HunyuanVideo-1.5

One standalone TPU inference script, generate_hunyuan_video_1_5.py, covers both T2V and I2V for a given resolution (pass --image_path for I2V, omit it for T2V). Covers the 4 core (non-distilled, non-sparse-attention, non-super-resolution) checkpoint variants: 480p/720p × T2V/I2V.

Script	Model	Params	Task	Checkpoint variant
generate_hunyuan_video_1_5.py	HunyuanVideo-1.5	8.3B DiT	T2V	transformer/480p_t2v/
generate_hunyuan_video_1_5.py	HunyuanVideo-1.5	8.3B DiT	I2V	transformer/480p_i2v/
generate_hunyuan_video_1_5.py	HunyuanVideo-1.5	8.3B DiT	T2V	transformer/720p_t2v/
generate_hunyuan_video_1_5.py	HunyuanVideo-1.5	8.3B DiT	I2V	transformer/720p_i2v/

Checkpoints

Conditioning requires three separate text/vision towers on top of the DiT itself: a Qwen2.5-VL-7B-Instruct text-only tower, a byT5-small (Glyph-SDXL-v2) glyph/color encoder, and (I2V only) a SigLIP vision encoder. --checkpoint_dir should point at tencent/HunyuanVideo-1.5's downloaded root (transformer/, vae/, text_encoder/{llm,byt5-small,Glyph-SDXL-v2}/); --siglip_checkpoint_dir (I2V only) at black-forest-labs/FLUX.1-Redux-dev's downloaded root — a gated repo, request access first.

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 .

Text-to-video

python examples/generate_hunyuan_video_1_5.py \
  --checkpoint_dir "./checkpoints/HunyuanVideo-1.5" \
  --resolution 480p \
  --prompt "A golden retriever running on a beach at sunset, cinematic, high detail" \
  --height 480 --width 832 --num_frames 121 --num_steps 50 \
  --output_path "out/output_hunyuan_1_5_t2v.mp4"

Image-to-video

python examples/generate_hunyuan_video_1_5.py \
  --checkpoint_dir "./checkpoints/HunyuanVideo-1.5" \
  --siglip_checkpoint_dir "./checkpoints/FLUX.1-Redux-dev" \
  --resolution 480p \
  --image_path "./assets/dog.jpg" \
  --prompt "The dog starts running toward the camera" \
  --num_frames 121 --num_steps 50 \
  --tensor_parallel_size 4 \
  --output_path "out/output_hunyuan_1_5_i2v.mp4"

Omit --height/--width for I2V and the output resolution is derived from the conditioning image's own aspect ratio — a portrait image produces a portrait video, not a squished landscape one.

Tensor parallelism

--tensor_parallel_size (default: every local device) Megatron-shards the DiT's double/single-stream Q/K/V/output/FFN Dense layers across chips; the rest (VAE, Qwen2.5-VL, byT5, SigLIP) is simply replicated across the same mesh. Must divide heads_num (16). Required in practice — the 8.3B DiT's bf16 weights (~16.6 GB) don't fit replicated alongside the other components on a single TPU v4 chip. There is no --sequence_parallel_size or weight offloading for this family yet; VAE decode has its own spatial tiling (--vae_tile_latent_size, independent of TP).

CLI reference

Flag	Default	Notes
--checkpoint_dir	required	tencent/HunyuanVideo-1.5's downloaded root.
--siglip_checkpoint_dir	None	black-forest-labs/FLUX.1-Redux-dev's downloaded root. Required for I2V.
--resolution	480p	480p or 720p — selects the checkpoint variant and the default --shift.
--image_path	None	Conditioning image, for I2V. Omit for T2V.
--height / --width	480 / 832	Must be divisible by 16.
--num_frames	121	Wants 1 + 4k for an exact VAE round-trip.
--num_steps	50	Flow-matching Euler sampling steps.
--shift	resolution default	480p: 5.0 both tasks; 720p: 9.0 T2V, 7.0 I2V.
--guidance_scale	6.0	Real CFG — these checkpoints have no embedded/distilled guidance path.
--dit_dtype	bfloat16	Cast target for the DiT's weights (checkpoints ship as float32).
--tensor_parallel_size	every local device	Megatron-shards the DiT's double/single-stream Dense layers; the other components stay replicated. Must divide heads_num (16). See Tensor parallelism above.
--vae_tile_latent_size	16	Latent-space spatial tile size for the tiled VAE decode. Shrink (e.g. 8) if VAE decode OOMs — more likely at tp > 1, where the replicated components leave less per-chip headroom.
--fps	24	Output video frame rate.

Scope

• No --sequence_parallel_size or weight offloading yet. Tensor parallelism (above) is enough to fit the 4 core checkpoints at 480p and 720p on a v4-8; a larger frame count or batch would need one of those additions. See Sharding, parallelism, and JIT on TPUs.
• No distilled/step-distilled/sparse-attention or super-resolution checkpoint variants.
• VAE decode is spatially tiled but never temporally tiled (matching the reference VAE, which has no temporal-tiling path).

Status

Verified end-to-end against real checkpoints, all 4 core variants. 5-run averages on a TPU v4-8 (see the Benchmark Explorer for the full, current set):

Variant	Task	Resolution	Frames	Steps	s/step	Peak HBM/chip (GB)
8.3B (480p)	T2V	832×480	121	30	119.5	29.1
8.3B (480p)	I2V	544×720	121	30	112.9	31.4
8.3B (720p)	T2V	1280×720	121	30	221.0	30.3
8.3B (720p)	I2V	832×1104	121	30	219.2	32.0

All rows use tp=4. This architecture's joint global self-attention over the full image+text token sequence (no windowing) makes it the slowest per-step model in the benchmark table.

---

See Understanding video diffusion architectures on the blog for how HunyuanVideo-1.5's token-concatenation conditioning (rather than cross-attention) compares to the other model families. See HunyuanVideo (1.0) for the 13B model in the same block family, and the Benchmark Explorer for measured numbers.