LTX vs Wan

The License Is the Main Difference

Wan 2.2's Apache-2.0 license lets anyone use, modify and redistribute the model for any purpose, with an express patent grant and only notice and attribution requirements. The model card asks users not to generate illegal, harmful or misleading content, and Alibaba claims no rights over generated output. There are no revenue thresholds, field-of-use limits or regional exclusions.

The LTX-2.x Community License Agreement (dated August 11, 2026) is royalty-free and worldwide for most users, but it is not open source in the Apache sense. Its key terms, quoted from the license:

• Revenue threshold: "Entities with annual revenues of at least $10,000,000" are "Commercial Entities" and "are required to obtain a paid license for any use" other than narrowly defined non-commercial purposes such as testing and evaluation.
• Competitive use: you may not use it "in any product, service, or application that directly competes with Licensor's commercial products or services" without a separate commercial license.
• Training other models: for commercial use, you may not use it to train, improve or fine-tune "any other machine learning model", except permitted derivatives.
• Derivatives: fine-tuned weights and models trained on its outputs "must be distributed exclusively under the terms of this Agreement".
• Provenance: you may not circumvent its "watermarking, content provenance or latent disclosure functionalities".
• Output and law: Lightricks claims no rights in the output you generate. The license is governed by New York law, and its only geographic limits relate to comprehensively sanctioned countries and US restricted-party lists.

For a startup or creator under the threshold, LTX-2.5 is effectively usable at no cost. For a larger company, or anyone building a video product that might compete with LTX Studio, it requires a commercial agreement with Lightricks, and companies should have counsel read the full text. Wan 2.2 needs no such review beyond standard open-source compliance.

Capabilities: A Generation Apart

LTX-2.5 is a year newer and it shows. It generates synchronized audio and video in one model, supports multi-shot generation in a single pass, and defaults to 1024×1536 at 24 frames per second, with a documented 4K UHD path (3840×2176) using its upscalers. Lightricks ships pipelines for keyframe-guided generation, audio-to-video, IC-LoRA control and dubbing, plus a distilled version of the transformer that runs in far fewer steps for fast iteration.

Wan 2.2 is narrower. Its A14B text-to-video and image-to-video models generate five-second clips at 480p or 720p, using a mixture-of-experts design in which separate expert networks handle different stages of denoising. Its TI2V-5B model combines text- and image-to-video in a dense 5-billion-parameter model with a high-compression VAE, generating 720p at 24 fps. The core models produce video without sound; Alibaba has released separate open variants for speech-to-video and character animation, including Wan2.2-Animate-2 in August 2026.

On Artificial Analysis's text-to-video-with-audio leaderboard as of October 20, 2026, the hosted LTX-2.5 Fast and Pro versions rank 23rd and 24th. That puts them well behind the closed leaders, but they are among the few entries on the board whose weights anyone can download. Wan 2.2 isn't on that board because it doesn't generate audio, while its closed successor, Wan 3.0, ranks first.

Hardware and Deployment

Both are heavy. LTX-2.5's components (transformer, Gemma-based text encoder, video and audio VAEs and upscalers) total tens of gigabytes in bf16, and Lightricks documents fp8 quantization with CPU or disk offloading for GPUs with limited memory, along with other quantized formats. Its weights are gated on Hugging Face: you sign in and accept the license before downloading.

Wan 2.2's A14B models need about 80 GB of VRAM to run on a single GPU, which means data-center cards, but the TI2V-5B model is designed to run on a 24 GB consumer card such as an RTX 4090, taking several minutes per five-second clip. That makes Wan 2.2's small model one of the most accessible ways to experiment with local video generation on a workstation (see GPU computing and edge AI). Both have mature ComfyUI support, and Wan 2.2 also works with Hugging Face Diffusers.

Ecosystem and Momentum

Wan has a longer open history. Wan 2.1 (February 2025) and Wan 2.2 were both released under Apache-2.0, which lets anyone build and redistribute fine-tunes and LoRAs without restriction, and Alibaba has kept publishing specialized open Wan 2.2 variants. The open line has slowed, though: Alibaba's general-purpose Wan releases since 2.2 have been API-only, so Wan 2.2 may remain the high-water mark for open Wan.

Lightricks has moved in the opposite direction. It released LTX Video (2B) as open weights in November 2024, open-sourced the 19B LTX-2 in January 2026 and followed with LTX-2.5 in August, alongside its commercial LTX Studio product. Its license reflects that business model: smaller users get the weights at no cost, and larger ones pay. For teams that want an actively developed open-weight model with audio, it's currently the strongest option; for teams that need maximum legal freedom, Wan 2.2 is.