Loading full resolution…
![Training frame full resolution]()
Frames extracted from 11 Osmo Pocket clips only — isolated from the v6 training exhibit and other projects. Face-excluded · creator-owned · model-card documentation.
Silent web encodes of the 11 Osmo Pocket clips that produced this 66-frame train set.
No source video available in this build of the catalog.
Click any frame for full resolution and the training caption used at fine-tune time.
How to read this page as training documentation.
Adapter training for anatomical consistency under viewpoint and contact deformation. Multi-angle identity stills + junction / occlusion examples so the model is not locked to a single pose.
Captions use trainer tokens
(prvpns, optional prvpene for contact scenes).
Restricted technical documentation for the published LoRA card. Do not redistribute frames. AI showcase images on model pages are generated samples; this catalog is training material only.
| Product | Osmo v8 TRAIN set only — not a train export |
|---|---|
| Base | Krea2 · Qwen-Image / flowmatch |
| Network | LoRA rank 16 / alpha 16 |
| Steps | 3200 · ai-toolkit · bf16 · adafactor · lr 1e-4 |
| Resolution | 768 × 1024 |
| Triggers | prvpns (identity) · prvpene (contact) |
| Weight | Free Smell turbo couples 0.42–0.44 · Gonzalomo ~0.55 · solo close 0.85–0.90 |
| Mix | ~96% contact-weighted slots · stills + curated video frames |
Adult technical material (18+). Private consensual source photography / video stills. Intended for model card transparency and reproducible fine-tuning documentation.