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
Trained more on various things with the continued goal of improving aesthetics. Differences from v4 may be more or less subtle depending on what exactly your're prompting for.
FAQ
Comments (32)
V5 is better in following the prompt, but somehow makes anatomy broken: https://civitai.com/posts/4363419
The same happens with other negatives - I guess V5 needs different negatives than V4 and V3? I've uploaded images with the same parameters to V4 and V3 for comparison.
Link 404s?
a lot of the negatives you were using have never meant anything in any version of Pony, though, anyways. It DOES recognize all the actual legit Booru tags that start specifically with "bad", "missing", or "extra", however. Like "bad hands", "missing limb", "extra legs", and so on.
Stuff like the weird "fewer digits" and such I see all the time means nothing to it though.
Anyways my actual recommended prompting style is reflected in all the showcase images, as always.
@diffusionfanatic1173 how Civitai's sharing works? It's a public post. Anyway, let's try the image itself https://civitai.com/images/19535210
@diffusionfanatic1173 I've tried the two prompts without all that fluff - still, V5 outputs partially broken anatomy.
So far it looks like a trend going from V3 to V5: additional training helped with styles (especially V5 - it's actually good at retro artstyle , for example), but drifts the model further and further away from Pony 1.5, which affects compatibility with older negatives. I get that I'm supposed to use the new negative, but V3 is more convenient in that I can (not always, but often enough) use prompts/negatives from examples that use Pony XL and derivatives - they are massively more popular.
@dobomex761604 The prompt you linked is frankly ridiculous even if it works sometimes, whoever wrote it originally doesn't know what they're doing, even the positive doesn't make any sense, like "source_furry" and "source_questionable" are fighting against them trying to depict a human-on-human explicit sex scene, and so on. It's not possible for me to retain perfect compatibilty between every version but I'm sure there's a way to generate this image on V5, I'll try it.
I can't reproduce any of your images at all in ComfyUI, they're similar but never the same, A1111 noise scheduling must be different a bit
@diffusionfanatic1173 I'm using Krita with the plugin, and that plugin installs ComfyUI. So it's essentially ComfyUI without nodes, but with a canvas sheet and "ease of use". Plus, I have to use "GPU Low" mode - I only have 3GB of VRAM.
I've also tested your Aerith prompt with all three versions because I was hesitant of score_3_up, score_4_up, score_5_up vs score_4, score_5, score_6 which I've mentioned before - seems like it's all over the place, both options work so far. It is very clear, however, that V5 is the most "anime" of all these, which is very interesting to test further.
The huge prompt is definitely ridiculous, but that's where SDXL and 12 GB VRAM shine. That prompt still breaks V3 sometimes.
@dobomex761604 Yeah V5's dataset does have a lot of art content, though also some that was aiming to improve realism when specifically prompting with "raw, photo, realistic". I might release a V5.5 later today, not big enough to call it V6 but slightly better than V5 I think.
@diffusionfanatic1173 I find it interesting: do I understand correctly that art (anime/comix) can reduce the quality of anatomy in generated images? Is there a way to map/filter an anime-focused dataset with "anatomy first" in mind?
I remember back in the day, when Anything v3 just "released", I had to spend hours just to get the character I wanted the way I wanted, but I didn't really focus on anatomy. Nowadays, with loras and how good models are (and how well-optimised lowVRAM modes are) it becomes more interesting to just get good pictures. That's a big part of why Pony became popular - universal, less hassle with prompts to get a good anatomy.
I'm not sure if it's possible, but having a new booru-focused model (like Pony, but strictly booru tags and no scores) would be massive. That's a lot of time and a lot of compute, tho.
@dobomex761604 I have the opposite experience adding flat shading artists to prompt increased performance for me, especially concidering that anime loves thick lines (thick lines are good at hiding blurriness and lack of detail in image). It feels like Pony 6 SD1.5 hates mixing artstyles that are too different from each other (realism+anime=bad). Currently at V5 realistic women look like orcs tbh.
(Simple style comparison)
https://civitai.com/images/19839883
https://civitai.com/images/19839952
https://civitai.com/images/19840038
@dobomex761604 the 2D part of ZootVision is already exactly this, it's a much more traditional anime base with typical "masterpiece, best quality, high quality, normal quality, low quality, worst quality". I dunno how many images I need to post in the Zootvision gallery before people understand that the whole point of it is that you can control the look by putting different things in the positive or negative.
@Cum_Miser Pony (the original or this version) DOESN'T know artists in any intentional meaningful way, ZootVision does. In fact I'm going to go post a text file with every artist tag ZootVision knows with more than 100 appearances in the original dataset, on the ZootVision page.
@dobomex761604 https://civitai.com/images/19772975 I replied here, this person does have a tag the model knows about but you had the name backwards, you'll find it in the tag list for artists I uploaded on the Zootvision page.
@diffusionfanatic1173 with all due respect, ZootVision suffers from the same problems that are lingering through probably all Anything v3/NAI-derivatives: getting characters that don't have enough presence is very, very painful without LoRAs (which is much easier on Pony). Plus, the unfortunate reality of inbred anime mixes has brought us to such a strong generic "default style" that it reduces the influence of artist tags.
I've tested ZootVision with the same tags I used on Pony ( henriiku_(ahemaru) , yoshio_(55level) , hara (harayutaka), yuiga naoha), and ZootVision is visibly less strong with them - again, the default style is too influential. That's why making a new anime model is important.
