Vidu Q4 Image to Video: Prompts, API, and Full Workflow

👁 119 reads

Vidu Q4 Preview does not generate video from pure text-it always starts with a still image. That is actually an advantage, as it gives you complete control over framing, characters, and lighting before spending a single cent on video compute. In this guide, we walk through preparing source images, writing effective motion prompts with audio cues, calculating API generation costs, and automating runs via code. For general details about the model, check our overview of Vidu Q4.

Step 1. Prepare your starting image

In Image-to-Video mode, the first frame of your video is identical to your input image. The model will not correct poorly framed shots, so all compositional decisions happen at this stage. Keep these best practices in mind:

• Match target aspect ratio. The output video inherits the dimensions of your source image. Crop to 16:9 for YouTube or 9:16 for vertical reels before uploading.
• Leave room for movement. If a subject needs to exit the frame or the camera pulls back, ensure there is ample negative space. A tight headshot cannot cleanly transition to a wide shot.
• Avoid baked-in text. Unless text needs to stay fixed, avoid on-image lettering. Video diffusion models often warp characters during motion, especially non-ASCII glyphs.
• Match or exceed output resolution. For 1080p outputs, use an image at least 1920 px wide; for 4K video, supply at least 3840 px.
• Real people and licensing. Only animate real faces with permission. Commercial use also requires a paid plan rather than the Free tier.

You can create source frames using any modern image model. For tools with strong prompt adherence, browse our guide to AI image generators, or evaluate them side-by-side in our AI image generator arena.

Step 2. Structure your motion prompt

An Image-to-Video prompt should not describe what is already visible in the static frame; it must describe what happens next. We follow a strict 5-part structure:

1. Action: Who does what, written in present tense (“She lifts the cup and smiles”).
2. Camera movement: Static shot, dolly-in, orbit, handheld, or quick cuts.
3. Dialogue: Exact spoken words in quotation marks, noting the speaker and language.
4. Audio design: Background ambience, sound effects, or explicit negative constraints (“no music”).
5. Technical constraints: “no subtitles, no on-screen text”, “real-time speed, no slow motion”.

English prompts yield the most consistent motion adherence across diffusion video models. Vidu also enables automated prompt enhancement by default (the is_rec flag in the API). When you require strict, literal execution of your prompt, make sure to disable it.

Step 3. Tested prompt templates

Each prompt below assumes a specific source frame, described in the first column. We use these identical scenes to evaluate competing models across lip sync, physics, and camera control.

Step 4. Select length and resolution

Generation costs scale linearly per second and spike significantly with resolution. Our standard recommendation: run fast tests at 540p or 720p, produce final clips at 1080p, and reserve 4K strictly for production deliveries.

Pricing sources: fal.ai (30% discount through November 30) and WaveSpeedAI, matching standard direct Vidu pricing ($0.005 per credit). Enabling native audio generation does not incur extra charges.

Step 5a. Using the web interface

1. Navigate to vidu.com/vidu-q4 and sign in.
2. Switch the generation mode from “Reference to Video” to “Image to Video”.
3. Upload your starting image, paste your motion prompt, and choose duration and target resolution.
4. Click generate, then download the resulting MP4 from your library.

On the Free tier, Vidu Q4 at 540p consumes 4 credits per second; paid plans (starting at $10/month) charge 2 credits per second. See our full breakdown in the Vidu Q4 overview.

Step 5b. Running via API

The official Vidu REST API accepts your starting image either as a direct URL or as a base64-encoded string. Here is a standard curl request from the API docs (replace with your personal bearer token):

curl -X POST https://api.vidu.com/ent/v2/img2video 
  -H "Authorization: Token YOUR_API_KEY" 
  -H "Content-Type: application/json" 
  -d '{
    "model": "viduq4-preview",
    "images": ["https://your-domain.com/frame.jpg"],
    "prompt": "Slow dolly-in, beans pour into the grinder, crisp ASMR grinding sound, no music.",
    "duration": 8,
    "resolution": "1080p",
    "audio": true,
    "is_rec": false
  }'

The endpoint returns a task_id. Poll status via GET https://api.vidu.com/ent/v2/tasks/{task_id}/creations. Note that generated file URLs remain valid for 24 hours, so persist the file locally right away.

If you prefer running via fal.ai, here is a complete Python snippet using standard requests (set your FAL_KEY environment variable first):

import base64, os, time, requests

headers = {"Authorization": "Key " + os.environ["FAL_KEY"]}
image_data = "data:image/png;base64," + base64.b64encode(open("frame.png", "rb").read()).decode()

task = requests.post(
    "https://queue.fal.run/fal-ai/vidu/q4/image-to-video",
    headers=headers,
    json={
        "image_url": image_data,
        "prompt": "Slow dolly-in, crisp ASMR grinding sound, no music.",
        "duration": 8,
        "resolution": "720p"
    }
).json()

while requests.get(task["status_url"], headers=headers).json().get("status") != "COMPLETED":
    time.sleep(10)

result = requests.get(task["response_url"], headers=headers).json()
open("output.mp4", "wb").write(requests.get(result["video"]["url"]).content)

Full disclosure on our initial API test: this script mirrors our benchmarking harness. During our October 8 testing run, fal.ai rejected our queue submission due to depleted account credits, so we have not yet captured the final returned payload. The schema parameters (image_url, prompt, duration, resolution, seed) strictly match fal.ai documentation for this endpoint, and we will update with full end-to-end execution logs shortly.

Common generation pitfalls

• Describing the image instead of the action. “A woman in a red coat on the street” gives the model zero direction because the visual elements are already present. Instruct what she does next.
• Cramming too much action into short clips. A 5-second clip can handle one clear movement and one camera adjustment. For complex sequences, render 10-16 seconds or cut between two distinct generations.
• Omitting “no music”. Multimodal video engines love inserting intrusive background music tracks unless explicitly instructed otherwise.
• Unintended text warping. If your source image features signs or logos, check the final frame-camera motion frequently degrades letters into gibberish.
• Defaulting directly to 4K. Prototype pacing and camera moves at 540p first, where compute costs are over 8x cheaper per second.
• Letting API download links expire. Direct Vidu storage links expire after 24 hours. Always download files automatically via your queue worker.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Need text-to-video instead?

Vidu Q4 does not support direct text-to-video prompt inputs. If you need to generate video without preparing a still frame first, consider models built for direct text generation-in our comparative testing, Wan 3.0 leads the category, closely followed by MiniMax H3 Max. You can read our detailed side-by-side analysis in Vidu Q4 vs Seedance 2.0 vs MiniMax H3, or review our workflow guide for Wan 3.0 Image to Video.

Sources: Vidu Q4 API Documentation (Image to Video), fal.ai Vidu Q4, WaveSpeedAI, vidu.com Pricing. Verified October 8, 2026.