---V7 Notes ---
Prompt adjustments
Improvements to skin textures
More 2.5d/Realism flexibility
Bug fixes and general improvements.
**NOTE - I know that the early access price went up a bit, that's to accommodate for the new system here. I'm still dialing this in and trying very hard to keep this as reasonable as I can. It's a balancing act between this and how much buzz I spend training some things as well as featuring things.
---V6 Notes ---
Better Pony Prompting Adherence
Improved Textures
Fantasy refinements
General Improvements
---V5 Notes ---
Improved Textures
Fantasy Fine Tuning
Better Animals
General Improvements
*Note - This update does not dial back people a ton, so please continue to use specific prompts when you don't want people to appear.
---V4 Notes ---
Anthromorphic Update (variety and quality of people)
Downtuned animal generation
Lighting Improvements
Fantasy Elements Introduced
*Note - This update had a lot of images trained on people. It's going to try to put people in things unless you are very specific in removing them.
---V3 Notes ---
General Improvements
Improved skin textures
Improved animal generation
bug fixes (training to fix bugs) - or however we should say that with models.
--- V2 Notes ---
Improved textures
Improved lighting effects
Better control of light/dark scenes
Better fluid/water rendering
--- V1.021 Notes---
Technically this model is saved as 00Crow-and-Pony (double o) --> reasons...
This is a new Pony model for me, which is trying to leverage what I learned while I've been migrating my Arrogant Bastard model from 1.5 to Pony. This model makes no attempt to bring in any of my 1.5 work, and instead relies on new generations from Arrogant Bastard to build its DataSet. I did create a number of Loras for this, and I strongly consider this version to be a very solid 1.0. I'm pretty happy with it. I'll continue to update my other models as it makes sense, and I am in no way abandoning them. The truth is that this just does different stuff than my other models, even though this is a realistic model also. Please let me know what you think or if you run into any problems. I pay attention and try to incorporate your ideas, and I do go back to fix bugs in future versions.
Description
FAQ
Comments (11)
As always, a great job... hands are a wee fiddly and getting a fairy's wings to not sprout from the back of her head (vs her body) is arduous at best... but that's a limited use scenario (lol). Thank you for your continued efforts!
I’ll work on the wings.
Its good, and the first, which has booru adherence. Good job.
ty
very good model I am pleasantly surprised thank you to the designer I attach the photos taken for a start
Tested the model real quick. The image quality is excellent, but prompt adherence is sort of a problem. The model is Pony-based, but it doesn't speak the Pony tags language. Some tags it understands, but many, which are very easily interpreted by Pony (i.e. Skirt lift) it simply doesnt know. IHowever it knows very well some tags Pony doesn't know at all (tested "black woman" - works perfectly, which Pony can't do). Looks like you need to prompt more in SDXL style for this model, but it doesn't really understand very well (or I wasn't able to find a way to prompt correctly) the style and composition keywords that SDXL knows, so typical style extensions don't work.
Overall - looks like a great model qality-wise, really one of the best NSFW-capable photographic models I tested out, but requires learning it's own prompting style. And expect this model's understanding of prompts to differ from both Pony and SDXL, don't try to speak Pony to it.
P.S. - just forget it's a Pony model and do SFW fantasy portarits in it - they are AWESOME :D
It's always a work in progress and some updates push things one direction or another. I don't typically test NSFW content when I work on models - I know it's pony - there are some other things pony has been good for, for me. That said, I'll make sure that pony prompt adherence gets a better pass on the next version. Thanks for your kind words!
@QuadPipe Yeah, I actually like Pony more than SDXL myself for many SFW generations as well, but it is the tag language that was actually the advantage for me, where I could look up a tag on Danbooru and know if the model knows it or not and what the strength of the tag in the dataset is. Without that advantage, since I need to guess the model language anyway, the Pony base looses much of its charm. However your particular model looks real promising. I'll tinker more and maybe comment later.
Oh, one thing I noticed and would like to commend - the model inpaints much better than all other Pony-based models I tried. In this it is quite on par with the SDXL, which is a very nice bonus. Even to the point that it makes me wonder - why did you make it based on Pony, since it's now much closer to how SDXL behaves somehow? :)
@Thrordainul There are a few reasons. One since pony was trained on the core SDXL, you can get bleed through when you merge stuff in. There's always a bit of ebb and flow to it. It depends on what I'm trying to improve/fix.
Pony has some quality tagging that is really good and I was trying to see how far I could take it. So the Crow&Pony started as an experiment and it got good at a few things that my SDXL models aren't great at. I'm actually working on the update to this, that should launch sometime next week. I did do some inpaint work, but it was more experimental because I was trying to decide whether it was worth creating a separate inpaint model or baking some stuff in.
Grab a couple of my example prompts and mess with those, the negatives I'm using aren't random. I'm always working to reduce negative requirements - but you can't get away from some of it.
I'm also using DPM++ 2M SDE Heun with Karras on this a lot at 50 steps
Try 4x_foolhardy_Remacri for your upscaler at 0.27 denoising, 1.5 upscale, and 30 steps.
That's where I've found some interesting results. I do tend to use score prompts, but they aren't as necessary honestly. Next version will have much better prompt adherence.
If you checkout my quadpipe model, it's crazy good at prompt adherence and I'm using some tricks I learned form that one in the next update. Start with one of my images there too, and try the same sampler/scheduler and upscaler settings.
https://civitai.com/models/996342
@QuadPipe Thanks, I'll try out the SDXL model :) I'll probably also experiment with different schedulers, since this model is so unusual. I typically use DDPM with SGM_uniform, after rather extensive testing it gave me best quality at good speed per step as well as no difference in quality in Pony models between 30 and 60 steps. As for upscaler - yeah, it's a whole field that opens so much opportunity that I'm kind of scared to go there :)) I'm generatig locally and switching a model for upscale takes a lot of time, so I'm lazy and upscaling with the same model that I used for generation.
In the few test generations I made with Crow Pony score prompts did not affect the outcome at all.
@Thrordainul Upscaler can make a huge difference, especially at lower denoising - where you're really just enhancing. I've tried a ton of them and some are good at some things but bad at others. The Foolhardy has been a good daily-driver for me.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.