Pruna VACE

Overview

Pruna VACE is Pruna AI's optimized version of VACE (All-in-One Video Creation and Editing), a video model from Alibaba's Tongyi Lab. In the original release, one VACE model handles reference-to-video generation, video-to-video editing and masked editing.

On Wiro the model runs in two ways: from a text prompt alone, or from a prompt plus one to three reference images. Pruna highlights character consistency through reference images. What is in your references carries over into the clip: a person, an animal, an object or a background. The prompt describes the action and the camera.

What you can build

• Short clips that place a person from a portrait, or a product from a product photo, in a new scene.
• A series of mascot or brand-character shots, made by reusing the same reference images with different prompts.
• Vertical 9:16 clips for social feeds and 16:9 shots for landing pages and ads.
• Text-only scenes, like the wide rocky coastline at sunrise or the woman by a rush-hour bus window in the example prompts.

Inputs

The prompt is required. Pruna advises being specific about the characters, what they do, the setting and the camera angle, and naming a style such as "cinematic" if you want one. The example prompts read like shot descriptions, with camera moves such as a slow pan or a slow zoom out.

You can add up to 3 reference images in JPG, PNG, GIF, WebP or HEIC. Pruna recommends clear, high-quality images with consistent lighting and angle. One example prompt animates a single photo of two people in the same grassy field it was taken in, adding a slow camera move and small gestures such as the man adjusting his tie. The default prompt combines two references, a photo of a deer and a photo of snowy, bare trees, into one clip of the deer walking past that tree line.

Resolution is 480p (the default) or 720p. Aspect ratio is Auto, 16:9 or 9:16. Auto picks whichever of the two is closer to your first reference image, and uses 16:9 when there is no image.

The advanced settings cover the frame count (1 to 81, default 81), sample steps (1 to 100, default 50), the solver (UniPC or DPM++), guidance scale (1 to 10, default 5; higher follows the prompt more strictly), sample shift (0 to 32, default 16) and a seed. Speed Mode has three levels. Lightly Juiced is the default and the one Pruna rates best for quality and consistency. Juiced and Extra Juiced give up some quality.

Outputs

Each run returns one video clip (the samples are MP4). At 480p the frame is 832x480, or 480x832 in portrait. At 720p it is 1280x720 or 720x1280. Clip length follows the frame count. The samples run at 16 fps, so 81 frames make about 5 seconds. Clips have no sound: the coastline and bus example prompts ask for gulls and ambient noise, but their samples are silent.

Recommended settings

Draft at 480p with Extra Juiced, which Pruna suggests for test runs. Keep the seed fixed while you rework the prompt, so the differences you see come from the text. Once the shot looks right, switch to Lightly Juiced (the default) and 720p for the final render.

Limitations

• One run makes at most 81 frames.
• Two frame shapes only: 16:9 landscape and 9:16 portrait.
• At most 3 reference images per run.