LTX-2 1.5.0: 4K renders in about half the time, plus keyframes you can pin or merely guide

LTX-2 1.5.0 landed on GitHub on Oct 18. doubleAI published the story behind it on Oct 19, and @ltx_io announced it this morning. doubleAI pointed its Maxwell 'code execution model' at LTX's already heavily tuned pipeline and came back with more than a dozen changes, from how work is split across tiles and GPUs down to individual GPU kernels. Their headline result: a five-second 4K clip on two B200s went from 250 seconds to 134, a 1.87x speedup. They say the soundtrack is identical bit for bit and the video is visually indistinguishable (41 dB PSNR against the original). The changes are in LTX's open code. Maxwell itself is only in early access.

The cleverest fix is easy to explain. LTX-2.5 generates audio and video together, and the roughly hundred audio tokens attend to tens of thousands of video tokens in every block. With the video split across two GPUs, the old code copied about 262 MB of video keys and values between the GPUs in every block, at every step. The new code has each GPU compute a partial answer from its own half and swaps those partial answers instead (about 1 MB), then merges them exactly, using the same trick FlashAttention uses inside one GPU. CPU work also moved onto the GPU and thousands of tiny launches were batched, so speed now barely depends on your host CPU. That matters on cloud boxes with weak CPUs.

Most people won't run two B200s, so here's what changes on smaller setups, according to the changelog: faster diffusion-VAE decoding that reserves less GPU memory, lower memory while encoding the MP4, no per-step CPU sync in the samplers, faster LoRA switching, and fewer, larger tiles for the 2x spatial upscale step, using fp8 maths on GPUs that support it (compute capability 8.9+). One caveat: Lightricks says output is no longer bit-identical to earlier releases, mostly because of the new upscale tiles and fp8. Re-check any seeds you rely on before you update a production pipeline.

The creative change: each image you condition on can now be marked replace (that frame is pinned exactly) or guide (it steers the shot without being copied). On the CLI that's --image PATH FRAME_IDX STRENGTH replace|guide. Keyframe interpolation moved into the main two-stage pipeline (use guide frames, plus --refine-audio if you want the old stage-2 audio pass), and the separate KeyframeInterpolationPipeline was removed. Scripts that call it, or that pass frame_idx to ImageConditioningInput, will break: switch to ImageConditioningInput.at(). Weights are under the LTX Model License, free for organisations under $10M in annual revenue. The speed numbers are vendor benchmarks we haven't reproduced. Spotted, not tested. Credit Lightricks and doubleAI.