FurTricks — LTX-Video 2.3 Full Fine-tune
FurTricks is a full fine-tune of Lightricks LTX-Video 2.3 (22 billion parameters) on a curated dataset of furry and anthropomorphic character videos sourced from E621.
⚠️ NSFW / Not Safe For Work.
Model Details
| Property | Value |
|---|---|
| Base model | LTX-Video 2.3 22B (ltx-2.3-22b-dev) |
| VBVR merge | LiconStudio/Ltx2.3-VBVR-lora-I2V merged into base before training |
| Training type | Full fine-tune (all 18.9B trainable parameters) |
| Training steps | 2,000 (Phase 1 of 3) |
| Resolution | 768×448, 9–121 frames (multi-bucket) |
| Dataset | ~5,000 E621 videos with AI-generated captions |
| Captioner | Qwen3-Omni-30B with structured tag + narrative format |
| Precision | BFloat16 |
| Hardware | 4× H200 141GB GPUs via Modal |
| Framework | Lightricks LTX-2 Trainer + DeepSpeed ZeRO-2 |
Intended Use
This model is intended for adults only for the purpose of generating anthropomorphic / furry character videos. It understands E621-style tag vocabulary natively and responds well to structured prompts describing:
- Character species, anatomy, and appearance
- Actions and scenes
- Camera angle, lighting, and composition
- Audio cues
Prompting Guide
The model was trained on captions in the following structure:
[e621 tags]. [Subject]. [Action]. [Camera]. [Lighting]. [Audio].
Example:
anthro, dragon, blue_scales, male, muscular. A blue-scaled anthropomorphic dragon.
Swims gracefully through a sunlit underwater cavern, fins rippling.
Smooth tracking shot following from the side. Soft caustic lighting from the surface.
Ambient underwater bubbling sounds.
For best results, lead with relevant E621 taxonomy tags before the descriptive sentence.
Limitations
- This is an early checkpoint (2,000/5,000 steps of Phase 1). Coherence and motion quality will improve in later phases.
- Complex multi-character scenes may be inconsistent.
- Audio generation is not enabled in Phase 1 outputs.
- The base model's knowledge of real-world places, text, and non-furry subjects is partially retained but not the focus of this fine-tune.
- More steps will be uploaded as we train more.
License
The model weights inherit the Lightricks Research License from the base model. Non-commercial research and personal use only.
Acknowledgements
- Lightricks for LTX-Video 2.3 and the open-source LTX-2 trainer
- LiconStudio for the VBVR I2V LoRA
- E621 community
- Downloads last month
- 139