openclaw - 💡(How to fix) Fix [Feature]: Add LTX-Video support to video_generate [1 comments, 2 participants]

Summary

First-class support for Lightricks' LTX-Video models in the shared video_generate capability.

Problem to solve

LTX-Video is one of the few open-weight video models with near-real-time generation (DiT-based, ~4–5s for a short clip on a single GPU) and strong text/image-to-video quality. It's currently absent from the supported providers list in docs/tools/video-generation.md and docs/tools/media-overview.md. Users who want fast, locally-runnable video generation, or hosted LTX via fal/Replicate, have no documented path — they have to reverse-engineer model IDs through the bundled fal plugin and there's no preset, no defaults, and no failover entry for LTX in agents.defaults.videoGenerationModel ordering.

Proposed solution

Any one (or combination) of the following, in order of effort:

1. Document LTX through existing plugins. Add LTX model entries (fal-ai/ltx-video, fal-ai/ltx-video-13b, Replicate equivalents) to the model tables in docs/providers/fal.md (and Replicate / Together docs if applicable), with recommended defaults for duration, resolution, and reference-image flow.
2. ComfyUI workflow template. Ship a reference LTX workflow under the comfy plugin so self-hosted users can run it via video_generate without authoring their own graph.
3. Dedicated lightricks provider extension. Wraps the official LTX hosted API (when stable) for users who want a direct integration rather than going through fal/Replicate. Same shape as extensions/runway — single API key, single capability surface (text-to-video, image-to-video, single-reference flow).

Path 1 is the lowest-risk starting point and unblocks users today. Path 3 is a maintenance commitment and probably belongs as a third-party plugin unless someone steps up to own it.

Alternatives considered

• Leaving it as a "configure fal yourself" exercise: works for power users but invisible to the agent's default model-discovery and to docs readers.
• Pure plugin-SDK route: viable, but LTX is broadly enough adopted (HuggingFace, fal, Replicate, ComfyUI ecosystem) that even minimal docs in core would help.

Impact

• Affected: any OpenClaw user wanting fast/cheap video generation, especially self-hosters using ComfyUI.
• Severity: low-medium — workaround exists via raw fal model IDs, but discoverability and defaults are missing.
• Frequency: every video-generation user evaluating provider options.
• Consequence: users either don't know LTX is reachable or pick a slower/more expensive model by default.

Evidence/examples

• LTX-Video repo: https://github.com/Lightricks/LTX-Video
• Hosted on fal (fal-ai/ltx-video, fal-ai/ltx-video-13b) and Replicate.
• Existing comparable provider entries: docs/providers/runway.md, docs/providers/minimax.md.

Additional information

Happy to start with a docs PR for path 1 if maintainers think that's the right scope. Flagging that path 3 should likely live as a third-party plugin per CONTRIBUTING.md unless there's a clear maintainer ask for first-party.