UPDATE:
Some of you seems having trouble to run checkpoint. According to feedbacks that i am receiving, outputs are getting odd. I've re-uploaded the checkpoint just in-case.
I've tried the checkpoint with 1.5.x, 1.6.9 and 1.9.x a1111 api's. Everything worked fine with multiple schedulers.
If you are still getting errors, remove everything on the "override settings" if you are using A1111. Thgen if you need to please make sure you have updated drivers, updated and correct cuda (cudnn) versions. Also check gradio versions as well. On my end, everything works without a problem.
ıf you do all actions that i've listed and still getting error, please provide a bad sample image, so i can directly read the metadata of your bad sample and check what is wrong. :)
I started to make this model as improved of my freel_photo. Eventually it got serious and changed into something good.
(Attention: let's say i sent 100 sample of photos, about 10 of them used adetailer. )
You can achieve really photorealistic images. About hand issue somewhat solved considered early trial that is freel_photo and it produce better hands but you still need lora, adetailer for perfection.
Added more head variations. It is not only producing great variety of head types, i finally balanced highly biased asian faces weights to caucasian faces. It still produce those faces but not biased as some of checkpoints out there. So you might not need to prompt to exclude them all the time.
at this point i'm satisfied about light, skin tones, skin textures. I might do little bit more tweak in the future on those weights.
Trained and merged with my test checkpoints, freel_photo. Trained some custom weight layers from scratch to experience those to understand better that how it works. :)
I have a plan to add more
face features,
young - old correlation.
better male subjects, anatomy ( yet it is improved on this point to get, better nsfw poses, somewhat better penetration textures and somewhat better penis)
thinking to add more landscapes, more interiors to get lifelike varieties for portraits.
I humbly ask you to give me a feedback, ideas about portrait composition, background subjects.
If you like the checkpoint or don't like the checkpoint, comment about it, that would be nice to have. So i can improve it.
Description
+ much better lighting, textures.
+ somewhat better compositions
+ better nsfw about body structures but not perfect.
FAQ
Comments (21)
having issue with version 14 with image becoming terrible quality at last second of generation. It happens with and without Clip Skip to 2
Hello, i will try to help you if i can. Most likely you have sdxl vae instead of usual sd 1.5 vae or doing overriding of something that you might missed.
Can you share a sample image that you're referring to ? so i can copy what settings you have when you get the error. I might manage to see what is wrong. If it is related to collapsing weight pipeline, i might able to change the specific weight line and release a fix. Thank you for feedback!
Same problem here, I tried with the same prompts and settings as the example images (without VAE), and results are terrible
same problem
@Error666 Thank you for providing settings. Can you cancel the downcast alpha settings below the image that you've provided ? If it won't work, i will re-upload the checkpoint because on your settings it shouldn't write new file < nouveau dossier > . That might implicate, checkpoint is corrupted when i uploaded to the civitai.
i'm re-uploading the file just in case. When it is available, please download and try it one more time. I'm currently testing it and it hasn't give any error for me so far. If error still continues, please, update your driver, your api-gui whatever you're using such comfy or a1111 etc.
before trying anything time-consuming, can you try choosing any or each schedule besides automatic ?
@Reelai It is the same without the the downcast alpha and clip skip. "Nouveau Dossier" is just "new folder" in french, it's just a subfolder where I put the models that I want to test, there are no problems with that.
I think there is a problem with the model, I have tested hundreds of them and I have never seen this. Hope you can fix this, I love your FreelPhoto !
Sorry for my bad english
-Strangely it's work normally on Forge-
Edit: Oops sorry, bad handling... same problem on Forge...
I will test on Comfyui later
@Error666 Bonjour! Je suis désolé pour ce problème. J'ai essayé 3 versions différentes de l'api. Tout fonctionne. Je suis en train d'uploader un nouveau checkpoint, cela pourrait résoudre le problème mais je n'en suis pas sûr. :)
@Error666 FYI: I didn't have "downcast alpha", just the Vae (84000) and Clip Skip 2, I " x " them out, used my Quicklist (vae and clip skip are setup) change as needed, ran A1111 to generate, worked, no pinkies pics.
@mrj622592 what exact steps did you take?
hi Reelai, I am still having issues with this model being pink at the last second even after using the newest model and upgrading my card
Hi, Can you remove everything on the "override settings" if you are using A1111 ? I'm almost sure that override settings are causing the pink-ish output at the end.
@Reelai what do you mean by that?
@Reelai I dont have any override settings and still have the pink issue.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.