LTX-2.5

One standalone TPU inference script, generate_ltx2_5.py, covers both T2V and I2V for the two released 22B checkpoints (dev, distilled). Scope is deliberately narrower than LTX-2.5's full reference: video-only (no audio generation), single-stage (no LatentUpsampler two-pass refinement).

Script	Model	Params	Task	Checkpoint file example
generate_ltx2_5.py	LTX-2.5	22B (dev)	T2V, I2V	ltx-2.5-22b-dev-transformer-bf16.safetensors
generate_ltx2_5.py	LTX-2.5	22B (distilled)	T2V, I2V	ltx-2.5-22b-distilled-transformer-bf16.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

Three separate checkpoint files from Lightricks/LTX-2.5:

• --dit_checkpoint_path: diffusion_models/ltx-2.5-22b-{dev,distilled}-transformer-bf16.safetensors — bundles the DiT and the embeddings connector.
• --vae_checkpoint_path: vae/ltx-2.5-video-vae-conv-bf16.safetensors (default --vae_variant conv) or vae/ltx-2.5-video-vae-bf16.safetensors (--vae_variant diffusion, a transformer/neighborhood-attention decoder — the official demos' decoder).
• --text_encoder_checkpoint_path: text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors — bundles Gemma-4, its embedded tokenizer, and the feature-extraction projection.

Text-to-video

python examples/generate_ltx2_5.py \
  --dit_checkpoint_path "./checkpoints/LTX-2.5/diffusion_models/ltx-2.5-22b-dev-transformer-bf16.safetensors" \
  --vae_checkpoint_path "./checkpoints/LTX-2.5/vae/ltx-2.5-video-vae-conv-bf16.safetensors" \
  --text_encoder_checkpoint_path "./checkpoints/LTX-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors" \
  --prompt "A majestic red panda climbing a bamboo tree in the snow, 4k" \
  --sampler dev \
  --tensor_parallel_size 4 \
  --offload_dit_weights --offload_chunk_size 8 \
  --output_path "out/output_ltx2_5_t2v.mp4"

--sampler distilled (the default) runs 8-step ancestral-Euler sampling with no CFG; --sampler dev runs 30-step plain-Euler with real CFG (--guidance_scale 3.0) plus a guidance-rescale correction (--guidance_rescale 0.7) that avoids CFG's characteristic washed-out over-saturation.

Image-to-video

python examples/generate_ltx2_5.py \
  --dit_checkpoint_path "./checkpoints/LTX-2.5/diffusion_models/ltx-2.5-22b-dev-transformer-bf16.safetensors" \
  --vae_checkpoint_path "./checkpoints/LTX-2.5/vae/ltx-2.5-video-vae-conv-bf16.safetensors" \
  --text_encoder_checkpoint_path "./checkpoints/LTX-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors" \
  --image_path "./examples/assets/cat.jpg" \
  --prompt "The cat looks up at the camera" \
  --sampler dev \
  --tensor_parallel_size 4 \
  --offload_dit_weights --offload_chunk_size 8 \
  --output_path "out/output_ltx2_5_i2v.mp4"

Conditioning works differently from LTX-Video's I2V: instead of clamping a per-token timestep, LTX-2.5 threads an explicit per-token "denoise mask" through the DiT's own timestep input, the sampler's x0 estimate, and its stepped output — --conditioning_strength (default 1.0) controls how strongly the conditioning frame is enforced.

Tensor parallelism

Unlike LTX-Video (where tp=1's weights fit and only activations force higher tp), --tensor_parallel_size 4 (the default) is required here just for the 22B DiT's and 12B Gemma-4's own weights to fit at all.

CLI reference

Flag	Default	Notes
--vae_variant	conv	conv (ResNet + pixel-shuffle-upsample) or diffusion (NATTEN-based, the official demos' decoder — noticeably slower, see Status).
--tensor_parallel_size	4	Must divide num_devices, the DiT's num_attention_heads (32), and Gemma-4's num_attention_heads (16).
--conditioning_strength	1.0	I2V only.
--dit_dtype	bfloat16	Every released checkpoint's scale_shift_table/prompt_scale_shift_table (AdaLN modulation tables) ship in float32 and are always kept at float32 regardless of this flag — downcasting them was a measurable quality bug.
--sampler	distilled	distilled (8-step ancestral-Euler, no CFG) or dev (30-step plain-Euler, real CFG + guidance-rescale).
--guidance_scale / --guidance_rescale	sampler default	Defaults to 1.0/0.0 for distilled, 3.0/0.7 for dev.
--height / --width	704 / 1216	Must be divisible by 32.
--num_frames	121	Wants 1 + 8k for an exact VAE round-trip.
--offload_dit_weights	off	Needed at the reference's own resolution — not because the DiT's weights don't fit (they do, comfortably), but because a fully fused 48-block forward pass doesn't free per-block activations across blocks. See Weight offloading.
--offload_chunk_size	1	8 is the largest value confirmed to fit at the reference resolution (tp=4, both checkpoints).

Status

Verified end-to-end against real checkpoints, both dev/distilled recipes and both VAE decoder variants, at the reference's own 1216×704/121-frame resolution:

Variant	VAE decoder	Steps	Compile (s)	Generation (s)	Peak HBM/chip (GB)
dev	conv	30	87.7	217.9	16.7
distilled	conv	8	87.9	37.3	15.3
dev	diffusion	30	479.5	2859.6	16.1
distilled	diffusion	8	478.2	2680.8	14.8

The diffusion VAE variant is a genuinely different, slower decoder (its own internal denoising pass, tensor-parallel-sharded across the same mesh) — reach for it only when matching the official demos' exact output characteristics matters more than speed; conv is the faster default for most use. See the Benchmark Explorer for the full, current set of measured rows.

---

See LTX-Video for the earlier, smaller, unrelated LTX model, or Understanding video diffusion architectures and Weight offloading on the blog for background.