HunyuanVideo 1.5 · Text to Video

Tencent's HunyuanVideo got there first and then lost the room. The December 2024 version made local video generation feel real, and then Wan showed up, took image-to-video, and by mid-2026 the standing answer to "should I run Hunyuan?" was a shrug - the model family's own mentions fell to single digits a month. So here's the awkward pitch for this node: the easiest way to find out what HunyuanVideo 1.5 looks like is to not install it at all.

What you're actually renting

HunyuanVideo 1.5 is the November 2025 rebuild: 8.3B parameters instead of 13B, a sparse-attention trick (Selective and Sliding Tile Attention) to skip redundant work, and a small super-resolution pass on the end. Tencent's own floor is 14GB with offloading, so it's not unreasonable to run locally - the problem is that it's a third choice locally, not a first one.

This node runs the same model on PopNow's servers. You get a tencent/hunyuan-video-1.5/text-to-video request, a real VIDEO object back, and a bill. No checkpoint download, no VAE wrangling, no 8GB of attention internals fighting your text encoder for space. If you want to know whether Hunyuan's look suits a project, this costs cents and twenty minutes rather than an evening of downloads.

How the request is built

The pack isn't a hand-written wrapper per model. At startup it pulls the PopNow catalog and generates one node class per model straight from that model's published input schema, which is why this node has exactly the fields the API accepts and nothing else. Whatever you set that differs from the model's default goes into the request; anything left at the default is left out, so the model fills it in. Files you wire in get uploaded once, keyed by content hash, and reused for twelve hours.

The result comes back as a genuine ComfyUI VIDEO - the pack decodes the returned MP4 into ComfyUI's video type - so Save Video works, and so does feeding it into another node's video input.

The inputs that matter

prompt is the whole job. The author's own short version: describe the scene, the action and the camera. Hunyuan responds to motion words more than style adjectives.

negative_prompt is the one genuinely interesting field here. None of the Vidu nodes in this pack have one - this node does, and it works. Blur, extra limbs, text, watermark, the usual. Keep it to a handful of terms.

Then the shape decisions: aspect_ratio is 16:9 or 9:16 and nothing else, resolution is 480p or 720p, and duration_seconds is 5 or 8. That's the whole menu, and it matches the model - 1.5 was built for short clips.

seed starts fixed, which means queueing the same graph twice returns the same clip instead of billing you twice. Flip control-after-generate to randomize when you actually want a variant, because each run is charged.

Outputs: video, plus task id - a STRING that identifies the run in PopNow's API console, and the thing support will ask for.

Install

Manager, or the registry:

comfy node install popnow-comfyui

Or clone it into custom_nodes:

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

There are no model downloads and nothing extra to pip install - the pack only uses packages ComfyUI already ships, and the audio/video decoding goes through PyAV, which ComfyUI already bundles. What it does need is a PopNow API key: create one at popnow.ai/api/console/keys, then either set POPNOW_API_KEY in the environment or paste it into Settings > PopNow. Restart twice the first time - the node list is rebuilt from your account's catalog at startup, so the first boot after adding a key can still look incomplete. If you previously used the single-file custom_nodes/popnow_api.py from popnow.ai's docs, delete it; this pack's PopNow Generate node replaces it.

Where people get burned

The price depends on the resolution pair you pick, and the badge says "from". The cheap tier is the 480p one, and the 720p rate is roughly double per second. An 8-second 720p clip is therefore around four times the entry price - set Settings > PopNow > Max credits per run before you experiment, and any run quoted above the ceiling stops before credits are held.

Cancelling does not stop the job. Interrupting the node stops the wait; the job keeps running on PopNow, and if it succeeds you're charged. The result lands in the API console logs instead of your node.

Content rejections and hard failures are refunded, so a content_rejected isn't money down the drain - but a 40004 ("the model cannot serve that combination of settings") means you asked for something the model doesn't do, usually a duration or ratio it doesn't support. Change the field the message names.

And if the model isn't in your node list: restart. The list is your account's catalog as of startup, not a hardcoded menu.