I couldn't be bothered to rebuild this workflow so I paid the £3 on Patreon. Its the first workflow I have ever paid for, but I am happy to support The Nerdy Rodent he makes such good content!
YES! Problems from weeks you solve with this video. My research on sampler/schedulers with diffrent models are on an extended level and now you give me this intresting function. This is the right time for this! Thanx a lot 🙏
I am really liking these settings for a more realistic looking image with my Flux model: Sampler = dpmpp_2m Custom Sigma Schedule: 14 Steps tensor([1.000, 0.9413, 0.8777, 0.7760, 0.7500, 0.6700, 0.6045, 0.5600, 0.5300, 0.4200, 0.3700, 0.2800, 0.2400, 0.2200, 0.0000])
Setting the first sigma to less than 1.0 will let you do similar things like img2img for any model, being able to absorb colours from the base latent with values near 1.0 but not quite, and more like img2img when hitting 0.5-0.7 ish. Mileage varies based on model and step counts.
Super interesting - even with regard to stripes, again! 😂 I recently found that shuffling the MLP of T5 only (!) over a large amount of layers (2,3,4,5,8,9,10,11,12,13) is able to reliably produce those "stripes of complaints" in about a third of random_seed. So, I started with your proposed methods and then refined to 20 sigmas (because good images happen from 20 steps on, and because I am an madcap nerd if you will), and tried to fix the deterministic stripes (for a given known seed) induced by shuffling T5's MLP - where the stripes are not a background nuisance, but kinda the main focus of the image. Ignoring the fact that the meaning of the prompt is obviously lost with this shuffle, and focusing on stripes only: "Heun" and "dpmpp_2s_ancestral" were the only ones that fixed the Shuffle-Stripes at 1024x1024. So I tried 2048x2048 - and without shuffle. With Euler and sigmas as predefined, stripes / patchy artifacts ensure. With "dpmpp_2s_ancestral" and roughly following a soft version of "beta" (don't keep it quite as high initially, decay to zero just a little slower, spread it out over one more step) - flawless image at 2048x2048! 🥳 Only at 2800 x 2048, the patchiness reappears. But it's a bad image at this non-square super-res anyway, and you'll wanna upscale to archive this - don't directly sample such image, lol. Pretty awesome tutorial again, TY! 🙏🤓
For the real Nerds that want to experiment on deeply technical issues like custom sigma schedules ComfyUI is really the best choice, and a great tool, its just to the layman it can seem intimidating.
I couldn't be bothered to rebuild this workflow so I paid the £3 on Patreon. Its the first workflow I have ever paid for, but I am happy to support The Nerdy Rodent he makes such good content!
😃
YES! Problems from weeks you solve with this video. My research on sampler/schedulers with diffrent models are on an extended level and now you give me this intresting function. This is the right time for this! Thanx a lot 🙏
Glad it helped!
off to your room, you're rounded. nice.
I am really liking these settings for a more realistic looking image with my Flux model:
Sampler = dpmpp_2m
Custom Sigma Schedule:
14 Steps
tensor([1.000, 0.9413, 0.8777, 0.7760, 0.7500, 0.6700, 0.6045, 0.5600, 0.5300, 0.4200, 0.3700, 0.2800, 0.2400, 0.2200, 0.0000])
well done, i know most of these but great tips for many people.
So it is the scheduler that's been messing up the images! Thanks for the info.
Setting the first sigma to less than 1.0 will let you do similar things like img2img for any model, being able to absorb colours from the base latent with values near 1.0 but not quite, and more like img2img when hitting 0.5-0.7 ish. Mileage varies based on model and step counts.
Thx Nerdy. U rock
Super interesting - even with regard to stripes, again! 😂
I recently found that shuffling the MLP of T5 only (!) over a large amount of layers (2,3,4,5,8,9,10,11,12,13) is able to reliably produce those "stripes of complaints" in about a third of random_seed.
So, I started with your proposed methods and then refined to 20 sigmas (because good images happen from 20 steps on, and because I am an madcap nerd if you will), and tried to fix the deterministic stripes (for a given known seed) induced by shuffling T5's MLP - where the stripes are not a background nuisance, but kinda the main focus of the image.
Ignoring the fact that the meaning of the prompt is obviously lost with this shuffle, and focusing on stripes only: "Heun" and "dpmpp_2s_ancestral" were the only ones that fixed the Shuffle-Stripes at 1024x1024.
So I tried 2048x2048 - and without shuffle. With Euler and sigmas as predefined, stripes / patchy artifacts ensure. With "dpmpp_2s_ancestral" and roughly following a soft version of "beta" (don't keep it quite as high initially, decay to zero just a little slower, spread it out over one more step) - flawless image at 2048x2048! 🥳
Only at 2800 x 2048, the patchiness reappears. But it's a bad image at this non-square super-res anyway, and you'll wanna upscale to archive this - don't directly sample such image, lol. Pretty awesome tutorial again, TY! 🙏🤓
Yeah, going over 3000 is tricky 🫤 Still, I’m ok with the 2-3k range!
Amazing
Nice.
🛌🛌🛌very comfy video
😉
Hi 👋
👋
Good, you're thinking like a NERD 😂
I don't think that Comfyui is too popular for a lot of people. You don't seem to have many views.
For the real Nerds that want to experiment on deeply technical issues like custom sigma schedules ComfyUI is really the best choice, and a great tool, its just to the layman it can seem intimidating.