This model is build for all anime lovers.
I continued training this model on a 1M NSFW anime dataset. Training is still ongoing, but it's already functional—enjoy!
Training is far from finished, with only about 25% completed so far.
Important
Prompts are absolutely crucial. If your prompt is too simple, the results will look just like basic anime model outputs. But if you include highly detailed scene descriptions, the outcome will be on a completely different level.
Prompting Tip: A blend of rich natural language (environment, action, character, style) paired with strategic prompt tags yields the best results; overly minimal prompts often lead to overly simplistic images.
An anime-style digital illustration in a dark fantasy gothic aesthetic with a dramatic low-angle perspective that emphasizes depth and scale, featuring a demon girl with long flowing white hair and large curved black horns seated on an elevated stone ledge; she has striking heterochromia—one glowing red eye and one cyan blue eye—along with pointed elf ears and red facial markings, looking down at the viewer with a mischievous smile; she wears an intricate black lace outfit with sheer lace stocking tights, dark red gloves, and shiny black Mary Jane heels with blue jewel details; in her gloved hand she holds an ornate gothic staff topped with a glowing red orb crackling with violet lightning trails; the backdrop showcases a massive, detailed full moon illuminating the dark night sky, framed by gothic castle spires, arching stone architecture, floating pink petals, and a dramatic foreshortened composition that enhances the immersive visual impact.Recommended Settings:
Resolution: 1024–2048px
Steps: 8-20
CFG Scale: 1
Sampler: Euler, Euler A / DPM++ SDE (or
er_sde)Scheduler: Simple, Normal
Supported Quantizations / Precision:
BF16, FP8, INT8, and INT4 formats are all supported!
V1.0
The first stable release. It performs exceptionally well across various art styles and already shows the potential to rival models like Pony in NSFW content. It works great with natural language prompting and understands most mainstream characters, though artist tags are not supported. Training progress is at roughly 5%.
V1.5
A recent checkpoint with impressive performance. It excels in multi-character scenes and character details while mastering a wider variety of art styles. I've made it my go-to anime model—it outperforms models like Pony and IL. It understands most mainstream characters, but artist tags are barely supported. Training progress is at roughly 10%.
V2.0
A major leap forward. V2.0 delivers noticeably sharper NSFW comprehension and rock-solid prompt adherence, alongside significantly upgraded multi-character handling and amped-up anime aesthetics.
Heads up on a few trade-offs:
However, it does come with a drawback: character forgetting may occasionally occur. Using 4-step LoRAs is not recommended, as the current model weights have diverged significantly from the base weights, which may lead to a degradation in image quality.
(I hope everyone can submit more generated images or provide your valuable feedback within the corresponding version. Since I'm a solo creator handling everything on my own, my testing scenarios are limited—every bit of your feedback truly helps improve the model.)
Description
FAQ
Comments (8)
Good job. Continue your work and one day your model will get it's own base model page.
how shall I describe the style and the IP characters?
Character Name / Franchise Series / Hair Accessories (e.g., hair clips, hairband) / Clothing Style (e.g., uniform, dress)
I do have some issues to report: while version 2.0 runs faster on my platform, its image fidelity and LoRA support are not as good as those of version 1.5. The same applies to prompt adherence; for instance, version 1.5 could successfully render details like "middle-parted bangs" and "pantyhose," whereas version 2.0 fails to do so.
The v2.0 weights have drifted pretty far from the base model. It's way more clued-up on NSFW stuff now, but it's gotten worse in a few areas—it's a trade-off, really.
Yessss!!! v2 is HERE!!
This is for v2. int4 seems more stable generation, anatomy and saturation than fp8, fp8 will often appear some weird block ,anatomy problem and sometimes slower. Euler a/eular is better than er_sde/dpm2pp. Pose stable than v1.5 but lose some detail and background somtimes turn white or black. This is my exp, thanks.
I am a foreigner, and I wrote this review using a translator. The training results for your base model this time are far better than before, and the quality has improved significantly. Thank you so much for sharing—please keep up the great work! Love you!