HunyuanVideo 1.5 · Image to Video

HunyuanVideo 1.5 is one of the few models in this pack you could also run yourself. Tencent rebuilt the 13B original down to 8.3B with a sparse-attention architecture and a super-resolution pass, and the official floor is 14 GB with model offloading. The community lands it at 10 to 14 GB.

So this node is asking a straightforward question: do you have 14 GB and an evening, or 18 credits and two minutes? At 3.6 credits per second, a 5-second clip is 18 credits - eighteen cents.

The KB's HunyuanVideo panel is honest about where the model sits: it got there first in December 2024, Wan took the ecosystem in early 2025, and by 2026 Hunyuan is a distant third in local usage. That's precisely the argument for calling it hosted. A model you'd struggle to justify downloading is a model you're happy to rent for eighteen cents.

What it does

tencent/hunyuan-video-1.5/image-to-video on PopNow's servers, wrapped as a ComfyUI node. Your opening frame goes up, a VIDEO comes back as a normal value you can Save Video or push through more nodes. No GPU, no weights, no offloading dance.

It keeps the aspect ratio of your source image. That's the one compositional guarantee you get, and it's the right one: you chose the framing when you made the still.

The inputs

image_url is the opening frame. Everything about composition is decided here.

prompt is optional, which is unusual for a video node and worth using. The tooltip says leaving it empty lets the model animate the frame on its own. That's not a degraded mode - on a clean subject shot it's often better than the motion description you were about to write, because you're not fighting a vague instruction with a vaguer one. Write a prompt when you want something the frame doesn't imply: "the camera pulls back slowly, her hair moves in the wind, overcast daylight".

resolution gives 720p (default) or 480p. Two tiers, no 1080p. The 480p tier is the draft tier - cheap per second and fast enough to audition motion before you commit.

duration_seconds is a dropdown with 5 or 8 in it, and that's it. Not a slider, not any integer - those are the two lengths the model serves. Billed per second, so 8 costs noticeably more than 5 and the difference is usually not worth it for a camera move that lands in the first three seconds anyway.

seed starts fixed, so re-queueing the same graph gives you the same clip instead of a second charge. Randomize control-after-generate to explore - each roll is billed.

Outputs: video and task id.

Install

Manager → search PopNow → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then a key from popnow.ai/api/console/keys into Settings > PopNow, or in the environment:

POPNOW_API_KEY=sk-popnow-... python main.py

Zero dependencies and zero model downloads - the pack only uses what ComfyUI already ships. Restart again after adding the key so the node list is built from your account's catalog.

Where people get burned

The local alternative is genuinely there. If you have the VRAM, the open weights are free and unlimited, and you'll iterate faster without a round trip. The Wan video essay is the better read if you're choosing a local video base - Hunyuan lost that fight for most workloads.

Eight seconds is not a scene. Both lengths are short; plan a cut, not a shot that develops. Anything with a beginning, middle and end needs two runs and an edit.

Uploads are capped. Images are re-encoded to PNG, or JPEG if the PNG is over 10 MiB. A 6000px still gets squeezed before it ever reaches the model - downscale to something sane first and you'll get the same clip without the wait.

Cost adds up in seconds, not calls. A dozen 8-second takes at 720p is a couple of dollars, which is nothing - until you do it for three shots and two of them need to be re-done. The node quotes each run before anything is held, and Max credits per run in Settings > PopNow is the seatbelt. Cancelling, as ever, stops the wait and not the charge.