/vt/ - Virtual Youtubers

>>58604386
cont. if you go with a higher actual CFG (anything approaching 30), you need to turn the threshold percentile up. I only use 17 cfg and 90-95 threshold most of the time. by the time you're at 30 you want at least 95%. anything above 30 you're going to get closer and closer to 99% or the image will fall apart.

had a hard time finding an example. this one is an example of something else - using a "down" scheduler to make ti start at high cfg and end real low. gives you a pastel effect without prompting it as the image drifts towards dreamy/low CFG while the gen progresses

>>52696616
>Dynamic CFG but I lever learned what it did.
CFG is the "guidance" (literally) for your proompt. It's normally a static number, you put 7 or 8 or 11 or whatever in the "scale" box and it's 7 or 8 or 11 or whatever the entire time. Dynamic CFG lets you introduce funny math functions into it, "start high, end low, slowly", "start low, go up really fast but only at the end", "start low, slowly go up", etc.
>SAG
An entirely different kind of proompt guidance. Everything had me thinking that the two would not work together...but they do, kek.
>>52697103
>>52697176
I use it like a complete retard, but it works for me.
>Sampler: DPM++ 2M SDE Karras, CFG scale: 17 >SAG Guidance Scale: 1.11, SAG Mask Threshold: 1
>Dynamic thresholding enabled: True, Mimic scale: 8.5, Threshold percentile: 93.3, Mimic mode: Power Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4, Scheduler value: 4
And I almost always use it the same exact way, mostly necessitated by early issues with SDE. Sometimes I'll play with either of the SAG values since that's newer to me. You have more leeway in how you can use Dynamic thresholding if you use a different sampler. If you do "the opposite" (start high, end low) with a really high clamp % you can make some really weird washed out but very "sharp" looking shit. Picrel:
>CFG scale: 22, Dynamic thresholding enabled: True, Mimic scale: 11, Threshold percentile: 100, Mimic mode: Half Cosine Down, Mimic scale minimum: 3, CFG mode: Half Cosine Down, CFG scale minimum: 3

>>45119284
still shows up even when I'm deep frying CFG with dynamic thresholding

>>44912718
MultiDiffusion and Tiled VAE have yet to work a single time for me. Either I'm retarded, or they're borked, so I'll be commenting them out of the Colab next chance I get to mess with it. YW. Cleanup for the launch form was long overdue. Make sure you thank ChatGPT too, the dumb cunt helped a bit when it wasn't sending me on a wild goose chase.
And yes, dynamic thresholding. It's very cool.
>Good job!
10q ;D
>>44913929
sampler matters. euler a will put an anchor on anything, kek
>>44913094
I much prefer comfy to speed running fucking threads...