hunyuan

This node runs Hunyuan Video - Tencent's open video model - on Civitai's cloud, inside your ComfyUI graph. It's the rare recipe node in this pack that isn't a closed black box: Hunyuan has open weights, and this node exposes a model socket and a loras socket, which means you can point it at a specific Hunyuan checkpoint and LoRA stack from Civitai and have it run on their fleet.

There's some history worth knowing before you use it. Hunyuan got there first in late 2024 - the original 13B was the model that made local video generation feel real, famously uncensored out of the box - and then Wan showed up and took the ecosystem. By 2026 the KB's hunyuan-video panel describes local usage as collapsed, with Tencent's newer 8.3B 1.5 model doing the heavy lifting. For you that mostly means: if you specifically want Hunyuan's cinematic look, or you have a Hunyuan LoRA you like, this cloud route is a way to use it without the 45–60GB VRAM the original demanded. If you just want good video from Civitai, the other engines in the menu might serve you better.

How it works

The node submits a videoGen workflow with the Hunyuan engine to the Civitai Orchestration API and long-polls until the clip is done, then downloads the MP4 back as a native VIDEO. The interesting part is the model plumbing. model is a CIVITAI_AIR socket and loras is CIVITAI_LORAS - you fill them by wiring in a Civitai Model Selector and Civitai LoRA Selector node, which have the Browse Civitai buttons. Because the job runs on Civitai's fleet, the checkpoint is loaded server-side; nothing downloads to your machine.

Inputs and outputs that matter

• prompt (required) - the clip.
• cfg_scale (required) - default 4, range 0–100. This is the classic guidance scale; push it up for stricter prompt adherence, down for more model freedom.
• width / height (required) - default 0 (engine default). Set your output resolution explicitly.
• frame_rate - default 25.
• duration - 1–30 seconds, default 5.
• steps - default 20, range 10–50. More steps, more quality, more time (and Buzz) spent.
• seed - 0 for random.
• model - CIVITAI_AIR socket; wire a Model Selector to pick which Hunyuan checkpoint runs.
• loras - CIVITAI_LORAS socket; wire a LoRA Selector to stack LoRAs.
• api_config - optional auth override.

Outputs: video, workflow_id, raw_json.

Installing it

Standard pack install: ComfyUI Manager → Custom Nodes Manager → "Civitai Comfy Nodes" (publisher civitai) → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Only requests; no local model downloads (the fleet downloads the checkpoint).

Common issues

• Need auth + Buzz. Connect the Civitai Auth node, set CIVITAI_API_TOKEN (headless), or use the sidebar connect panel.
• The model socket is the trap. Recipe nodes have no Browse button - you must wire a Civitai Model Selector's air output into model. An unwired model socket may silently fall back to an engine default, so if you specifically wanted a custom checkpoint, double-check the selector is connected.
• Hunyuan is heavy on the fleet too. The 13B-class models are big; jobs can sit in queue. workflow_id lets you track them in your Civitai generation history.
• Early preview. The pack is under active development - behavior may change without notice. Pin versions.
• Cost. Cloud video generation bills per run. Watch the per-run cost shown on the node.