In this photorealistic XL-Pony-Illustrious hybrid model I tried to adopt multiple features, styles and concepts from other resources I used internally. It was trained on a big dataset of carefully selected SFW and NSFW images of different themes. Due to multistage training and additional tuning some PONY and ILL concepts might be washed out. Let me know if something is missing - I will try to bring them back.
v0.7a 1216px
Recommendations:
Don't put style / quality tags to the beginning of the prompt. If you put "beautiful face" at the start of the prompt you probably will get an extreme close-up of the "beautiful face" ;)
Text-to-Image (drafts) - use Euler A or DDPM, CFG 2.0 .. 3.0, 16..24 steps at 832x1216 px.
Image-to-Image (HiRes Fix) - DPM++ 2S a, Karras, CFG 1.0 .. 1.5, 12..16 steps, Denoise 0.45..0.67 at 2.5..3.5 MPx resolution.
More steps - more contrast / sharpness.
I2I at 4 Mpx resolution - set lower denoise because at high resolution elements duplication may occur.
Pony tags may work, but this version trained with more natural SDXL captions.
Tags like "anime style", "painting", "brush strokes", "digital art" may give a drawing style instead of realism. This version can render arts / anime if you ask so.
v0.5 1280px
Recommendations:
ComfyUI user can drop my samples into Comfy window. My complex workflow contains Qwen 3VL based best image selector - you can remove this option. Main feature of the Comfy workflow is the multi-stage rendering.
Text-to-Image stage usually gives the draft with good composition but bad details (so I recommend Euler A sampler for T2I). One or two Image-to-Image stages is crucial to polish the draft. It will give smooth background, more realistic textures and better detailing. Good results in I2I can be achieved in 2.0..2.5 MPx resolutions. You can also try 4 Mpx resolution.
For the CivitAI generator users I can recommend use the T2I with Euler A for drafts, then apply HiRes Fix with more contrast samplers (DMP++ *).
My apologies, I didn't manage to make it possible to get good results using the single stage render. All my attempts to train single T2I ckpt resulted in different critical flaws.
This version is trained with addition of DMD2 so it have reduced the number of steps required. I see some good drafts rendered with Euler A at 16 steps (or maybe 12) posted by other users, but I recommend 20.
v0.5 1280px
Recommendation for CivitAI Image Generator: start with Euler or Euler A (DPM++ samplers are too contrast for the first run); then use HiRes Fix mode and DPM++ 2M Karras sampler (if you get too contrast image, then try Euler A for HiRes Fix); see my samples made with CivitAI Gen.
This version is trained in high resolution and uses big number of steps. For speed, apply DMD2 4steps CFG 1.0 of DMD2 8steps CFG 1.5 LORA.
You can render from scratch at resolution of 1440px by longest side and then use ADetailer or HighRes Fix at resolutions of up to 4 MPx.
Use Euler A for Text-ti-Image render - this will produce smooth textures.
Use other contrast samplers with higher CFG to refine the textures.
Ultimate SD Upscale in ComfyUI with tile size of 1216 is a recommended way to produce the high quality images.
If you want to render images with only the Text-to-Image aproach then please give this version more steps. Also, you can post undercooked images if you need my assistance. I will try to help you with better render parameters.
FEEDBACK appreciated!
PS: eyes may be unstable for characters at the distance; I'm working on fixing it asap.
v0.3 params for CivitAI generator
https://civarchive.com/posts/27865267
I hope you will engoy using this. And I hope you will convert your pleasure into BUZZ donation ;) because I need more fuel to continue my work on training dataset improvements.
š„° THANKS for attention, posts and comments! š„°
Description
More training of CLIP (adherence improvement, use CLIP Skip 2) and Unet (anatomy, composition, details).
Added DMD2 with a very little weight to reduce the number of steps required. Text-to-Image with Euler A requires 16-20 steps for good drafts.
FAQ
Comments (11)
I like the output, but:
v0.6 & v0.4 Prompt adherence is very weak. The longer prompt the more inaccurate output it is.
Even other Pony CP do handle prompts better.
Is it only me?
I got an opposite report
Here is the example of much better adherence on v0.5
https://civitai.red/posts/28837926
And I used the same prompt on v0.6 with the equally good result in sense of adherence.
Post images so I can debug.
It is really possible that adherence on some of the Pony tags might be lower. This versio0n is heavily trained on natural language.
Let's debug together
@iwantobeleafĀ The prompt is simple, WF fallows Your guidelines AFAIKT,
realistic, photorealistic,milf,grey clothing, sports bra, redhead, sweatpants, bob haircut, green eyes, large breasts, pink lips, cleavage,chubby belly, pants pulled down,
And there is no way for gall to pull her pants down :) in conditions set in WF.
results: https://civitai.red/posts/28876545 (WF embedded)
Added more examples to link above using different checkpoints. I was unable to produce pull paniets effect whit this model without basically forcing it with > 2 tag strength and it did it incorrectly. Some other models did this, with 1.0 tag weight, 99% or 80% of the time, much more accurate prompt adharance.
It could be that this CP is overfit and / or training data didn't contained it and CP "forgot" this.
Also Your example already show this issue: https://civitai.red/images/132109930
score_9, score_8_up, score_7_up, source_photo, wide-angle lens, dark alley with distant light, 1girl, Young Adult Early Adulthood, athletic body, Mocha skin, Strawberry Blonde hair, Hoop Skirt, bridge, front view, rating_safe
- dark alley with distant light - it's not on image
@N0n4m3Ā Confirming that this particular concept (pants pulled down) is hard to achieve. Looking for better words combination...
I believe you can add some words describing what you must see when the characters pulls the pants. What will be visible after that. This did the trick for me (see the prompt below). I added natural description. Your concept along is weak and has low chance to be followed. But with additional words above I get 1 or 2 from 3 with pants down.
"1girl, (absurdress, bare hips, feminine_pubic_hairs, labia, bare inner thigh, no_underwear, no_panties, bare feminine pubic, hairy_pubic, cameltoe:1.2), sweatpants down, realistic, photorealistic, milf, grey clothing, sports bra, redhead, sweatpants, bob haircut, green eyes, large breasts, pink lips, cleavage, chubby belly.
Mature woman pulled her sweatpants very low so we can see her bare hips, bare thigh and her hairy pubic. Messy pubic hairs are visible above her sweatpants pulled down to the knees."
PS: need to work out this in upcoming training dataset; thanks for letting me know and feel free to post other weak concepts if you see some.
Here are the examples after the CivitAI's prompt enhancement:
@iwantobeleafĀ Thanks for the reply, but this prompt adherence is more deeper problem IMHO. The issue is that concepts are forgotten in Your mix and are replaced by other ones. It can be OK but makes this CP less flexible and moves towards natural language with I do suspect is what was used in training data. Again completely fine if that is intentional but it causes now collision meaning that it will be hard to describe output.
@N0n4m3Ā drift to natural.. yes, exactly