Sorry, I should have worded that a bit differently. That came across a LOT harsher than I intended for it to. I edited my comment to remove the part about buzz farming because you are right that was unfair.

Its less about the version posting or any kind of hype chasing especially in this instance, considering pony is a pretty old category and doesn't draw up the hype that it used to. Its more that its a very common theme in the realism category that people are charging more and more for early access to their newest models just for them to not be that much of an improvement over the previous version, and that still have the same fundamental issues that they once had.

Your model isn't bad by any means, its just not up to the current quality standard of realism models, but then again none of the big name ones really are.

You at least have the decency to put out models that works, and you do a fair bit of training still which I am very happy about. Not all of the realism category creators do. In fact a few of the big name realism model creators at this point have gotten so lazy that their models now require significant refining or additional Loras to look any good.

Honestly at this point I am just looking for better resources and can't seem to find any. All the different SDXL branch categories are all suffering from the same stagnation at this point. There are so many different areas of improvement that people could attempt, but nobody (especially the big name model creators) seem to be trying anything new, or improving things that were already a known issue.

It seems SDXL is more detailed and photorealistic than Pony, is that correct ?

@DesignerSingapore SDXL and Pony can also be merged easily enough to combine elements from both, I had fun merging v14.1 and v15 of this line with Big Lust (https://tensor.art/models/956601564842165687/Lustful-Pony-v3.0)

</shameless plug>

@Cyberdelia Huh, works when I'm logged in, not in private browsing. Might need an account there. I did post my merges here first but they kept getting unpublished

Do you mean compared to base Pony? Base Pony is mostly 2D, anime/cartoon style. It can be decent at more realistic styles with LORAS, but kinda sucks if you want something approaching photorealism. Which is when you want more specialized checkpoints like cyber.

And it isn't in the description because it's kind of obvious when comparing the sample images of both models.

@oneeyed2 Thank you for the explanation :)

LOL, this is clearly a man who has never used pro noob.

@PixivTdog which model is that?

@bhookhabudha 

You can always dress the characters at the prompt. Specifying the outfit/clothes does a fantastic job.

It all depends on what you want to do.

What about for Text to Image generation using this model?

You need to understand ControlNets, try using OpenPose and SDXL Depth ControlNets, for the input image, it needs to be connected to a VAE encoder, which will then yield a LATENT that you can use instead of the Empty LATENT, it really is doable to build your own Workflows if you just start asking your questions to Google Gemini or ChatGPT

Same here. The last version that gave me consistently good results was v150 and the last one that gave me visually stunning or creative results was v130. But it's nice to have choices and I'm glad that Cyberdelia is still actively working on it.

@Cyberdelia Thanks!

Wrong VAE? This checkpoint has a VAE built in, but you might be overriding it with another? Try setting it back to "Automatic", & delete your entire negative prompt field just to check as well.

This sounds like what happens if you try to use an SD 1.5 VAE with SDXL/Pony though.

You probably have something else either in your positive or negative prompt that's overriding it, & cancelling it out. A LoRA(s) can also aggressively change prompt content & override content. If you're lazy/can't figure it out try adding this below (it will increase prompt weight; higher value means more - but don't set it too high or it'll overpower everything else in the image!

(anal:1.2)

or

(anal:1.4)

etc Go higher until it works

@ShadowCell some times this works, i leaned it from a famous anime series; (((anal over 9000)))

@pink0909 I'd strongly advise not going over (anal:2.0). Anything higher is liable to have more unwanted arseholes in it than the average subreddit forum post.

@ShadowCell i was joking , if you dont know dragon ball :D

https://huggingface.co/spaces/nicras/JoyCaptionBetaOne
if u r running ComfyUI you can run it locally else just ask grok to write a system prompt for a local VL model like Qwen3-VL (https://ollama.com/library/qwen3-vl) else Grok is pretty good at captioning nsfw images, absolutely no limit

Your image upload gallery is empty. Upload something with the specified problem, & I'll tell you where the prompt is going wrong.

Any time you have issues.....

Step 1. Remove all LoRAs & test just the checkpoint by itself.

Step 2. Delete the entire negative prompt field.

Step 3. Delete the entire positive prompt field, & start over to see if does what it's supposed to with a simple prompt just to test it.

Step 4. Stick this in your negative prompt field "cgi, anime, cartoon, illustration, painting, 3D, airbrushed".

