aivideo.zone LTX-2 v1.5.0: replace and guide keyframes, interpolation in the main pipeline, fp8 on newer GPUs
v1.5.0 release notesLightricks' LTX-2 v1.5.0 release and its notes, on GitHub
Lightricks published v1.5.0 of the LTX-2 code on 18 October 2026. It is the code that runs the LTX-2.5 open weights, not a new model, and it changes how a still image steers a clip.
Replace or guide
An image binding now takes a mode. On the command line it is --image PATH FRAME_IDX STRENGTH replace|guide, with an optional CRF value at the end. Replace puts the image into the clip as that frame; guide lets it steer the frames around it without being copied in.
Replace has a constraint: the frame index, counted in pixel frames, must be 0 or 1 plus a multiple of 8, so 1, 9, 17 and on. In Python, ImageConditioningInput.at, .replace and .guide take the place of the old frame_idx argument.
Keyframes in the main pipeline
Keyframe interpolation, making a clip that passes through several given images, is now done by TI2VidTwoStagesPipeline with guide bindings, and the separate KeyframeInterpolationPipeline is removed. Audio can be refined in the same run with --refine-audio, or refine_audio=True from Python.
Faster on newer cards
The performance work was done with doubleAI. Matrix multiplications run in fp8 on GPUs of compute capability 8.9 and above, the RTX 40 series and L40 generation onwards, and H100. The DFR epilogue is tiled at no more than 1024 px with 64 px of overlap, and the Gemma text encoder is faster across several GPUs.
Not bit-identical
Lightricks says outputs from v1.5.0 are not bit-identical to earlier releases. The same seed and prompt will give a slightly different clip.
A bug fixed
Long 1080p clips could turn grey from frame 128 onwards under PyTorch older than 2.13, because RMSNorm offsets wrapped past 232. That is fixed.
Where this fits
September's 1.4.0 added long clips in overlapping windows and a one-pass HDR master; see LTX-2.5's September.
Two things to do before upgrading. Pin the old release in any project where a client has approved a shot, because the new numerics will not reproduce it. And move scripts off KeyframeInterpolationPipeline now: it is gone, and the guide binding in the two-stage pipeline is its replacement. The fp8 gain only arrives on Ada-, Hopper- or newer cards; an RTX 30 series or an A100, at compute capability 8.6 and 8.0, does not take the fp8 path.
LTX-2 v1.5.0: replace and guide keyframes, interpolation in the main pipeline, fp8 on newer GPUs: questions
What is new in LTX-2 v1.5.0?
Image bindings that either replace a frame or only guide the clip, keyframe interpolation inside TI2VidTwoStagesPipeline, fp8 speed-ups on compute capability 8.9+ GPUs, and a fix for grey frames in long 1080p clips.
Which frames can a replace binding use?
Frame 0, or any frame that is 1 plus a multiple of 8 in pixel space: 1, 9, 17 and so on.
Will my old seeds give the same video?
No. Lightricks says v1.5.0 output is not bit-identical to earlier releases, so keep the old version for any clip you need to reproduce exactly.
Source: Lightricks on GitHub checked against the source
More AI video news
Helix World preview
Noiz AI and HKUST have released Helix World preview, an interactive audio-visual world model: 768×512 AI video at 24 frames a second generated live on a single H800 as you steer with WASD or the arrows, with 48 kHz spatial stereo sound. Built on LTX-2.3, weights on Hugging Face, demo at helixworld.org.
LTX-2.5 gets fourteen LoRAs and HDR
Lightricks spent September turning LTX-2.5 into a post-production kit: fourteen LoRAs for jobs like mattes, deblurring and day-for-night, long AI video in overlapping windows, and an SDR-to-HDR pass that writes EXR.
ComfyUI 0.32: LTX-2.5 and Grok Imagine
ComfyUI 0.32 shipped on 11 August with support for LTX-2.5, Lightricks' new open AI video model, partner nodes for Grok Imagine Image 2.0 and Qwen-Image 3.0 Pro, and a higher minimum PyTorch.