Hello ♥
for whatever reason you want to show me appreciation, you can: ❤️Ko-Fi❤️
Special Thanks to
badassdragoon2 georgebanjog Mr_fries1111 barytdruck124
Please take a look at their amazing work!
Update v1.0:
if you see when V0.5 was released, you can imagine how hard it was for me to create a better version that is really better in all areas.
This is the result of probably 100 different versions.
more Details
better colors at low cfg
better faces (and more variance) (not perfect)
For people and faces, you should not use DDPM. Use DPM++ 3M SDE or Euler
Update v0.5:
Small improvements
- better skin, teeth and eyes in humans
- more details
I am still working on the following improvements
- deformed faces / displaced eyes
- faces from a distance
- xxx
Update v0.3:
It's just a small update to fix some issues from v0.1.
It doesn't change every subject, but some improve.
I haven't had a picture in about 1000 pictures where v0.3 wasn't better, so the upgrade is definitely worth it.
I can notice the following improvements:
a bit sharper
more realistic
a bit more creative
a bit more possibilities to get nicer hands/faces on full body photos
________________________________________________________________________________
This is a merge of the following Models:
Boltning
JuggernautXL
WILDCARD Lightning
Recommended Settings:
Sampling Method: DDPM (my absolute favorite), Euler, DPM ++ 3M SDE
Sampling Steps: 10 (more very rarely make a difference.)
CFG Scale: 1-4 (lower, more realistic, i prefer 1-1.8)
Upscaler: SwinIR _4x (this is my all time favorite)
(Hires steps: 15; Denoise: 0.28, upscale by 1.5 - 4 (depends on your GPU))
📌 USE sdxl_vae
All Images generated in SD Forge with Nvidia RTX 4090
Please test other Settings and share with the communty, so i will learn about the Settings and can optimize.
Recommended Loras:
Please experiment with the weights. Loras are really gamechanging in this Model!
TESTS
Here you can find and download my tests:
Thank you so much for using this Model. Hope you are get insane results and create some BANGERZ ♥
Description
converted to f16
FAQ
Comments (11)
so how many steps and cfg used for this?
please check Model Description.
cfg: 1-2
steps: 10 (+)
sampler DDPM or Euler
After testing the Checkpoint for more than 24 hours now on fooocus and invoke and talking directly to the creator on reddit yesterday I can safely say that this is hands down one of the absolute best models I have ever used.
I personally recommend using Invoke for this model, it seems to give out better results than fooocus (I don't use A1111 or comfy yet so I can't comment on those two)
Initially I didn't know this was a lightning model and I assumed it was a regular one and I already found it very impressive. Once I started using it as a lightning model tho I went from not caring about lightning models to considering this one of the best models of all time bar none, lightning or not.
I personally use CFG 1.8 with DDPM(no scheduler just the sampler, in foocus that would be Sampler DDPM and Normal scheduler) as recommended at 15 steps, I know the creator recommends 10 but I feel 15 is even better, no need to go higher.
So what have I deducted so far:
- Extremely detailed and anatomically correct characters
- The model can handle multiple characters of varying sizes and distances incredibly well
- It handles distant and smaller faces very well
- Environment detail is exceptionally high and it's a damn good model for inpainting and outpainting
So far I've been able to achieve results every bit as good as my favorite models if not better, especially on Invoke, maybe invoke has a better VAE by default I don't know but the clarity, sharpness and precision are simply outstanding.
At 15 steps on my 2070 Super it takes a mere 9 seconds to generate fully fledged, realistic, high quality photos at 1024x1024 and similar sizes.
I've tried using lightning and turbo models in the past and they'd always choke on multiple characters, hands and feet, smaller details, environmental details, not this one, this makes mistakes like merged or extra limbs once in a while but nothing major and nothing that can't be fixed with a quick inpaint.
Thank you so much for the outstanding model my friend, I'm glad I saw your post yesterday or reddit :)
I don't know what to say.
Thank you very much for this detailed report. i wish i could pin it so that more people could benefit from your experience.
I am incredibly pleased that you are so satisfied. i wish i could see some pictures of you.
I hope you still get infinitely good artwork!
Could you put this available to use on here to test it?
what do you mean?
@AIDigitalMediaAgency there is usually a pencil option to try here on civit ai the model, I think you must enable it so we can test here your model before download it you know,if you see other checkpoints you will see the pencil option to try it
@kd231002573 just the big checkpoints are aviable on civit to generate here
@AIDigitalMediaAgency oh I didn't know that, anyway I will wait for a fp16 version of your model with less weight to try it out when is ready 🤙🏻
@kd231002573 there is the fp16 verson of version 0.3
@AIDigitalMediaAgency lol Im so blind, thanks,will test right now