multimodalart's picture
multimodalart HF Staff
Upload README.md with huggingface_hub
dae86b5 verified
|
Raw
History Blame Contribute Delete
1.41 kB
---
title: LTX-2.3 Cinemagraph
emoji: 🎞️
colorFrom: yellow
colorTo: pink
sdk: gradio
sdk_version: 6.13.0
python_version: "3.12"
app_file: app.py
pinned: false
short_description: Still image to looping cinemagraph (LTX-2.3-22B LoRA)
models:
- Lightricks/LTX-2.3
- Lightricks/LTX-2.3-22b-LoRA-Cinemagraph
startup_duration_timeout: 1h
---
# 🎞️ LTX-2.3-22B Cinemagraph
Turn a **still image** into a looping **cinemagraph** β€” one element moves
(water, neon, clouds, reflections) while the rest of the scene stays frozen.
- **LoRA:** [`Lightricks/LTX-2.3-22b-LoRA-Cinemagraph`](https://huggingface.co/Lightricks/LTX-2.3-22b-LoRA-Cinemagraph)
on the [LTX-2.3-22B](https://huggingface.co/Lightricks/LTX-2.3) base.
- **Pipeline:** native single-stage `TI2VidOneStagePipeline` from the latest
[LTX-2](https://github.com/Lightricks/LTX-2) codebase (no pinned commit),
bf16 everywhere β€” **no fp8 quantization, no LoRA fusion**.
- **Settings** follow the model card: 30 steps, guidance 4.0, STG scale 1.0 on
block 29, LoRA strength 0.7–3.0, best at 512Γ—704.
- **Text encoding** (Gemma-3 + LTX-2.3 connectors) is offloaded to a remote
encoder Space (`linoyts/ltx23-gemma-encoder-api`) so Gemma never loads here.
Be explicit about what should move and what stays still, avoid camera language,
and use "seamless natural loop". The `CINEMAGRAPH_MOTION` trigger word is added
automatically.