WorldFoundry
English中文

LTX-Video

Method overview: LTX-Video’s Video-VAE plus a 3D transformer that denoises in latent space.

Lightricks LTX-Video 0.9.8 distilled I2V on the shared native LTX stack. Local T2V labels are upstream metadata, not a WorldFoundry route.

Edit this pageGitHubMarkdown

LTX-Video is Lightricks' earlier realtime latent-diffusion video model. This catalog card is the 0.9.8 distilled image-to-video route — not LTX-2.x and not LTX-2.5. The single runnable variant is ltx-video-i2v. Family-level text-to-video tags describe upstream capability; the native runtime raises if you ask it for T2V.

Official paper teaser: LTX-Video realtime latent-diffusion samples from the paper.
Official paper teaser: LTX-Video realtime latent-diffusion samples from the paper.
Method overview: LTX-Video’s Video-VAE plus a 3D transformer that denoises in latent space.
Method overview: LTX-Video’s Video-VAE plus a 3D transformer that denoises in latent space.

The paper is LTX-Video: Realtime Video Latent Diffusion. Code lives at Lightricks/LTX-Video. The project site is ltx.video. There is no official teaser or method figure checked into this docs tree — paper.png is the title thumb only.

Build your run

Pipeline
LTXVideoI2VPipelineworldfoundry.pipelines.ltx_video.pipeline_ltx_video_i2v
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  ltx-video-i2v \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Animate a still on ltx-video-i2v: required prompt + image, optional video, seed in actions, artifact ltx-video-i2v.mp4 through LTXVideoI2VPipeline. Diffusers is not the inference backend. T5 loads as a Transformers architecture via NativeModuleLoader; the DiT, Video-VAE, Euler schedulers, spatial upscaler, and joint-state runner are the same native stack later LTX-2.x cards reuse.

There is no local T2V or V2V on this page. Listed I2V eval is AIGCBench — a suite pointer, not an ingested score.

Inference

WorldFoundry pins the distilled two-pass recipe: 7 + 3 = 10 fixed-sigma Euler steps, guidance 1.0 (the runtime rejects any other step count or CFG). The spatial upscaler runs automatically between passes at first-stage scale 2/3; it is not a Studio refiner toggle. Sampler is the checkpoint / fixed-sigma Euler pair, not UniPC.

Recorded defaults from runtime_defaults.yaml:

KnobWorldFoundry defaultConstraint
Size768×512height and width divisible by 32
Frames121upstream also wants frames = 8k+1
FPS30frame_rate
Steps10 (7 then 3)required
Guidance1.0required
Image noiseimage_cond_noise_scale=0.15first-frame conditioning
Offloadcpu_offload=trueblock offload

Upstream 0.9.8 cards advertise 1216×704 @ 30 fps and recommend guidance 3–3.5 on the dev (non-distilled) line. Those numbers are not this wrapper's defaults. Catalog VRAM is unset — do not invent a floor.

Running in WorldFoundry

Launch worldfoundry-eval run ltx-video-i2v in worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage $WORLDFOUNDRY_HFD_ROOT/Lightricks--LTX-Video with both ltxv-13b-0.9.8-distilled.safetensors and ltxv-spatial-upscaler-0.9.8.safetensors plus the T5 encoder/tokenizer shards. Missing weights are reported at runtime.

Recorded Hugging Face repos: Lightricks/LTX-Video (the 0.9.8 distilled files WorldFoundry actually opens), plus Lightricks/LTX-Video-0.9.5, Lightricks/LTX-Video-0.9.7-dev, and Lightricks/LTX-Video-0.9.8-13B-distilled. Every card is tagged license other. The GitHub tree is Apache-2.0 — review the HF card before commercial use.

Integration is recorded. Status: Static load only — checkpoint structure is validated, not a GPU artifact in this workspace.