LTX-2 video model

The video‑generation model from Lightricks called LTX-2 has arrived on 24th of October 2025. The press release says “open‑source foundation model” and that weights/code will be released later this year.

LTX‑2 is a next‑gen AI video engine that claims to generate high‑fidelity, synced audio + video clips (up to 4K resolution and high frame‑rates) from text, images or video inputs. It’s built for both rapid ideation (“Fast” mode) and higher‑fidelity production (“Pro” mode), and is marketed as efficient enough to run on consumer‑grade GPUs (meaning running on a 5 series is realistic once optimized.)

Two versions available (Fast / Pro) right now for API use; a future “Ultra” mode is mentioned.

Pro model compared to fast:

• Designed for higher visual fidelity with fast turnaround.
• Balanced mode: strong quality (higher resolution, higher frame-rate) but still practical for daily content creation, marketing teams, iterative workflows.
• More suitable when you want a deliverable-ready video rather than just a rough draft
• Likely requires more computational resources (or higher API cost) than Fast but less than the future Ultra mode.

👉 LTX-2 became 'open source' in January 2025. 

Licence states that it is free for community use; paid license required for large commercial entities; not OSI open source.

Here’s what you’re allowed to do with it. Run it on your own machine. Use it in ComfyUI. Fine-tune it. Put it in apps that don’t compete with it. Make money from what it creates. Build internal tools. Share tweaked versions as long as you use the same license.

Here’s what you can’t do. Make a competing AI product. Use the outputs to train other models. Change the license on your version. Ignore the rules. Use it for big commercial stuff (like over $10M) without paying.

What’s included. Full model weights and a smaller version. LoRAs and IC-LoRAs. A modular trainer for tuning. Inference that runs fast on NVIDIA RTX cards.

👉 LTX-2 VRAM Needs

LTX-2 needs lots of VRAM. If VRAM runs out, NVIDIA and ComfyUI made a system that shifts some work to RAM. 

Say you’ve got a 5090 with 32GB VRAM.

A 4s 720p 24fps video takes ~25s.

Double the length to 8s, now it’s 3 minutes with weight streaming.

Tip: use low settings first, then raise them when ready.

Nvidia recommends this:

• Cut length. 4s on 16GB+, 3s on 12GB+.
• Lower res. 720p for 16GB+, 540p for 12GB+.