LTX-2.3 OmniCine Preview v0.1 (LoRA, VibeBoard mirror)

Mirror of TenStrip / Singularity LTX-2.3 OmniCine Preview v0.1 (CivitAI), hosted here for stable, anonymous Docker-build-time hf_hub_download access from the VibeBoard GPU worker.

What this is

A LoRA adapter for ltx-2.3-22b-distilled-1.1 that improves motion realism — specifically:

Significantly improved finger / toe stability and limb consistency
Reduced "AI Stiffness" with natural micro-expressions
Identity preservation during subject occlusion (e.g. back-turned moments during rotation)

Validated on VibeBoard 2026-05-20 against the v701.3.0 HDR distilled pipeline. Default scale 0.9.

Tradeoff

The model spends temporal compute budget on anatomical accuracy instead of dramatic motion → spin / rotation speed is slightly reduced in exchange for physically realistic motion. For most production work this is a favorable trade.

The author notes "high-dynamic physics (extreme motion) is still being refined" — VibeBoard testing on a fast-twirl prompt confirmed the LoRA still helps despite that caveat, but per-scene strength tuning may be worthwhile via the user-facing slider.

License

LTX-2 Community License — inherits from the LTX-2.3-22b-distilled-1.1 base under the LTX-2 license's copyleft clause. Pre-$10M-revenue commercial use is permitted. See the original CivitAI page for any author-specific notes.

Source

Original: https://civitai.com/models/2610733?modelVersionId=2931450
Author: TenStrip
Training: ~11.5K steps ("Early Alpha", ~10% of intended training)
File: omnicine-v01-preview.safetensors (2.51 GB)

Usage in VibeBoard

Auto-stacked on the LTX-2 HDR distilled pipeline when omnicine_scale > 0 (default 0.9). See gpu-worker/handlers/ltx2_handler.py and the Shot Studio Advanced Tuning UI.

Downloads last month: 32