Step 5. Positive prompt; try "score_9, score_8_up, score_7_up, Realistic, DLSR, photorealistic, RAW photo"

Step 6. Add LoRAs/prompts back in chunks (can do half at a time). If you add 50% of it back & it doesn't cause issues again; then those parts are probably fine. You can find the problem via process of elimination pretty easily this way. Only very rarely will it be more complicated.

Bonus images - I feel v16 was done kind of dirty from this batch (especially image 1), so I'm uploading some random seed images with otherwise the exact same prompt data.

They're looking way better so far! ^^

(Image Links to v16 will go here when ready)

These links won't work for now (still uploading), but I'm off for a sleep for now, so just check them later today in a few hours & they should be working by then.

https://civitai.com/posts/26529662

https://civitai.com/posts/26522978

https://civitai.com/posts/26529795

Just some basic stuff for now; I'll upload more later.

Tbh, this is absolutely baffling to hear.....lol. It almost seems like satire. Most people complain about precisely the opposite; that every human AI tends to generate looks like super models, & that getting them to produce just average looking people is the hard part lmao. Are you using a crap scheduler by any chance?

Even incredibly lazy & half baked prompts should still generate well above average atrractive individuals..... it seems like I'm stating the obvious too much when I say this....lol, but have you tried something even as simple as just "A beautiful woman, perfect figure, perfect skin...." ,etc? Literally even the most half-assed prompts should still generate attractive people by default.

It was me that suggested those other prompts in the comments you refer to; & they were for a user complaining about everything looking too cartoony (not so much unattractive). Refer back to the bit I talk about troubleshooting any issues.

Refer to the comment right next to this one (the one I made with image comparison galleries). Just copy & paste those prompts/settings into whatever you're using & see what you get? Hell, you can probably just download & save one of my images, drag it into whatever you're using, then just import all the prompt data that way.

I'm not sure if you just have a very weird sense of what constitutes as attractive, or what here.....?lol.

Upload some images where you have this issue. If all the prompt data is there, a ton of people in here will be able to diagnose the problem almost instantly (probably literally in 20secs or less just looking at the prompt data). Hell, it's practically like the Matrix for a lot of the chronic gooners in here; we don't even see the 'code' anymore....just "An attractive Blonde Woman in a red dress" automatically, when we look at the prompt data by itself at this point lul.

Here. Bung this in your positive prompt. Literally any seed will do. Either use Euler A Karras or DPM2. I did those comparisons with no LoRAs, & designed to be very simple. If it works, just use it as a template; swap "blonde" with whatever else you want. Swap "outdoor pool" with whatever also, etc. Probably best to not try & write entire prompts if you're new; just copy other peoples' templates until you get a good feel for it (you'll save yourself a lot of frustration starting out).

"Score_9, realistic, princess looking at viewer, dark blonde hair in a bun, perfect face, 20 years old, large breasts, sweating, hot, extremely detailed, intricate, elegant, deep contrasting background, beautiful light, inner glow, natural look, without make-up, black swimsuit, outdoor pool"

I use no neg prompts at all and all of the women are adorable.

I think 5.5 CFG is too high for a pony model, have you tried lowering that?

Is it possible this user has so little VRAM, that even modest tokens can start to push it way beyond its limits? lol (They might even be on a laptop/integrated GPU with no real VRAM haha).

What GPU/hardware you using? I wish people would upload images with prompt data; it's way easier to see what's going wrong.

PS - I once tried to help someone having 'weird problems' with Ray Tracing in Cyberpunk 2077.....turns out they were trying to use it running the game 720p on a 2050 potato or some shit lmao. I assume nothing at this point (you just never know).

this happens to me sometimes, and i fix it by grabbing a piece of prompt, and just shift it a bit. i'll just cut/paste a few words and change their order and often it starts genning fine again. it's as if some configurations of tokens it doesnt like, so i'll change it slightly until it fires

This can randomly happen with SDXL derived models. If you're using forge go into the settings and change emphasis mode to "no norm" and it'll stop. The only real downside is your old generations won't reproduce exactly unless you switch back. Apparently it's related to some underlying math issue when calculating emphasis.

you can see here: https://civitai.com/articles/4044 and https://github.com/hako-mikan/sd-webui-prevent-artifact for more info, or just change the setting and you'll be fine.