Which nodes does a Hunyuan Video 2 graph need?

Four functional blocks, in order:

• Model loader — the GGUF loader for the checkpoint, plus a text encoder (CLIP) loader.
• Conditioning — positive and negative prompt nodes feeding a video conditioning combine node.
• Latent — a video latent node that sets width, height, and frame count.
• Sampler — the KSampler, then a VAE decode, then a video combine and save.

Wire them in that order. Keep the negative prompt generic: “blurry, watermark, oversaturated, extra limbs.” Overloading the negative side slows sampling without helping much.

What sampler settings reduce flicker?

Flicker — the shimmer between frames — is the number one complaint with local video models. Three settings tame it:

• Steps: 25-30. Below 20 the model under-resolves and flickers; above 35 you gain little and wait longer.
• CFG: 6-7. Higher values sharpen but increase temporal instability.
• Scheduler: a uniform or karras-style schedule tends to be more stable than exponential alternatives.

Also drop frame count before you drop resolution. A 49-frame 720p clip is more stable than a 97-frame one at the same VRAM.

How many frames can a 24GB card handle?

On a 4090 with a Q8 build, 49 frames at 720p is comfortable. Pushing to 97 frames risks a swap to system RAM, which slows everything to a crawl. If you need longer clips, render 49-frame segments and stitch them. Enable model offloading only if you are short on VRAM; it trades speed for memory.

Can I batch renders overnight?

Yes, and it is the main reason to run locally. Queue a folder of prompts, set the queue to run in order, and let it work while you sleep. A Q8 build on a 4090 averages a few minutes per 5-second clip, so a night yields a usable batch. Check the first output before walking away — a bad seed wastes hours.

When should I use a hosted tool instead?

If your GPU has 8GB or less, or you need audio baked in, a hosted model like Hunyuan Video 2 through a cloud endpoint — or a competitor like Pika 3 for stylized clips — will save you the wiring. The local graph pays off once you are generating dozens of clips and want repeatable settings.

Quick Summary

• A Hunyuan Video 2 ComfyUI graph needs loader, conditioning, video latent, and sampler blocks.
• 25-30 steps, CFG 6-7, and a karras schedule cut flicker.
• A 24GB card handles 49 frames at 720p comfortably; stitch for longer clips.
• Overnight batch queuing is the strongest reason to run it locally.
• Under 8GB VRAM or need native audio? Use a hosted service instead.

What is Hunyuan Video 2?

It is a diffusion transformer trained for video, released as open weights. That means you download the checkpoints, point a runner at them, and generate offline. Commercial use is permitted under Tencent’s community license, with the usual restrictions around certain jurisdictions. It handles text-to-video and supports image conditioning for animating a still.

How much VRAM do I need?

Full-precision inference is heavy — think 60GB+ for comfortable headroom, which realistically means an A100 or a multi-GPU setup. Most people use a quantized build instead:

• GGUF Q8: around 24GB VRAM, near-full quality.
• GGUF Q5/Q4: 16-12GB VRAM, some quality loss on fine detail.

A 24GB card like an RTX 4090 runs the Q8 build at usable speeds. Below 12GB, expect long renders and occasional out-of-memory crashes.

How do I set it up step by step?

1. Install Python 3.10 or 3.11 and a recent CUDA toolkit.
2. Create a virtual environment and install PyTorch matching your CUDA version.
3. Download the model weights from the Hugging Face repo (they are large — 30GB+ for the full set).
4. Clone the ComfyUI or the official inference repo.
5. Load the GGUF checkpoint via the GGUF loader node.
6. Set resolution to 720p to start; 1080p roughly doubles render time and VRAM.
7. Queue a 5-second test prompt and let it run.

First render after a cold start is slow because the model loads into VRAM. Subsequent prompts reuse it and are much faster.

Is local rendering actually cheaper than paying per clip?

Depends on volume and electricity. If you already own a 24GB GPU, the marginal cost per clip is basically power — a few cents. Hosted services charge per render, and at 100+ clips a month that adds up past the cost of a used 4090. But if you render five clips a month, paying a hosted tool is far cheaper than hardware. Local wins on volume and privacy, loses on convenience.

How does it compare to hosted models?

A hosted option like Seedance 1.5 beats Hunyuan Video 2 on out-of-the-box polish and native audio, and it needs no setup. Hunyuan Video 2 gives you weights you own, no per-render fee, and room to fine-tune. Pick it when control and volume matter more than the last 10% of visual quality.

What are the downsides?

Setup is not a one-click install — dependency conflicts between PyTorch, CUDA, and the runner are common and can eat an afternoon. Renders are slower than a cloud service unless you own top-tier hardware. And there is no built-in audio, so you add sound separately. If you want a clip in five minutes with zero config, this is the wrong tool.

Quick Summary

• Hunyuan Video 2 is Tencent’s open-weight text-to-video model on Hugging Face.
• Full precision needs 60GB+ VRAM; GGUF builds run on 12-24GB cards.
• Setup runs through ComfyUI with a GGUF loader and a compatible PyTorch/CUDA stack.
• Local rendering wins at high volume and for privacy, loses on convenience and setup time.
• No native audio — you add sound in post.