LTX-2.3 Singularity OmniCine V1: base model, first/last frame mode, timeline prompts
Singularity-LTX-2.3 OmniCine V1: LTX-2.3 fine-tune (~100k accumulated steps) for I2V, first/last frame and reference-to-video. Base: ltx-2.3-22b-distilled-1.1 transformer-only fp8_scaled; ComfyUI workflow in repo, first & last frame mode. Prompts in English: scene/style, timed action, timed camera, environment, timed dialogue, audio/technical (e.g. no subtitles).
Singularity-LTX-2.3 OmniCine V1 by WarmBloodAban (author of the Singularity MiniMax-H3 models) is a fine-tune of LTX-2.3 aimed at image-to-video, first and last frame control and reference-to-video, trained for about 100,000 steps counted with gradient accumulation. The author claims better fingers and limbs in fast motion, timestamp-controlled shot and camera cuts, more expressive faces and lip-sync without burned-in subtitles, steadier physics, anime/CGI/realistic styles, and reference images used for identity and style rather than as a fixed first frame. Slight motion blur can remain in extreme action.
Running: the recommended base is ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors. A ComfyUI workflow is in the repository files, and the author recommends its first and last frame mode for the most scene control. The card does not state whether the release is a LoRA or a full checkpoint (it is tagged lora) or give a strength, step count or CFG.
Prompt structure (keep everything in English): [Scene & Style] one-sentence look; [Action Timeline] "0-X seconds, action or emotion"; [Camera Timeline] "0-X seconds, camera movement or composition"; [Environment] light source, contrast and grading; [Dialogue] "0-X seconds, Character says: \"...\""; [Audio & Technical] background sound, film grain, and exclusions such as "no subtitles". The card example: cinematic wuxia, dim interior, 0-10 seconds tight close-up of a confused young man, static camera with slight handheld movement, candlelight bokeh, a quoted line, voice low and slow, precise lip-sync, film grain, no subtitles.
Licence: the card metadata says apache-2.0; the LTX-2.3 base model it runs on has its own licence.
Based on the WarmBloodAban/Singularity-LTX-2.3_OmniCine_V1 card pinned to revision 8ea615dc49e6ea6637be6557b188f76c0ee6ef14 and observed on 2026-10-27.
Nothing was run here; the demo videos and workflow JSON were not inspected, and quality claims are unverified.
The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.
Structured fields
{
"citations": [
{
"url": "https://huggingface.co/WarmBloodAban/Singularity-LTX-2.3_OmniCine_V1/raw/8ea615dc49e6ea6637be6557b188f76c0ee6ef14/README.md",
"note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 10358:10720, 10512:10720, 10956:11061, 11219:11577, 11285:11577, 11430:11577, 598:784, 4:131; exact excerpt SHA256 e2f69d1d9970932428902d4e50b4d90912aa1f7352747e478baa2a8d6fe713a2,22328b4e9b77d0066aa9048b5f38eb6cd27cde09fbb56e5d09f68791b96bc668,59cf4d94e0069189aa0bb1ace89fe2a3d"
}
]
}Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | permitted | unknown | Original attributed paraphrase and short functional values only. Source text, images, weights and other assets remain separately governed. |