LTX-2 v1.5.0 Adds Keyframe Support and Performance Enhancements

What happened

The LTX-2 model version v1.5.0 adds support for explicit latent-replacement and keyframe-guide bindings through the ImageConditioningInput class. This update also enhances the TI2VidTwoStagesPipeline with keyframe interpolation capabilities, allowing users to guide video generation with image conditions.

Performance improvements include faster transformer passes, LoRA switches, and denoising samplers without per-step CPU synchronization. The DFR spatial epilogue now uses fp8 GEMMs on GPUs with compute capability 8.9+.

Why it matters

These updates address issues like gray frame tails in DiffVAE decodes and synchronization problems during multi-GPU operations. By improving the efficiency of image and video generation, the update supports more stable and faster workflows for developers and researchers.

Key details

The ImageConditioningInput class now uses ImageReplace or ImageGuide bindings instead of frame indices, with specific methods to set these bindings. The MGPUController caps intra-op CPU threads to 8 per rank to prevent stalling on many-core hosts.

Multi-GPU pipelines use the default Accelerate-sharded Gemma for faster prompt encoding. The BroadcastCoordinator has been moved to ltx_core.multigpu.broadcast, and BroadcastGemmaWrapper allows optional coordinator parameters.