As for "Pony (the original or this version) DOESN'T know artists" - sure, it doesn't have all tags...because they were scraped out as tags. For example, Satoshi Urushihara wouldn't work on Pony, but works on ZootVision. However, it is caused by scraping out tags, and the ones left are simply ones they've missed while scraping out. I'm sure the artworks were left in dataset.
In any case, these are your models, and I deeply appreciate creating Better Pony Diffusion - it's an absolute blast. However, in practice I see that focusing on artist styles doesn't lead anywhere. I know that a lot of people don't like Pony's syntax (scores and such) - but it's hard argue with effectiveness. Pony and derivatives are simply more flexible and stable, which is especially important for SD 1.5.
@diffusionfanatic1173 oh my, why sd1.5 tokenizer has to be like that( "urushihara satoshi" but not "satoshi urushihara"?
@dobomex761604 you need to use the exact syntax in the CSV I uploaded for Zootvision styles, they always start with the word "by". Less than 25% of the original basis of Zootvision is even vaguely NAI-derived also, the customized model I started training Alpha on initially has much more in common with Furryrock believe it or not.
Underscores are never correct (except for Pony scores that were actually trained with them) also and always make sure you properly escape brackets, that's how the captions work, otherwise you're putting undue emphasis on parts of the name you don't mean to.
Do you have examples of characters you found to be a lot easier on Pony?
One other thing I should mention, "masterpiece" in the positive is highly likely to overpower all of the artist styles in ZootVision, it's better (if you're specifically using styles) to just do the negatives like usual, but not "masterpiece" and whatnot in the positive.
@diffusionfanatic1173 Motoko Kusanagi, as always XD Due to multiple redesigns it's hard to get one exact, and the amalgamation of designs looks horrible. Plus, all SD models know her (even from the base SD 1.0), so the base "concept" also influences the result. Pony is significantly more stable in that. Misato Katsuragi is in the same spot.
Essentially, if a character is not too famous, but still iconic, you'll have trouble getting them without LoRAs. The more popular a character is in the present (Yae Miko, for example), the easier it is to get them.
@diffusionfanatic1173 Yes, I remember about the influence of masterpiece . I remember in some models (like ERA, for example) it was suggested to not use it at all. In ZootVision, however, it improves the overall quality too much to ignore it.
@diffusionfanatic1173 I've just tried to generate futas on zoot6 zeta, but dicks were all wrong. I kinda gave up. The list of artists is impressive though.
@Cum_Miser I've never tried that lol, it can certainly do men, what were you getting exactly?
@dobomex761604 you can definitely forgo masterpiece if you use the right negatives, and artist styles that aren't super weak. Several images I posted yesterday were examples of this.
Anyways I have a V6 cooking of this with one dataset that's just combined 2d / 3d NSFW of all kinds, and another that's just the same dataset from ZootPhotoMaxxer XL. About 1500 more images total, then I'm probably not gonna touch this thing again, the CLIP was too fried to begin with for me to be able to improve the prompt adherence and such any more than I have I think.
Only other thing I might do is see if I can figure out what he did to the VAE to make the encoding wrong relative to all other SD 1.5 models, so that I can swap out the shitty WD one for Furception, which is basically the best for fine details regardless of content type.
wait how is ZootVision Motoko Kusanagi worse lol? I just compared "masterpiece, Motoko Kusanagi" on ZV vs "score_9, Motoko Kusanagi" on BP, no negative for either, same seed, the ZV one was a straightforward pretty accurate 2d representation while the BP one was a bizarre looking pseudo 3d randomly muscular shirtless depiction lmao
@diffusionfanatic1173 If by "a straightforward pretty accurate 2d representation" you mean Netflix series - that's a blasphemy. Also, I'm getting random and inconsistent mixtures of that with long hair (???) most of the time on ZootVision (and any other model - that's why I use that LoRA).
Better Pony is at least consistent with eye color and the overall shape of her hair (far from perfect, but at least close).
muscular shirtless yes please.
@diffusionfanatic1173 just checked, Better Pony (and Pony in general) sucks without negative, but it wasn't supposed to be run without negatives lol. But the overall look is pretty consistent and is exactly what it should be.
Afaik, in datasets designs win by volume: if a certain character has most images with a particular design, that design will be (should be?) more prevalent. However, the majority (pun intended) of artworks on -booru with Motoko are the series design, like https://img3.gelbooru.com//images/4c/71/4c716d0e96c45e58eddaf69b71e47d9d.jpg
However, in Anything v3 and NAI (and all later anime models based on them) this design is not prevalent - instead, it looks like Kuvshinov's design or something else. And that's why I've mentioned the base SD 1.0 which already "knew" Motoko - that design is still affecting all non-Pony models.
It seems like Pony did the best job at going away from the roots of SD 1.5 - maybe that's a part of why it's so flexible.
@diffusionfanatic1173 Sorry, I don't understand what you mean.
@Cum_Miser I just meant like what were you getting for futa prompts exactly lol. It knows the tags but I've honestly never tested them that much in particular
@diffusionfanatic1173 Balls, but no dick, or worse. https://civitai.com/images/20541961 (I will delete post soon)
I've slightly adjusted my "basic negative" recommendation in the main description. As far as I can tell, Pony 1.5 must have had a very large number of images in the original dataset that inaccurately had "greyscale" and "monochrome" assigned to them, resulting in the generated image being much more saturated colorwise than it should be if you use those at all in your negative (because presumably you're then negating all but the most vibrant images basically, not just true single color ones as it should work).
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.