tencent / HunyuanVideo-1.5-i2v

Overview

HunyuanVideo-1.5-i2v is Tencent’s image-to-video model from the HunyuanVideo-1.5 family. It starts from a single still image and uses diffusion to synthesize a short video that keeps the source identity and layout while adding motion. Under the hood, it pairs a Diffusion Transformer with a 3D causal VAE for spatiotemporal latents, plus selective attention optimizations to make longer clips more practical. It can also run a separate super-resolution stage to upscale results, so you can draft fast at lower resolution and finish in higher quality.

What you can build

• Animate a product photo into a short marketing loop with subtle camera motion
• Turn concept art or an illustration into a moving shot while keeping the original style
• Create establishing shots from a single keyframe for storyboards and animatics
• Generate motion tests for VFX planning (smoke, water, cloth, crowds) from one frame
• Make social clips that start from a branded hero image and evolve with prompt guidance

Inputs

• A single reference image used as the first frame of the clip. Upload a JPG, JPEG, PNG, WebP, GIF, or AVIF image.
• A text prompt that describes the scene, motion, camera movement, and visual style you want in the generated video.
• An optional negative prompt that lists elements you don’t want (for example, extra limbs, warped faces, text artifacts, logos, or heavy motion blur).
• A resolution preset that picks the base generator and (when available) an upscaling path. Typical presets include 480p for speed, 720p for best balance, and 1080p via super-resolution.
• A diffusion step count that controls the quality–cost tradeoff. Higher step counts usually improve detail and stability.
• A target clip length provided as a total frame count. The frame count must follow a 4n+1 pattern (for example 1, 5, 9, 13, 121).
• An optional seed value to reproduce a result. Using the same inputs and seed should keep outputs more consistent across runs.

Outputs

• An MP4 video clip that begins from your reference image and evolves with prompt-driven motion.
• The response also includes the seed used for the generation, which helps you rerun or iterate while keeping randomness controlled.
• Frame rate is typically 24 fps, so 121 frames is about 5.0 seconds of video.

Recommended settings

• Balanced default: 121 frames at 24 fps and around 28–30 diffusion steps.
• Highest quality: use 50 diffusion steps, especially for 720p generation.
• Fast previews: use 480p and fewer steps for iteration, then rerun at 720p for final.
• Longer clips: increase the frame count, but keep the required 4n+1 frame pattern.

Limitations

• Clip length is constrained by a fixed frame-count rule (4n+1). Invalid counts will fail or produce unstable results.
• 1080p output typically relies on a separate super-resolution pass rather than native 1080p generation.
• Fine text, small logos, and dense UI elements can degrade or drift across frames.
• Low-quality inputs (heavy JPEG artifacts, tiny images, screenshots with compression noise, or strong sharpening) can cause hallucinated details and flicker.
• Large viewpoint changes can reduce identity consistency, even when the first frame is stable.
• The Tencent Hunyuan Community License restricts usage by territory, including explicit exclusions for the European Union, the United Kingdom, and South Korea.

Safety & compliance

• Don’t use this model to impersonate real people or create deceptive content. Get consent for any recognizable person.
• Avoid generating illegal sexual content, child sexual content, or instructions for wrongdoing.
• Follow the Tencent Hunyuan Community License terms, including the territory restrictions and redistribution rules.