tencent / HunyuanVideo-1.5-t2v

Overview

HunyuanVideo-1.5-t2v is Tencent’s text-to-video model for generating short, high-quality video clips from written prompts. It uses a diffusion transformer backbone with a 3D causal VAE, so it can model motion across frames and keep details stable. Many deployments add an extra super-resolution stage, which can refine edges and textures at higher output presets. It’s useful when you need a fresh clip from scratch without filming, storyboarding, or stock footage hunting.

What you can build

• Product b-roll clips for ads, landing pages, and app store videos
• Cinematic establishing shots for concept trailers and pitch decks
• Stylized loops for music visuals and social posts
• Animated scenes for pre-visualization and story tests
• Background plates for motion design and compositing
• Bilingual content concepts, including in-video text attempts (results vary)

Inputs

• A text description of the scene you want, written in English or Chinese. Longer prompts with clear subjects and actions work best.
• An optional “avoid list” description in plain text to reduce unwanted traits (for example: blur, shaky camera, extra limbs).
• A resolution preset selection (commonly 480p, 720p, or 1080p). Higher presets may run a super-resolution pass.
• An aspect ratio preset that sets the frame shape (for example: 16:9, 9:16, 1:1, 4:3, 3:4, and other supported ratios).
• A step count for the diffusion sampling process. More steps usually improve detail and prompt match.
• A target clip length given as a total frame count. The frame count must follow a 4n+1 pattern (1, 5, 9, 13, 17, and so on).
• An optional numeric seed for repeatable results. Use the same seed to re-run a similar generation.

Outputs

• An MP4 video file of the generated clip at the chosen resolution and aspect ratio.
• The seed used for the run, which helps you reproduce or iterate on the same result.

Recommended settings

• Fast previews: pick 480p, keep the clip short, and use a moderate step count.
• Higher fidelity: pick 720p or 1080p and use the maximum step count allowed by the run.
• Cleaner motion: describe one main subject and one main action. Avoid conflicting camera moves.
• Reproducible iteration: set a fixed seed while you tune the prompt.

Limitations

• The frame count must follow the 4n+1 rule. Other values can fail or produce broken clips.
• Longer clips need more frames, which raises the chance of drift and identity changes.
• Fine text rendering inside the video is not guaranteed, even with explicit instructions.
• Very short, vague, or contradictory prompts often cause weak motion or unstable objects.
• High-resolution presets may rely on upscaling stages, which can amplify artifacts.
• Seeds improve repeatability, but exact matching can vary across deployments and updates.

Safety & compliance

• This model is released under the Tencent Hunyuan Community License Agreement, with territory limits that exclude the EU, the UK, and South Korea.
• Don’t use the model to bypass safety guardrails or to harm people.
• Don’t generate content that exploits or harms minors.
• Don’t create election manipulation content, targeted harassment, malware, or harmful personal data disclosure.
• If you post generated content in public contexts, you must clearly label it as machine-generated.
• Don’t use the model, its outputs, or results to improve a separate AI model outside permitted derivative terms.