benefits: (1) If you don't have the beefiest GPU (mine has 6GB VRAM), it allows you to create larger, higher res images without crashing your UI; and (2) By taking that slight "un-diffusing" step at 50% generation and then continuing generation from there, it will very often fix (or at least significantly smooth out) any weird bits of anatomy or other visual gibberish that might appear in the initial generation. A good percentage of the time, I'll see the initial 50% of an image's generation (while it's still working at 512x683) have an extra leg, a missing arm, too many fingers, or whatever but then, once Hires Fix kicks in for the second half of the generation, those weird glitchy bits of anatomy get fixed. Any time I try running without Hires Fix checked, I find my images will often (though not always) come out significantly worse.
Super solid information! I did indeed misspeak on that one and appreciate you catching it!
I've done all of this in the past few months and I'm getting good results with the process outlined in the video. Just wanted to drop a comment and let people who have got Stable Diffusion set up and ready to add some mods, that this is legit.
About those canny low and high thresholds, technically they define how much high and low detail is filtered out of the black and white data image. You can allow almost all or none of it, but the other options are still going to determine the final image more. Setting the high, up high, leaves less allowance for the high data, and setting the low leaves less allowance for the low data - whatever they are, in the base noise image. For what I'm doing, I go with:
Low threshold: 85
High threshold: 100
As a place to start that won't ruin your day while you test the other settings. You want your waifu doing that swimsuit model pose, start with 85/100. Come back and tweak that when you've mastered what the other options do. If you're getting more or less of the background detail than you like, it's probably the high/low thresholds.
Some awesome tips! I'm loving the experiments!
Definitely going to be messing around with the threshold settings you mention Jason!
Appreciate you watching and taking the time to comment!
Excellent!
I'm going crazy , I don't have the control net tab, everyhting installed updated, and files put into folder, but no control net tab, can someone help me pls?
It’s a drop down inside of text2img or img2img not it’s own tab :)
@@HustleMillennial thx for your reply, yes I see it in videos but it is not there on my UI
just script nothing else, control net is checked in extensions
Yeah i hate the same issue. You likely have an old version of stable diffusion. If you set it up properly the first time you may be able to update it by adding a line into the .bat file to launch it.
👍
much love!
A wonderful presentation. Thank you for the efforts you have made to put together and make this video. It was clear, thorough and followable.
thank you for the tutorial, it was very helpful. cant wait to see the ones about inpainting and img2img