生成式影片與開放模型
LTX‑2.5 Updates Its Open Video Model With a Diffusion Decoder and Single-Pass Multi-Shot Generation
LTX‑2.5 releases trainable 22B weights and eight-step distilled weights, adding character, audio, and scene consistency across shots. The complete quick-start components total roughly 66 GiB, Diffusers support still requires installing from the main branch, and the model is subject to a community license with a revenue threshold.
Lightricks released the weights and inference tools for LTX‑2.5 on August 12. At its core is still a 22B DiT capable of generating video and audio in sync, but the generation pipeline includes three substantive changes. First, the model can produce multiple connected shots in a single inference pass, attempting to preserve character identity, environments, lighting, audio, and style across cuts. Second, it replaces straightforward VAE reconstruction with a diffusion-based video decoder to improve details in faces, text, textures, and fast motion. Third, detail-fidelity rendering allocates compute according to scene complexity, then completes refinement with a spatial latent upscaler and an optional temporal latent upscaler.
The release includes a fully trainable BF16 DiT, a distilled variant with a fixed eight-step schedule and CFG=1, an INT8 version specifically for ComfyUI, and NVFP4 weights targeting Blackwell. On the text side, it uses a custom Gemma 4 12B encoder, along with a duration head that predicts the frame count from the prompt. The transformer, text encoder, audio-video VAE, and upscalers required by the official Python quick start total roughly 66 GiB. Systems with insufficient memory can use FP8 casting and CPU or disk offloading, though this generally trades transfer latency for the ability to run the model. The folder-based component packaging also means legacy single-file loaders cannot be reused directly.
Engineering integration is not yet fully mature. The official `ltx-pipelines` requires Python 3.12 and recommends CUDA 12.7 with PyTorch 2.7. A Diffusers compatibility package is available but has not yet entered a stable release, so users must install it from GitHub’s main branch. Most LTX‑2.3 LoRAs are said to work directly, but the company acknowledges exceptions, making individual validation essential before production deployment. The weights also do not use a permissive Apache-style license: entities with less than $10 million in annual revenue may use them commercially at no cost, while larger organizations must sign a separate agreement, and transferring fine-tuned models may incur fees. Key areas to watch next include independent audio-video synchronization benchmarks, identity drift across shots, and the diffusion decoder’s real-world VRAM and latency overhead.