Replies: 5 comments 17 replies
-
You are missing out. I wouldnt recommend Euler. I render with 7 UniPC steps for drafting without controlnet and get a glimpse of what SD thinks of my prompt in no time, and then I can continue or select seeds with higher steps. With Controlnet I only need 3 UniPC steps in txt2img!! Thats the minimum, okay quality would begin at 4 steps, if controlnet is activated. Tip for UniPC: the lower the steps, the more you profit from lower CFG. In my 3 step experiments I had the cfg at 1,7 in the end to make it work (and it does not mess with the prompt interpretation). UniPC likes low CFG values in low steps This sampler is blazingly fast if you know how to tweak the maximum out of it. I recommend setting UniPC variant to "bh2" and UniPC covariant to "time_quadratic" in settings under "Sampler parameters". This helps with artifacts in low steps. If you still get artifacting or blur you need to lower cfg |
Beta Was this translation helpful? Give feedback.
-
Another tip to squeeze out every last megabyte of vRAM: Open up Task Manager, head on over to the Details tab, rightclick a column title and click Select Columns. Find Dedicated GPU Memory, tick it and click OK. Now sort the list by Dedicated GPU Memory to see what's eating up your vRAM. Open that program and see if you can disable Hardware Acceleration (similar to the browser setting in the OP). For example, Discord likes to nom on vRAM, so click the settings cog ⚙near the bottom left, go to Advanced and disable Hardware Acceleration there. |
Beta Was this translation helpful? Give feedback.
-
thanks for the write-up, i think this is useful for a lot of people!
|
Beta Was this translation helpful? Give feedback.
-
@vladmandic if you have time to sus this out: my test prompt, without quotes in your UI, hits the 74/74 token count I noted in my notes and originally in this post. It still does, but in A1111 and Anapnoe, 80/78. My only guess, off the wall, maybe something to do with the token merging thing I've noticed in settings? |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
First of all for Windows 11 users , this is important! On my RTX 3090 TI , rough estimate on improved it/s was about 15-20% - Requires restart.
![image](https://private-user-images.githubusercontent.com/113237671/242502502-c62e207d-33da-4189-a5b9-1e1aea4e5ffa.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjMxMDkwNTMsIm5iZiI6MTcyMzEwODc1MywicGF0aCI6Ii8xMTMyMzc2NzEvMjQyNTAyNTAyLWM2MmUyMDdkLTMzZGEtNDE4OS1hNWI5LTFlMWFlYTRlNWZmYS5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODA4JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgwOFQwOTE5MTNaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0zMzliYTBjMGJlYjkzMTVkZDRlYWFkYTRhMWFjMDJmODQ3Nzc3ZDkyMzc4MjkyMjhlMWEyMTE0YzhmZTdiMGYwJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.fh3qDSlTzXiTvkyvBSCmsx885hV6Ow6S_4Hv4eN0H2c)
EDIT: Oh yeah, turn off hardware acceleration in your web browser settings!!!
![image](https://private-user-images.githubusercontent.com/113237671/242515531-f0b7c1fc-16bb-4b75-915a-441b6801421e.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjMxMDkwNTMsIm5iZiI6MTcyMzEwODc1MywicGF0aCI6Ii8xMTMyMzc2NzEvMjQyNTE1NTMxLWYwYjdjMWZjLTE2YmItNGI3NS05MTVhLTQ0MWI2ODAxNDIxZS5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODA4JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgwOFQwOTE5MTNaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT00YTI0ZTFjNWRiNzQyNmRjYjU0ZmMxOGVkMjBkMGJjMWU2YmI5ZWRmYjRlZTQ1ZjhiY2JkMjBkYTgyNzA0MWRiJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.1BLRcxA5hnHMwfLLQoX1Dph6c9w1sCUkaDnFzU45W0U)
Secondly, this one may be much more obscure, but it is important to BALANCE your prompt and negative prompt. Gains here, for me, anywhere 5-15% it/s on loose testing, eyeballing.
EX test prompt - tokens for both prompt and negative prompt are 74/74. To compare, simply unbalance it, clear the negative prompted entirely.While the concept remains true, this isn't the prompt I was thinking of. I am trying to locate it. The prompt here is giving me 80/78. I could have sworn it was the 74/74, it's how I have it notated in my notes, but it's incorrect.Prompt: masterpiece, realistic, (((perfect))), photorealistic, ultrarealism, (official art), 1girl, bobcut, blue eyes, brunette, serious look on face, (sleeveless red dress), (white thighhighs), (hightops), hands on handlebars, ((riding on motorcyle)), speeding, looking at viewer, highway, scenery, dusk, raining, puddles, bokeh, intricate, highly detailed, motion lines"
Negative: nsfw, lowres, ((mangled)), inaccurate, mutated, mutant, tumor, growth, cancer, guro, injury, violence, (deformed), (bad anatomy), (disembodied limbs), bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
I have a prompt that was 314/50-something, was generating results at about 5.2 it/s, when I bumped negative to 301, it's currently running at 5.88 it/s (315/301).
TIPS on balancing
TL;DR -
74/74, but the closer they are the better the it/sLong-winded -
trick I used to bump the tokens up in the negative was by using BREAKS. BREAKS are how my prompt crossed 300 mark in the first place due to the 'padding' breaks put between tokens to isolate 'cross-contamination.' By cross-contamination I mean if you specify 'a man is wearing a black shirt and driving a red car,' sometimes you'll have a black shirt black car, red shirt red car, from the contamination between the tokens 'black' and 'red.' Putting a BREAK between them isolates them from each other a little better, improving adherence to your prompt. As such, each BREAK inserts 75 token's worth of padding.
So to bump my 50 to 300 for my negative, I did 'BREAK blah.' I can't dbl check atm but will update - Vlad's token counter wasn't recognizing my BREAKS unless I followed it with something, in this case 'blah.' Anapnoe's counts every BREAK even if it's the only thing in the prompt. Not sure if there's a technical difference here as far as balancing prompts go, but something to maybe note.
GO BACK TO USING EULER/EULER A!!!!!! IF it/s is super important to you. There was something important to mention about DDIM too but I can't remember, for the life of me, what it was. I'll edit if I remember. I feel like samplers are largely preference, I dunno. I know at one time the Ancestrals were considered more creative and were not deterministic, ie. you could put in any number of steps and it would just continue to develop to your step count, for better or worse. I know the DPM++ Karras series and the modified series are popular, but really all they seem to do is double-sample per step or something? I'll need to refresh myself on the samplers. I still don't know the point o UniPC still haha, except that it won't work with AND syntax in prompts.
Additional thoughts -
Be interested to hear what other people might have to share or general thoughts on this.
Beta Was this translation helpful? Give feedback.
All reactions