Pony FinalCut
Note this model is intended to be run at far higher resolutions then currently offered onsite. Consider petitioning @theally or @chipshajen or @Cosmic_Crafter (Or @Justin if your feeling lucky) to add support for such - Maybe they could "steal" from TensorArt for once.
Or maybe ask they didn't make high rez images look like garbage without clicking them three times. The perfectly sharp and single pixel hair at 3k has been thousands of hours of work.
832x1216 & 1280x1536 native generation size even 1280x1792 is possible for highly trained subjects
The best quality will come from forcing unet in BF16 & VAE/CLIP in FP32 use this tool to launch ComfyUI
Forge/Auto1111: set COMMANDLINE_ARGS= --unet-in-bf16 --vae-in-fp32 --clip-in-fp32 --cuda-malloc
Comfy: cmd /k python main.py --fp32-text-enc --fp32-vae --bf16-unet --use-flash-attention
This model took over 120KwH of energy to train. Not including the 50KwH used to train the CLIP models.
Note: Auto1111/Forge has an issue with stored BF16 VAE in V1.1 BF16 if not using the FP32 VAE command
Description
FAQ
Comments (20)
tile_resampling blurry photos with this model is awesome. The best I have seen so far. Bravo.
Thanks, Now with the JoyCLIP I have been able to push it to 2048x3096
@Felldude 1792x1280 pushes the limit of my RTX3060. My VRAM usage is hovering around 10.5gb and caps out at the end and takes about 1:10ish. I never had a issue before with Forge but I'm running that command line string now but I'm not sure if I notice a difference. I'm running a year plus old build of Forge though. Hybrid Joy looks awesome man!
@notjackdorsey Thanks - Yeah the difference from 1920x2304 to 2048x3096 is insane time cost usually I just use a 2x upscale of 1280x1536 then reduce by 0.75
You models are more than just a merge. For me, I think your pony's rock harder with realism and prompt following than any other horse in the race. Your whole pony collection deserve all of the gold buzz you get from people. That means your work is loved. Your pony lineup is oh so nearly perfect. You have done all the hard work. Your data is good and it shows in the results. The community all works hard to find their own magic sauce. You have made your own sauce and I love the taste. I just urge you to keep it on the burner, keep it simmering, keep on cooking, because there's still some spice to add, and unlock., but you are so close to having a perfect pony model.
Thank you - regarding keeping it simmering what I have determined is that I can only do so much without breaking the WD14 tags, characters etc that astra trained - right now if you drop LAION CLIP-G into the model the realisim is even higher and you get celebrity name recognition but loose most of PONY
@Felldude i use forge and i think all that talk about exchange and swap clip stuff wont work for forge. i just wanted to let you know how good your vin, proz, exotic, ultra, lites, supers, 2k, hybs, finals, and maybe a few i missed but they are all top notch. Funny enough the one i hated was actually later renamed by you with the title = do not use. Lately there have been Illustrious models posted with some nice realism models but they are not prompting like illustrious base should react to, which makes them useless. It no good having a realistic model if it wont listen to your prompts, you can not steer it to do something you want then its no good. I will note that your 2 illustrious releases for some reason are hard to prompt, they are a little funky with the faces, but no offense Illustrious is not meant to be for that. Maybe i do not like how they prompt, I do also love your ultratoon pony, you really have got the pony series rocking, i just wanted you to know i appreciate your pony work. i love flux and your flux models but but my system for me is to slow ( GTX1070 )which is why i like pony models.
Absolutely the most realistic pony model. Perfect with loras. But i come from Cyber, and sex prompts just simple not working good enough, very much hit and miss. It seems, models forget booru tags more and more, which is the easiest way for porn.
@rodricksammy327 I was aware of some of the issues with NSFW and was working on a fix - which worked extremely well. It is LMM guided CLIP-G and CLIP-L you can see the results in the top models here first https://civitai.com/articles/16652/fix-any-pony-model-new-llm-training
THANK YOU for uploading it as full precision.
Yep train away
For FP32 in A1111/FORGE:
set COMMANDLINE_ARGS=--no-half --precision full
is this the best approach?
Great model, btw.
I personally downcast the unet to BF16 : set COMMANDLINE_ARGS= --unet-in-bf16 --vae-in-fp32 --cuda-malloc --clip-in-fp32
When do you add the llm trained clips?
I might do a small finetune with the new clip but even that takes days to weeks. Visually I have tested across 8 models and shown improvement on 80+% seed to seed - But creators are not showing interest atm
@Felldude If nobody interested, at least release the clips. Thanx, looks promising.
@Felldude I already use most of your clips in my workflows, your stuff never fails to impress. Even if the vocal majority hasnt seemed interested, i believe its nosyly due to indtimidation and size of the model. I can run the fp32 model just fine on my rtx2080ti, and the results are amazing! I've dabbled a bit in training Lycoris,LoRa and DoRa, getting good results so far! Aai-toolkit seems to be the only out of box solution for training spesifically clip G or L that are embedded in the checkpoint. However, I was wondering if you have any tips on how to train with one of your clips, is the only option to merge them into the checkpoint im training?
@Bragonborn thank you - and I do have some of the clip models available for import on hugging face but merging them in and letting Kohya call the clip architecture and update the weights is the easiest
FP32 Model is GREAT! To be honest, this is the best realistic model I've ever used.
Thank you
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.