HunyuanVideo is a strong open-weight video model you can run yourself, and ComfyUI is how most people run it. Tencent's model generates cinematic text-to-video and image-to-video clips. The November 2025 release of HunyuanVideo 1.5 brought it to a size that fits on a single consumer GPU.
The catch is VRAM. Even the lighter 1.5 model needs real memory at full quality, and many tutorials assume you already own a 24GB+ card. This guide walks the full path: what HunyuanVideo is, the official ComfyUI workflows, the exact model files you need, and the GPU that runs it without quantization headaches.
HunyuanVideo ranks among the best open-source AI video generation models for people who want to generate locally rather than through a paid API.
Takeaways
- HunyuanVideo 1.5 runs on consumer GPUs, needing about 14GB VRAM with offloading and 24GB comfortably.
- ComfyUI supports HunyuanVideo natively, with official text-to-video and image-to-video workflows and no custom nodes.
- Full FP16 720p generation needs a 48GB card such as the RTX A6000, avoiding offloading or quantization.
- A rented RTX A6000 costs $0.35/hr, under $0.10/clip with per-minute billing.
- Thunder Compute is the cheapest provider for the RTX A6000 and A100 80GB.
What Is the HunyuanVideo Model?
Developed by Tencent, HunyuanVideo is an open-source video generation model, first released in December 2024 as a 13B parameter system. It produces text-to-video and image-to-video clips with strong motion quality, prompt alignment, and frame-to-frame stability. At launch it was the largest open video model available, and it rivaled closed systems on output quality.
The version most people run today is HunyuanVideo 1.5, released in November 2025. It rebuilds the model at 8.3B parameters using a 3D diffusion transformer with a 3D causal VAE, and it unifies text-to-video and image-to-video in a single pipeline. Despite being smaller, it delivers quality comparable to the original while running far more efficiently.
HunyuanVideo 1.5 generates 5-10 second clips at 480p or 720p, with a built-in super-resolution step that upscales to 1080p. It runs at 24fps and handles diverse styles, including realistic, anime, and 3D, along with on-screen text rendering in English and Chinese. A faster attention mechanism called SSTA improves efficiency on longer clips, with Tencent reporting a 1.87x speedup for 10-second 720p generation.
HunyuanVideo in ComfyUI: What You Need
ComfyUI natively supports HunyuanVideo, so a standard setup needs no custom nodes. The ComfyUI team added native HunyuanVideo 1.5 support on November 24, 2025, shipping official workflows for both text-to-video and image-to-video. The original 13B model has been supported even longer, with image-to-video landing on day one back in March 2025.
Three things get you running: a recent ComfyUI build, the HunyuanVideo model files, and a GPU with enough VRAM. The model files include the diffusion model itself, the text encoders (Qwen2.5-VL and byT5), the VAE, and a CLIP vision model for image-to-video. ComfyUI's official workflow points you to each file and places it in the right folder.
Getting started is simple once ComfyUI is current. Update to the latest version, download the official workflow file, and drag it onto the canvas. ComfyUI detects any missing models and prompts you to download them automatically.
HunyuanVideo VRAM Requirements for ComfyUI
HunyuanVideo 1.5 needs about 14GB of VRAM at minimum with model offloading enabled, and ComfyUI's native support targets 24GB cards for comfortable headroom. At full FP16 precision, a 720p generation uses roughly 28-45GB depending on clip length, which pushes past a 24GB card and forces you into offloading or quantization. Running FP16 on a 48GB card or larger sidesteps that entirely and keeps generation fast.
The table below maps the common precision options to their approximate 720p memory footprint and the card that runs each comfortably.
| Precision | Approx. VRAM (720p) | Runs Comfortably On | Trade-off |
|---|---|---|---|
| FP16 (full) | ~28-45 GB | 48GB+ cards (RTX A6000, A100) | Best quality, no offloading |
| FP8 | ~18-20 GB | 24GB cards (RTX 4090) | Minimal quality loss, ~10-15% slower |
| GGUF Q4 (offloaded) | ~9-11 GB | 12-16GB cards | Runs on small cards, noticeably slower |
| 1 Figures are approximate and vary with resolution, frame count, and step count. Text encoder placement (on GPU vs. offloaded) shifts totals by several gigabytes. | |||
Which HunyuanVideo Build to Load
Your VRAM decides which build to load. The native FP16 workflow gives the best quality and the simplest setup when you have the memory for it. FP8 builds cut VRAM by around 40% with minimal quality loss, and GGUF quantized builds from community maintainer city96 push the model onto smaller cards at the cost of speed.
Most workflows no longer require a wrapper. Kijai's ComfyUI-HunyuanVideoWrapper is still useful for early access to experimental features, but the native implementation now covers standard text-to-video and image-to-video generation. Start with the native path unless you specifically need a feature that only exists in the wrapper.
Setting Up the HunyuanVideo ComfyUI Workflow
The official workflows are the fastest way to a working setup. Grab them from the ComfyUI HunyuanVideo 1.5 announcement or the ComfyUI templates library. From there the process differs slightly depending on whether you are generating from text or from an image.
Text-to-Video Workflow
Text-to-video is the default HunyuanVideo workflow and the simplest starting point. Load the T2V workflow file, confirm that the diffusion model, text encoders, and VAE are downloaded, and write a prompt in the text node. Set your resolution to 480p for quick tests or 720p for final quality, choose a frame count in the 5-10 second range, and queue the generation.
Step count is the main quality-versus-speed lever. More steps produce cleaner motion but take longer, so start around 20-30 steps and adjust from there. The distilled variant of 1.5 runs well at lower step counts, which is handy when you are iterating on prompts.
