✅ Positive Prompt for Illustrious
Stable_Yogis_Illustrious_Positives
🚫 Negative Prompt for Illustrious
Stable_Yogis_Illustrious_Negatives-neg
==========================================================
✅ Positive Prompt for SDXL PONY
Stable_Yogis_PDXL_Positives
🚫 Negative Prompt for SDXL PONY
Stable_Yogis_PDXL_Negatives-neg
⚠️ Please use this model responsibly. While it is designed for high-quality and artistic outputs, AI generations can sometimes produce unpredictable results. I bear no responsibility for any unintended outputs—experiment and tweak settings to match your needs!
Description
Sampling Method: DPM++ SDE Karras / DPM++ 2M Karras / Euler a
Sampling Steps: 25+ Steps
Resolution: 896x1152 / 832x1216
CFG Scale: 4+
Clip Skip: 2
FAQ
Comments (14)
FYI distilled models (DMD) ignore the negative prompt.
Correction: if the model allows for increasing CFG beyond 1.0, negatives work fine, even on distilled models. That also appears to be the case for AimRim Pony XL.
Could yo clarify what you mean by this? Because in my experience they don't
@bm3011838 Without going into a deeper technical discussion, for which I don't have all the knowledge anyway, the way distilled models are generally used, at CFG 1, doesn't allow for the diffuser to take the negative prompt into account. It has no effect.
At least that's how all the distilled/DMD and Flux models I've seen are used, and if you try increasing CFG, bad things happen. It's definitely true for this one - AimRim Pony XL with DMD2.
If you've used successfully used any distilled models at higher CFG and/or with a negative prompt that has any effect at all, I'd really like to learn about them.
@bm3011838 OK, I'll have to correct myself, apparently you sometimes can get away with slightly increasing CFG even with distilled models. In those cases, a negative prompt does work. I'll correct my original comment, thanks for pointing it out.
HI guys Here is more clarification. https://civitai.com/posts/12938216
1: Image with Negative and CFG 1.1
2: Image with no negative and CFG 1.1
3: Image with no negative and CFG 1
All images have slight variation.
So Negative Prompt does make an impact.
Use CFG 1.1 or 1.01 to enable effects of Negatives.
dmd2 still not good enough. Its more realistic only with euler a, but thats glitchy. And you lose a lot detail.
@asadbekdevonta998 Just to ensure you get that correct, DMD2 models are meant be only used with Euler A.
Kindly always read the 👉 "About this version" section of the checkpoint. That section enlightens creativity !
@AIMRIM Actually no. The default sampler is LCM, because it has no artifact. And you can use higher cfg(1.5-1.6) and higher steps. But Euler is good too, but with low steps, cfg 1.2 max. And you lose a lot detail.
I write this for any DMD2 model, not only for you.
@lorenzitay956 DMD2 models are designed for LCM scheduling, but the repo does not clearly mention a specific sampler. I tested Euler A, and it gives sharper details—though it can be glitchy if CFG is too high.
LCM is smoother and works best with higher CFG (1.5-1.6), while Euler A is better for fine details at lower CFG (~1.2 max).
Interestingly, there are hundreds of high-quality DMD2 images using CFG 2.7, like these: Image 1 | Image 2. Check out my own comparison here: https://civitai.com/posts/12995584!
@AIMRIM I wrote the same:)
About high cfg pics. One of it is normal steps, normal cfg pic, other one is a lucky generation and ultrabright.
And for sex, 8 steps simple not enough.
I dropped Cyberrealistic for this. Since i use Pony, i always used Cyber.
What DMD lacks: emotions, fancy hairstyles, fancy eyes. I managed to set up a good setting with detailers and expressiveh lora(this should put way above its normal strenght).
I will refine it more, but now im ok with that.
Thanx for this.
Glad to hear you’re getting good results with AimRim Pony XL! Appreciate the feedback—I’ll keep in mind ways to improve emotions, hairstyles, and eyes on the DMD2 models.
Yes, its the nature of Dmd2. But its way more realistic and i better add detailers to it. You cant add more realism to a model, what its capable.
excellent checkpoint, I like the detail and how fast it is, we will continue testing.