LTX-2 v1.4.0 Enhances Video and Audio Generation Tools

What happened

LTX-2 v1.4.0 introduces new capabilities for video and audio generation, including precompiled diffusion-VAE kernels and tools for long-video creation. The update also includes features for handling audio and video conditions with latent indices.

The release expands the toolset for developers working with LTX-2, providing enhanced support for decoding, keyframe management, and spatial tiling. New pipelines and utilities allow for more efficient handling of large canvases and multi-stage processing.

Why it matters

The improvements in LTX-2 v1.4.0 address performance and flexibility in video and audio generation workflows, enabling better support for complex projects. These updates help developers manage large-scale tasks more efficiently, reducing computational overhead and improving output quality.

Key details

The release includes precompiled diffusion-VAE kernels for LTX-2.5, streamlining decoding processes. It also introduces the ICLoraPipeline, which allows users to define sequences of IC-LoRA stages with customizable settings.

New tools like spatial_tiled_ic_lora_stages and TiledDiffusionModel support large canvases through spatial tiling, and the --tile flag enables transformer tiling with a default window size of 1024x1536. The decode_with_keyframes feature allows for keyframe-aware decoding across multiple pipelines.