HunyuanVideo 1.5 super-resolution to 1080p vs hosted 1080p

HunyuanVideo 1.5 generates at 480p or 720p, up to 121 frames, and a separate few-step super-resolution network upscales the result to 1080p. It is an 8.3B-parameter open model you run yourself. Sume does not list a Hunyuan model, but it does bill hosted 1080p video by the second, so you can price the two routes.

Specs are from the model card and the repository, read on 2026-10-02.

What does HunyuanVideo 1.5 produce natively?

The card describes a diffusion transformer with a 3D causal VAE that compresses 16 times spatially and 4 times in time. It does text-to-video and image-to-video, with bilingual prompts. The repository says exports are 24 fps and the default is 50 steps for 121 frames, which is about five seconds. Step-distilled variants work with 8 or 12 steps.

How does the 1080p step work?

The card says an efficient few-step super-resolution network upscales outputs to 1080p, sharpening and correcting distortions. It is a second model run after generation, not a native 1080p mode. As with any generative upscale, check faces and text against the 720p original.

The card gives a 14 GB minimum GPU memory with offloading, and the repository reports about 75 seconds end to end for the step-distilled 480p image-to-video model on an RTX 4090. Neither number covers the super-resolution step, so measure your own full pipeline.

What would 1080p cost on Sume?

Sume's repo pricing table carries Wan 3.0 at provider list $0.20 per second for 1080p, and the docs say Sume bills list times 1.25, so $0.25 per second. A 5-second 1080p clip is $1.25, and a 5-second 480p clip is about $0.31 at $0.0625 per second. wan-3.0 accepts 2 to 30 seconds per the docs. It is a different model from Hunyuan, so the look will differ.

The comparison is not only price. Self-hosting buys a licence you can read and a model you control, at the cost of GPU hours, queueing and your own failures. A hosted job buys no setup and one invoice line, but the model is fixed to the catalog. The hosted or weights post walks through that choice.

Which should you pick?

If you already own a 24 GB-class card and want to iterate on prompts without a meter, run Hunyuan with the distilled model and decide on 1080p only for keepers. If you need predictable clips on a deadline, or longer than five seconds, a hosted route is simpler. Read the community licence in the repository for your territory and size before you ship outputs. Check GET /v1/videos/models in Video generation for the live catalog before you assume anything is listed.