ltx25-image-to-video
Installation
SKILL.md
ltx25-image-to-video
Run every graph with the bundled client: python scripts/comfy_client.py --url $COMFY_URL run workflows/<file> --set NODE.INPUT=VALUE --out out/. Upload input files first with python scripts/comfy_client.py upload <file>. Before the first run on a server, check it with python scripts/preflight.py --url $COMFY_URL workflows/<file>. Node packs, model files, VRAM and weight licenses for this skill are in references/requirements.md. Use open-weights models only, and never swap in a hosted or closed model. LTX weights are under the Lightricks LTX license, which is not an OSI-approved open-source license (it is open weights with use conditions), so read it before commercial use.
Pick a graph
| Need | Workflow | Set |
|---|---|---|
| I2V, two-pass with latent upsampling (sharpest) | ltx25_i2v_two_pass.api.json |
395.image, 364.text, 356.width/height/length, 338.noise_seed |
| I2V from a start frame, with an optional last frame | ltx25_start_frame_i2v.api.json |
401.image start, 402.image last, prompt in 900.string_a, 356.* size and length |
| Text to video | ltx25_t2v.api.json |
prompt in 900.string_a, 356.* |
Rules
lengthmust be 8n+1 frames (97, 121 ...). Width and height must be multiples of 32. At 24 fps, 121 frames is about 5 s.- Write the prompt for motion, not appearance: what moves, the camera move, and the pacing. The keyframe already carries the look.
- Keyframe chaining: use the last frame of shot N (extract it with ffmpeg
-sseof -0.05) as the start frame of shot N+1 when two shots must cut together without a visible jump. - Never change the sampler or sigma nodes. Leave the guider and
ManualSigmasas they are. - Google Colab: on a 24 GB L4, run the two-pass graph at 640x352 or the start-frame graph at 672x384.