Image-to-Video Workflow in ComfyUI
HunyuanVideo image-to-video animates a still image into a moving clip, and it is built into the same ComfyUI integration. Load the I2V workflow, which adds a CLIP vision model and an image input node on top of the text-to-video graph. Drop in your source image, add a short prompt to guide the motion, and queue it the same way.
Image-to-video is the mode most creators reach for, since it gives precise control over the opening frame and composition. HunyuanVideo 1.5 follows the guide image closely while still producing natural camera moves and physics. That makes it a reliable choice for product shots, character intros, and short social clips.
Running HunyuanVideo in ComfyUI on a Cloud GPU
Renting a cloud GPU is the practical answer when your own hardware cannot run HunyuanVideo at the quality you want. Full FP16 at 720p exceeds a 24GB card, and quantization or offloading trades away speed and quality to fit. A rented 48GB or 80GB GPU removes that ceiling and runs the native FP16 workflow with room to spare.
Thunder Compute is built for exactly this kind of on-demand work. You launch a GPU instance in about 30 seconds from a VS Code or Cursor extension, or from the web console, then install ComfyUI and start generating. Instances bill per minute with persistent storage included, so you only pay for the time you spend rendering.
The standout option for HunyuanVideo is the RTX A6000. Its 48GB of VRAM fits the FP16 720p workflow with no offloading, and at $0.35/hr it is far cheaper than the enterprise cards most guides assume you need. If you want more headroom for longer clips or heavier batches, the A100 80GB steps up while remaining the cheapest A100 among the major providers.
What It Costs to Generate HunyuanVideo Clips
Generating a HunyuanVideo clip costs a few cents on the right GPU. A typical 720p clip finishes in a few minutes, and on an RTX A6000 at $0.35/hr that works out to well under $0.10/video. Per-minute billing matters here, because video generation is bursty and you avoid paying for a full rounded-up hour between renders.
The table below compares per-hour prices for the two GPUs best suited to HunyuanVideo across major providers, so you can see where the model is cheapest to run.
| Provider | RTX A6000 (48GB) | A100 80GB | Notes |
|---|---|---|---|
| Thunder Compute | $0.35 | $1.09 | Cheapest on both cards, per-minute billing |
| Hyperstack | $0.50 | $1.35 | Reserved and on-demand options |
| RunPod | $0.53 | $1.39 | Managed pods, template-based setup |
| Vast.ai | $0.461 | $1.731 | Marketplace median, varies by host |
| Lambda | $1.09 | $2.79 | A100 priced in 8-GPU instances |
| Paperspace | $1.89 | $3.18 | Notebook-friendly platform |
| 1 Prices per GPU-hour as of August 2026 and subject to change. Vast.ai figures are the median of verified US and Canada hosts and exclude bandwidth, which is billed separately. Check each provider for current pricing. | |||
The cost math is what makes local-style generation on a rented GPU so appealing. Some hosted HunyuanVideo APIs charge around $0.40/clip. Renting a GPU and running the open model yourself drops the marginal cost to pennies once you are set up, and that gap adds up quickly at volume.
HunyuanVideo Licensing and Commercial Use
HunyuanVideo is released under the Tencent Hunyuan Community License, which allows commercial use with some conditions. The main conditions are reported regional exclusions and large-scale usage limits. The license is permissive for most individual creators and small teams, but not unrestricted. Review the current license text on the HunyuanVideo 1.5 model card before you build a commercial product around it.
Many "free online" generator sites gloss over these licensing details. Running the open weights yourself makes clear which model version and license you are using, which matters when your output goes into paid work.
Last Thoughts on HunyuanVideo in ComfyUI
HunyuanVideo in ComfyUI gives you flagship-quality open video generation on hardware you rent by the minute. With the official workflow loaded and a 48GB card like the RTX A6000, you can turn a prompt or image into a 720p clip for pennies. Launch an instance on Thunder Compute, set up ComfyUI, and start generating without buying hardware or wrestling with quotas.
FAQ
Can You Run HunyuanVideo in ComfyUI?
Yes. ComfyUI has native support for HunyuanVideo, including HunyuanVideo 1.5, with official text-to-video and image-to-video workflows. Update ComfyUI to the latest version, load the workflow file, and download the prompted model files.
How Much VRAM Do You Need for HunyuanVideo in ComfyUI?
HunyuanVideo 1.5 runs at a minimum of about 14GB of VRAM with model offloading enabled, and 24GB is the comfortable target for the native workflow. Full FP16 at 720p uses roughly 28-45GB depending on clip length, so a 48GB card such as the RTX A6000 runs it cleanly without offloading or quantization.
Does HunyuanVideo Support Image-to-Video in ComfyUI?
Yes. HunyuanVideo 1.5 unifies text-to-video and image-to-video in one model, and ComfyUI ships an official image-to-video workflow. It adds a CLIP vision model and an image input to the standard graph so you can animate a still image with a guiding prompt.
What Is the Difference Between HunyuanVideo and HunyuanVideo 1.5?
The original HunyuanVideo is a 13B model from December 2024 that needs high-end data center GPUs. HunyuanVideo 1.5, released in November 2025, is an 8.3B rebuild that runs on consumer and prosumer GPUs, adds native image-to-video, and generates more efficiently thanks to its SSTA attention mechanism.
Is HunyuanVideo Free to Use?
The model weights are open and free to download, and running them yourself only costs the compute you use. Commercial use falls under the Tencent Hunyuan Community License, which carries some conditions, so check the license for your specific use case.
How Much Does It Cost to Generate a Video With HunyuanVideo?
A HunyuanVideo clip costs under $0.10 on an RTX A6000 at $0.35/hr. A typical 720p clip takes a few minutes to render, and per-minute billing means you pay only for the compute you use rather than a full hour.