LTX-Video (0.9.8)

One standalone TPU inference script, generate_ltx_video.py, covers both T2V and I2V (pass --image_path for I2V, omit it for T2V) across all three released 0.9.8 sizes. Architecturally unrelated to the Wan/Cosmos scripts — a different VAE, a different (plain T5-XXL) text encoder, and a different RoPE convention — see Understanding video diffusion architectures for how it compares to the rest of vidax's model families.

Script	Model	Params	Task	Checkpoint file example
generate_ltx_video.py	LTX-Video 0.9.8	2B (distilled)	T2V, I2V	ltxv-2b-0.9.8-distilled.safetensors
generate_ltx_video.py	LTX-Video 0.9.8	13B (dev)	T2V, I2V	ltxv-13b-0.9.8-dev.safetensors
generate_ltx_video.py	LTX-Video 0.9.8	13B (distilled)	T2V, I2V	ltxv-13b-0.9.8-distilled.safetensors

Everything it needs (torch, transformers, pillow) is installed by default. On a Cloud TPU VM also add the tpu extra:

pip install -e ".[tpu]"    # or just: pip install -e .

Checkpoints

Each variant's DiT + VAE ship together as one flat .safetensors file from Lightricks/LTX-Video — pass its path as --checkpoint_path. The text encoder is a separate download, PixArt-alpha/PixArt-XL-2-1024-MS's text_encoder/tokenizer subfolders (a plain T5-XXL, shared by all three sizes) — pass text_encoder/model.safetensors.index.json's path as --t5_checkpoint_path.

Text-to-video

python examples/generate_ltx_video.py \
  --checkpoint_path "./checkpoints/LTX-Video-0.9.8-2B-distilled/ltxv-2b-0.9.8-distilled.safetensors" \
  --t5_checkpoint_path "./checkpoints/PixArt-XL-2-1024-MS/text_encoder/model.safetensors.index.json" \
  --prompt "A majestic red panda climbing a bamboo tree in the snow, 4k" \
  --num_steps 8 --guidance_scale 1.0 \
  --tensor_parallel_size 4 \
  --output_path "out/output_ltx_t2v.mp4"

--num_steps 8 --guidance_scale 1.0 matches the distilled checkpoints' own recipe (few steps, no classifier-free guidance). For 13B-dev, use real CFG and more steps instead: --num_steps 30 --guidance_scale 3.0.

Image-to-video

python examples/generate_ltx_video.py \
  --checkpoint_path "./checkpoints/LTX-Video-0.9.8-2B-distilled/ltxv-2b-0.9.8-distilled.safetensors" \
  --t5_checkpoint_path "./checkpoints/PixArt-XL-2-1024-MS/text_encoder/model.safetensors.index.json" \
  --image_path "./examples/assets/cat.jpg" \
  --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard." \
  --num_steps 8 --guidance_scale 1.0 \
  --tensor_parallel_size 4 \
  --output_path "out/output_ltx_i2v.mp4"

The conditioning image is resized to exactly --heightx--width (no aspect-ratio-preserving crop), VAE-encoded, and blended into the first latent frame at --conditioning_strength (default 1.0 — the first frame is the encoded image).

Tensor parallelism

--tensor_parallel_size (default 1) Megatron-shards both the DiT's and the T5 encoder's weights. Needed for the 13B checkpoints regardless of resolution (their bf16 weights alone don't fit replicated on a single TPU v4 chip's HBM) — and, less obviously, needed for 2B too at the reference's full 1216x704/121-frame resolution, where the self-attention activations alone don't fit at tp=1. There is no --sequence_parallel_size yet for this model family.

CLI reference

Flag	Default	Notes
--checkpoint_path	required	The flat .safetensors file bundling both the DiT and VAE.
--t5_checkpoint_path	required	PixArt-XL-2-1024-MS/text_encoder/model.safetensors.index.json.
--image_path	None	Conditioning image, for I2V. Omit for T2V.
--conditioning_strength	1.0	I2V only: how strongly the conditioning image is enforced.
--guidance_scale	3.0	CFG scale. Use 1.0 for the distilled checkpoints.
--dit_dtype	bfloat16	Every released checkpoint ships natively as bf16 — no fp32-weights requirement here, unlike Wan2.1.
--tensor_parallel_size	1	See Tensor parallelism above. Must divide num_devices, LTXDiT.num_attention_heads (32), and the T5 encoder's num_heads (64).
--num_steps	30	Use far fewer (e.g. 8) for distilled checkpoints.
--sampler	LinearQuadratic	Uniform | LinearQuadratic | Constant. Every released checkpoint's own embedded config uses LinearQuadratic.
--height / --width	512 / 768	Must be divisible by 32 (the VAE's spatial downscale factor).
--num_frames	97	Wants 1 + 8k for an exact VAE round-trip; the reference's own default is 121.
--fps	24	Output video frame rate.

Scope

Not implemented in this first port: multi-scale two-pass generation (the reference's default pipeline runs a low-res pass, an upsampler, then a high-res pass — this port always runs single-scale, single-pass), STG (spatio-temporal guidance) and cfg_star_rescale (plain CFG only), sequence parallelism, and V2V.

Status

Verified end-to-end against real checkpoints, all three sizes, T2V:

Variant	Resolution	Frames	Steps	Compile (s)	Generation (s)	Peak HBM/chip (GB)
13B (dev)	1216×704	121	30	134.7	156.0	15.3
13B (distilled)	1216×704	121	8	136.4	104.2	15.3
2B (distilled)	1216×704	121	8	83.5	47.3	8.8

See the Benchmark Explorer for the full, current set of measured rows.

---

See LTX-2.5 for the newer, larger, architecturally-unrelated LTX model, or Understanding video diffusion architectures on the blog for background on how these models differ.