If you do like this work, consider buying me a coffee :)
Use v6 of this model for free on Stable Horde
Use this model for free on Shakker.ai
The long awaited followup to ICBINP, this model is merged to try and get the most realistic images possible. It's the ICBINP you all know and love, but now with all of the benefits of SDXL (recommended 832x1216)
Update: v6
More mixing to try and get something even better than v2, and I think I've found a much nicer balance
The sample images are made in Forge with the DPM++ 2M Karras, 35 steps, 832x1216, cfg 3 - 4, clip skip 2, (some hi-res fix, but the best realism is without it IMO - use the Ultimate SD Upscaler instead) and no additional tooling so should be reasonably easy to replicate.
If you want to push this to make anime, change the sampler to DPM++ SDE 3M Exponential and crank the CFG up to 6-7.
I'm sure that some more experimentation and better workflows you will be able to achieve amazing results!
Description
Merged v2 with Fully Real XL, my own trained lora from a larger dataset than the first one, and the DPO lora, then merged back into v2, then more DPO lora and the Add Details lora to finish
More goodness for all :)
FAQ
Comments (22)
--no-half-vae doesn't work with inpaint(, Please, make a BakedVae or VaeFix version!
Are you using Automatic or None as the vae setting in a1111?
@jafdeth2030105 use the correct fixed VAE (https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main) and then you can ditch that startup argument altogether. Also be aware, you need to manually set VAE in automatic1111 - in my experience, the "Automatic" VAE setting does not work. Or does not work reliably.
EDIT: you'll also get a speed-up from ditching that argument.
if you download that VAE, i recommend you change it's filename to something like "sdxl-vae-fp16-fix.safetensors" because it shares a filename with the unfixed version which can potentially create confusion.
Have found the issue.. will make a v3.1 that has the VAE fix baked in
@residentchiefnz do you have an eta for 3.1 please?
@MagicaMentis apologies.. I made the 3.1, and proceeded to leave it on huggingface - I wasn't entirely happy with it as it was having an odd issue where the eyes would occasionally split apart vertically (i.e. 1 eye higher up the face than the other), so I never published it - if you want to have a play with it, it is available at https://huggingface.co/residentchiefnz/Testing/resolve/main/v31_gradAclip.fp16.safetensors?download=true
@residentchiefnz Ok, thanks!
@residentchiefnz Is this fixed in latest version? Is the VAE Baked in now? v4
@MagicaMentis I think I baked it in but I built half of it a couple of weeks ago and picked it up again so I dont 100% remember sorry
@residentchiefnz Ok, will give it a try anyway! Cheers.
@shapeshifter83 What's the argument to ditch it altogether? Just --no-vae?
@lucidzachary473 um, i'd have to go look at the A1111 page to confirm what "--no-vae" switch does, you might wanna do that. I can only confirm you can ditch "--no-half-vae" (and gain some VRAM and speed-up) if you're using the fixed VAE.
@lucidzachary473 Set the VAE to None in the settings
@shapeshifter83 thanks a lot, bro! it helps!
@residentchiefnz and much thanks for v3.1!
V2 is significantly better than V3 in terms of look and feel. V3 has that uncanny AI look that wasn't there in V2.
Thanks for the feedback
For people photography v2 is my favorite.
I tried a lot of checkpoints, but this one is the best for photography generation i ever used.
Congratulations and thanks for your amazing work.
Glad you like it :)
wow... I just generated ONE man's portrait and I am stunned. really great, for my taste it looks more natural than Juggernaut. Thanks!
Glad you like it, here's hoping it wasn't just a one-prompt wonder!!