Consistent characters across every shot
Feed LTX-2.5 one reference sheet with characters, props, and location. Every generated clip stays faithful to it. Open weights, self-hosted, no per-clip cost.
.svg)
Key capabilities
Reference-sheet conditioning
One composite sheet locks identity across the clip. Bigger panels on the sheet carry over more faithfully into the output.
Two-part prompt control
"Reference sheet:" describes what's on the panels; "Generated video:" drives the action. Look and action stay separately controllable.
Episodic & multi-shot sequences
Keep the same recurring character consistent across every shot in a series or season without regenerating the character each time.
Branded mascots and campaigns
Lock a branded mascot or product visually consistent across every ad in a campaign. The reference sheet acts as your visual anchor.
Consistent identity from one reference sheet
Lock characters, props, and location from a single reference sheet, then carry all three into every generated shot. No drift across takes.
Self-hosted, no seat licenses
Run the model on your own infrastructure. No per-seat fees, no vendor lock-in. You control the pipeline end to end.
How it works
Input
Technical characteristics:
- Reference sheet — one clean panel per element (character face close-up and turnaround, product-style props, location) on a black background with no text. Supplied as a static video looped to 121 frames minimum at output resolution and fps.
- Two-part prompt — "Reference sheet: [description] / Generated video: [action]"
- Base — LTX-2.3-22B plus ltx-2.3-22b-ic-lora-ingredients-0.9.safetensors
- Recommended settings — strength 1.4, 30 inference steps, guidance scale 4.0, validated with spatiotemporal guidance (STG mode stg_v, block 29, scale 1.0) for motion stability
Output
Technical characteristics:
- Short clip — characters, props, and setting match the reference sheet
- Best results — trained bucket 768x448, 121 frames, 24fps; other resolutions or much longer clips are out of distribution
- Self-host — via ComfyUI on GPU; only reproduces elements present on the sheet, with bigger panels carrying over more faithfully
What users are building with LTX
FAQs
How does it keep characters consistent?
You supply a reference sheet (characters, props, location) as conditioning; generation stays faithful to it.
What resolution and length should I use?
768x448, 121 frames, 24fps is the trained bucket where you'll get the best results. Other resolutions or much longer clips are out of distribution. Use strength 1.4, 30 steps, and guidance scale 4.0.
What makes a good reference sheet?
One clean panel per element on a black background with no text. Bigger panels carry over more faithfully into the generated video.
Can it invent a character that isn't on the sheet?
No. It only reproduces elements present on the reference sheet.
What does it cost to run?
Open weights on HuggingFace, self-hosted via ComfyUI under the LTX-2 Community License. No per-clip cost and no seat licenses.
