HunyuanVideo 1.5: which checkpoint to download (480p, 720p, SR)

HunyuanVideo 1.5's README lists 13 model names; at the read date 11 have download links and 2 (720P T2V cfg-distill and 720P T2V sparse-cfg-distill) say coming soon. For a first run, choose by the job: a 480p image-to-video draft starts with the step-distilled file (8 or 12 steps), 720p work with a CFG-distilled or sparse I2V variant, and 1080p comes from a super-resolution pass. The minimum GPU memory is 14 GB, and only with model offloading enabled.

What are the checkpoints?

How should you choose?

Start from the input you have and the size you need. The README states the default is 50 inference steps and that the best configuration varies by variant (8 to 50). The step-distilled 480p I2V model is the one with a published speed claim: about 75 percent faster end to end on an RTX 4090, within 75 seconds. I found no such claim for the other files, so time them on your card.

• No input image: pick a T2V file.
• Have a first frame: pick an I2V file at the same resolution.
• Need speed: cfg-distill, then step-distill at 480p.
• Need 1080p: run the super-resolution model after a 720p generation.

What about disk and download order?

Download one file first, not every variant. A 480p I2V step-distilled checkpoint is the quickest way to prove your install works, because it runs in 8 or 12 steps. Add the 720p model once the pipeline is stable, and the super-resolution model only if you need 1080p. The README lists the variants but not their sizes, so check each file's size on the model page before you fetch it.

Keep a note of which variant made which clip. The names carry the resolution and the distillation, which is enough to reproduce a run.

Is it on Sume?

No. The Sume catalog code (read 2026-10-09) has no HunyuanVideo row. For hosted image-to-video, frame_images takes a first or last frame; the listed alternatives and their limits are on GET /v1/videos/models and in the video docs. The model card's license tag is tencent-hunyuan-community; read the license file before shipping output.