WARNING: v2.0α is a 1-epoch preview. FLF-trained. Expect drift.

Trained on high quality 768p, native-res video and image stills.

Triggers

Always use: 2dan1m and 2D-animated.

Best mode: I2V or first + last frame (FLF / FL2VA).

Strength: 0.8-1.0. Guidance: 1. FPS: 24.

Start every prompt with:

2dan1m [Shot 1] 2D-animated,

Then count + act only (1girl, 1boy, cowgirl position, vaginal / fellatio / handjob / paizuri / ejaculation), name and describe the characters, and write the action. Camera can hold, pan, tilt, or push. For a real cut, add [Shot 2].

Use H3’s three fields: integrated_multimodal_description / overall_soundscape / non_diegetic_music. Silent clip: overall_soundscape: N/A.

Cel craft: hard cel shading, clean ink outlines, on twos, smear frames, background hold. Pixel stills: Pixel-art sprite animation, chunky pixels, limited palette, on twos, background hold.

Not for: photoreal, 3D, aftersex.

Download TEMPLATES.txt for paste-ready loops.

Trained terms (use these, not synonyms):

penis, vagina, anus, breasts, semen, vaginal, anal, cowgirl position, doggy style, missionary, standing sex, fellatio, handjob, paizuri, ejaculation, facial, cunnilingus, female masturbation, on twos, smear frames

Training

v2.0α

First-epoch FLF preview. Clips + stills. Trained using kohya-ss musubi-tuner fork on an RTX 6000 BW due to high VRAM requirements for high res video. Training is duration-gated: Native single frame stills, 768-class ≤5s, 512-class ≤10s, 384-class ≤15s. Audio in the train. Rank 16. ~500 samples.

v1.0 - 5sec

Short looping 5s at 768-class. Use this for high-res loops. Trained using ai-toolkit on 5090 with 512p-class ≤5s video. Audio in the train. Rank 16. ~330 samples.

v1.0 - general

Longer clips, multi-shot, up to 15s at 384-class. Trained using ai-toolkit on 5090 with 384p-class ≤ 15s video. Audio in the train. Rank 16. ~330 samples.