LTX-Video to LTX-2: Version-Safe Local and ComfyUI Video Workflows

LTX is a version-sensitive family: the original LTX-Video repository began with text-to-video and image-to-video, and later development moved toward LTX-2 with synchronized audio-video generation. A production tutorial must identify the checkpoint and software stack instead of instructing all users to download a vaguely named ‘LTX model’.

A documented development line

The official Lightricks repository records an initial LTX-Video 0.9.0 release on 21 November 2024. Its changelog lists additional checkpoint updates, improved controls and support for keyframes and extension by March 2025. The current project README points to LTX-2 as the development home for synchronized audio-video output and multiple-keyframe control, including integration with ComfyUI.

These history notes are enough to distinguish feature generations. They are not enough to certify one set of weight-license terms, GPU memory requirements or output settings for all checkpoints. One March 2025 note mentions an OpenRail-M commercial-use license for a particular release, which should never be generalized to every LTX-2 derivative without reading the exact model card.

Workflow graph discipline

Local video rendering often involves encoders, conditioning nodes, motion components, decoders and an output combiner. In a visual workflow editor, pin every node version, ensure that accepted input dimensions and FPS match, and save the graph with the exported video. Do not let auto-updated custom nodes silently change a reproducible project between renders.

For a character-led scene, build the opening frame and scene palette first, then animate a single motion. If the pipeline permits keyframes, use them to lock important visual transitions. If the pipeline produces audio too, check that its duration matches the rendered frames and that dialogue has not been inadvertently generated over approved narration.

Operational costs and caveats

Even with open model access, local compute is not free. Document GPU memory, sampling speed on your own hardware, disk use, job failure rate and engineer time before declaring savings. A cloud GPU job that runs for hours at an unknown retry rate can cost more than purchasing several hosted video clips. Never publish test timings or comparative rankings that you have not personally measured and recorded.

A robust archive includes model checksums, licenses, workflow JSON, input hashes, seeds, output previews and acceptance records. Because exact LTX-2 specifications and model terms can vary across checkpoints, this article is marked review for production-ready specification claims even while its version-history source is verified.

Worked production example

A motion designer opens a saved ComfyUI graph six months after it was made. The graph may name an obsolete LTX-Video checkpoint while the workstation now has LTX-2 libraries. Preserve the old dependency manifest, output examples and model checksum before updating. Run a small non-sensitive reference shot and check that audio channels, frame count and keyframe timing behave as intended. If the migrated version creates new sound, do not overlay it onto approved narration without review. Treat a software upgrade as a new pipeline version, not an invisible replacement of the previous one.