NOTE: This is an old model, my latest work is Ultra Maiden.
...
Neo Maiden 4.0 is one of the best realistic SD models, and one of the sexiest! You'd never guess from the previews, but that's just because I selected as unsexy example images as I could find. Gotta vibe with civitai's new groove ya know...
As is my tradition, the example images were made with random single-token prompts, but this time I also used a negative embedding I created especially for this model. The model was tuned to not need negatives at all, but this embedding reduces nudity and makes the output more cinematic, and it also makes default faces look more European.
Other than that I just used my standard basic generation settings: 12 steps DPM++ 2M at CFG 7, with the Align Your Steps schedule, generated at 576x832 px.
You can of course do whatever you like, but I recommend trying my settings first, just to see what the baseline performance is. If your fancy settings don't beat it, don't bother!
Description
FAQ
Comments (2)
Takes some finagling.
512x512-512x768 seem to work okay, 768x768 in specific situations, anything above that goes south fast. Very good lighting, good posing, good faces. My takeaway is that it is difficult to get the CFG and steps down and are very finicky, but I'm happy that it looks different from every other model/merge out there.
My personal settings for decent compliance with prompt:
Steps: 22-27 (24)
Sampler: DPM++ 2M SDE
Schedule type: Exponential
CFG scale: 3-4.5 (3.5)
Size: 768x768
Lora:<lora:Hyper-SD15-8steps-CFG-lora:0.5> is a must, otherwise you pretty much double your steps
Interesting. I'll admit prompt compliance isn't something I've been caring a lot about. My own prompting strategy is to just type something random and say "surprise me!", then generate a few seeds until something catches my eye (and isn't too broken). My tuning has been focused on getting the model to produce nice pictures from very simplistic prompts, with a lot of randomness to poses and backgrounds, without having to ask for specific details. So trying to control the output explicitly is probably harder than with most models.
I agree that the model is limited in how many pixels it can handle without issues. My example images are 576 wide and 832 high, and that's already pushing it. This is one thing I would like to improve in the future, if possible. My upcoming work-in-progress version seems to be getting slightly better at handling larger pixel formats. Slightly.
(update: v3 did indeed turn out slightly better at handling more pixels.)
My personal settings for fast and reasonably high quality results, with good success rate:
Steps: 12
Sampler: DPM++ 2M
Schedule type: Align Your Steps
CFG scale: 7
Size: 512x768 or 576x768 or 576x832
No LoRAs. Very short and simple prompt and no negative prompt.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.