Vidu Q4 Preview: 15 photos and 3 voices, one character
Vidu Q4 Preview takes up to 15 reference images and 3 voice clips, so your AI character keeps the same face and voice across clips.
You finally have an AI character whose face you love. Then you make the second clip and something drifts: the jawline, the hair, and worst of all the voice, which sounds like a different person every time. This week a new video model went straight at that problem. Shengshu Technology released Vidu Q4 Preview on October 7, and it lets you show the model up to 15 pictures and up to 3 voice clips of your character in one go.
- Reference-to-video with 1 to 15 images for characters, outfits, props and places, all in one generation.
- Up to 3 voice references (3 to 12 seconds each) that keep a character's voice and delivery the same, with lip-synced dialogue.
- 3 to 16 seconds per clip at 24 fps, from 540p up to 4K, with 10-bit colour at the higher tiers.
- Native audio: dialogue and sound effects are generated in the same pass as the picture.
Vidu Q4 Preview is the first public look at Shengshu's next flagship model. It runs in the Vidu web app, through the Vidu API as viduq4-preview, and inside ComfyUI, where two partner nodes for image-to-video and reference-to-video were merged on launch day. Perfect Corp. also added it to the image-to-video feature of its five YouCam apps on October 8.
Why this matters for AI avatar creators: until now, most tools let you lock a face with one or a few images and then gave the voice to a separate text-to-speech step. Vidu Q4 puts the look and the voice in the same request. You can give it a close-up, a full-body shot, the outfit and the location as separate pictures, add a short clip of your character's voice, and write the lines your character should say.
How to use Vidu Q4 now that it takes 15 references
The following images were generated using Nano Banana 2:
- Build a reference set once. Make a clean close-up, a three-quarter view and a full-body shot of your character, plus one image each for the outfit and the place. Reuse the same set for every clip so the model sees the same person each time.
- Refer to images by number. In the reference node, every image counts toward the 15 cap and you point to them by their order in the prompt, for example "the woman from image 1 wearing the cardigan from image 4".
- Give the voice, write the words. A voice clip only donates the voice, not the words. Write the dialogue in the prompt and assign each line to voice 1, 2 or 3. Keep the audio switch on, or the voice references are ignored.
- Draft small, finish big. The default is 720p and 5 seconds. Test openings and camera moves at low resolution, then render the keeper in 1080p or 4K.
Worked example — the rooftop host. One clear portrait like this, plus a full-body shot and an outfit image, is the kind of reference set that lets the same host appear in every weekly clip. Pair it with a 10-second recording of her voice and write her lines in the prompt.
Prompt (Nano Banana 2, 1:1):
Portrait of a young East Asian woman with a sleek black bob and a mustard-yellow knit cardigan, standing on a sunny city rooftop at golden hour, city skyline softly blurred behind her, gentle smile, looking just past the camera, natural skin texture, 85mm lens, shallow depth of field, editorial photography, no text, no logos
Worked example — two characters talking. Reference-to-video can carry more than one person. Give each character their own portraits and their own voice clip, then write a short back-and-forth and say which voice reads which line. This is where the 3 voice references matter: two characters, two voices, and one left for a narrator.
Prompt (Nano Banana 2, 1:1):
Two friends sitting across from each other at a small cafe table by a window, a Black woman with braids in a green jacket laughing, and a red-haired man in a denim shirt holding a coffee cup and talking, soft afternoon light, warm interior, candid moment, realistic skin, editorial photography, no text, no logos
The honest take: this is a preview, not the final Q4 model, and there are no independent tests yet, so treat the consistency claims as the company's own until you have tried your own character. Prices are also unclear. The two launch press releases quote different starting prices, Vidu's API bills by the second with more credits for higher resolutions, and a 30 percent launch discount runs until November 30. There is also no text-to-video mode in the preview: you always start from at least one image.
Receipts: Vidu Q4 product page, ArtRealmAI on the API and ComfyUI nodes, Perfect Corp. YouCam announcement, SlopTV on the two launch prices.
The bottom line for AI avatar creators
The hard part of running an AI character was never one good clip. It was the tenth clip looking and sounding like the first. Vidu Q4 Preview takes a big stack of pictures and the character's voice in one request, which is exactly the setup a recurring AI host needs.
It still pays to have a strong, consistent set of images of your character before you open any video tool. If you don't have that yet, start with your avatar.
FAQs about Vidu Q4 Preview
When was Vidu Q4 Preview released?
Shengshu Technology released it on October 7, 2026, as a public preview of its next flagship video model. The final Q4 model has no announced date.
How many reference images and voices can I use?
Up to 15 reference images and up to 3 reference voice clips in one reference-to-video generation. Each voice clip is 3 to 12 seconds long.
How long and how sharp are the clips?
Each clip is 3 to 16 seconds at 24 frames per second. You can choose 540p, 720p, 1080p, 2K or 4K.
Can it make vertical videos for Reels and TikTok?
Yes. Reference-to-video supports 9:16 as well as 16:9, 1:1, 3:4 and 4:3. Image-to-video keeps the shape of the photo you upload.
Where can I try it?
In the Vidu web app, through the Vidu API (model viduq4-preview), in ComfyUI with the Vidu Q4 partner nodes, and in the YouCam apps for image-to-video.