This is the SD 1.5 version of Pony V6 fine-tuned at native 1024px on ~5000 hand-selected images (some of them borrowed from my Zootvision model's datasets). Each image was captioned both with Florence-2 Large "More Detailed Mode" rich captions and also Booru tags from WD-VIT-V3. Use this the same way you'd use Pony V6 SD 1.5 normally, and just uh, enjoy the pretty objectively better overall aesthetics.
Important notes:
in A111, use Clip Skip 1 (NOT 2) and in Comfy just do not use the "Clip Set Last Layer" at all, with this
DO NOT use any VAE besides the one that is baked into the checkpoint, it will fry the image
You will probably get worse results from weird "in between" resolutions like 720x1280 than you will from standard ones like 768x1024 and 832x1216
Basic positive prompt: score_9, source_whatever, rating_whatever, your tags or natural language description here.
Basic negative prompt: score_3_up, score_4_up, score_5_up, sketch, (simple background:1.2).
Recommended steps / sampler CFG: typically Euler Ancestral at CFG 7.0 with around 25 - 35 steps is a good starting place. The DPM++ SDE GPU family of samplers can also be good with this at lower CFG (4.0 - 5.0) if you're going for realism in particular.
Generating at 512x512 is NOT recommended, as this model was originally trained by AstraliteHeart at 768px, and my additional training was entirely done at 1024px.
Description
Didn't plan to release a V3 really but noticed a few more things I wanted to train on and adjust. Can't push the realistic side of things much farther than this without beginning to degrade the model's original overall artstyle, though. The two Aerises in the gallery kinda show opposite ends of the current spectrum.
FAQ
Comments (8)
Did a V3! I think this is about as good as it gets for now. Might take another look at it sometime though.
New ZootVision out now too. Note that some stuff people might have wanted to use Pony V6 for, my model ZootVision is actually better at, e.g. ZootVision certainly knows a lot more obscure anime characters via Booru tags, and also knows like 20,000 'by artistname' format Booru artist tags.
Maaaaan, I don't have enough buzz to donate you more XD
This model is awesome - it's better than Pony 1.5 in almost everything, including horizontal aspect ratios! The only weird thing I see is smiles...exaggerated and uncanny sometimes. Good teeth, however. Very interesting, will test more!
Glad you like it!
In your review you mentioned artist tags, AFAIK Pony V6 1.5 never knew these any more than the XL version does. Also, try something along the lines of '(raw, photo, realistic, photorealistic, 3d, cgi:1.1)' in your negative if you want less realism.
All of that said, my checkpoint ZootVision is definitely better at what you seem to be going for overall IMO, and it does know thousands of artist tags. Did a quick example here just now:
@diffusionfanatic1173 Oh yes, I saw your recommendation of ZootVision in description, definitely will check it out!
As for artist tags, we've found the differences with a friend of mine who runs Pony XL (and merges now) daily. Pony v6 1.5 recognizes less artist tags, but the effect is more strong, especially for some tags (like kawakami_rokkaku which I had to use as a base anime style). In Better Pony it's offset by overall realism of the model, but I find results more interesting. Plus, Pony v6 1.5 is leaning towards pencil art style heavily (had to use `flat lineart` to fix it a bit). All in all, it seems to be just a difference in training and the base model (XL is a larger model after all).
For now I will use your models to see how far can SD1.5 go. Thanks again!
@dobomex761604 No problem! In theory it would probably be possible for me to train the same data from my Pony XL anime style Lora directly into the source_anime tag of Better Pony, to sort of make the 2D a bit more consistent. Maybe I'll do that in a future update.
@diffusionfanatic1173 I think "porting" (training?) styles loras from Pony XL to Pony 1.5 would be easier and more flexible (but you clearly know more about it).
So far I've only tried a few 1.5 loras on Pony 1.5, and results were...weak. Your model already works better in that.
@dobomex761604 yeah I would just straight up train Pony 1.5 on the same data as my Pony XL anime style lora basically, would work for sure
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.