FYI: Full Integration - Githubissues

megvii-research / HiDiffusion

[ECCV 2024] HiDiffusion: Increases the resolution and speed of your diffusion model by only adding a single line of code!

Apache License 2.0

771 stars 43 forks source link

FYI: Full Integration #12

Open vladmandic opened 6 months ago

vladmandic commented 6 months ago

I like HiDiffusion project a lot, results are VERY good!
So i've integrated it into SD.Next (with full credits)

Also added a bit more tunables instead of using fixed values (enable/disable raunet, msw-msa, set attention steps, tune t1/t2 values)
Attaching example XYZ grid created with different T1/T2 ratios for SD15 at 1k and 2k and SDXL at 2k:

sdnext-hidiff-sd15-1k-compressed

sdnext-hidiff-sd15-2k-compressed

sdnext-hidiff-sdxl-2k-compressed

attashe commented 6 months ago

Great! Thanks for your work

ShenZhang-Shin commented 6 months ago

Wow, This is excellent! I am going to try it.

recoilme commented 6 months ago

@vladmandic could you provide some information on how to use this or where it is highlighted? Didnt find any info in changelog about this...

vladmandic commented 6 months ago

it's in dev branch and enable just with checkbox in main generate screen. release is scheduled soon.