How to Generate AI Videos from Text or Images: Full Guide (2026)

From reference to reviewed video

Anchor the shot with a suitable reference, describe one clear motion, and review the whole result before extending the sequence.

AI video generation creates short clips from text prompts or still references. Framedance's AI Generator exposes text-to-video, image-to-video, reference-to-video, and video-editing routes in one interface; media slots, duration, ratio, resolution, sound, and other controls depend on the selected model. Generation consumes credits, so review the current displayed cost before submitting.

TL;DR

Generate videos from plain text (text-to-video) or animate a still image (image-to-video).

Top video models aggregated in one place — compare them on the same prompt and keep the best.

Chain your workflow: generate an image, click 'use as reference,' and turn it into a video without leaving the page.

Seeds, batch runs, and on-page history make iteration systematic instead of random.

Cost varies with model, duration, resolution, media route, and other billable parameters; check the current estimate before submitting.

How do you generate an AI video from text?

1. Open AI Generator, sign in, select the Video tab, and choose a method and model that match the shot.
2. Check the model card for its approximate per-video price and supported options.
3. Describe the shot: subject, action, setting, camera movement, and mood.
4. Hit generate and let the clip land in your on-page history.
5. Compare a few takes and reuse the seed of the best one for controlled variations.
6. If the motion is not right, rerun the same prompt on a different model — switching models is one click, not a new subscription.

How do you turn an image into a video?

Image-to-video starts from a picture — upload a photo, or use an image you just generated — and the model animates it, using your image as the visual anchor. When you care about composition, a character's look, or a product's appearance, this route is far more controllable than text alone.

When subject appearance matters, settle composition and identity in a still, approve it, then choose 'use as reference' and pass it to a compatible image-to-video model. Image and video costs do not have a fixed ratio, so check each estimate; the main value is catching appearance errors before unsuccessful video iterations.

What are people actually making with AI video?

Short-form content is the volume king: hooks, B-roll, and visual gags that would be unfilmable on a real budget. Marketers render products into impossible settings, musicians storyboard an entire music video before booking a single location, and studios lean on text-to-video for previsualization — cheap moving drafts of expensive shots.

Common tasks include previsualization, short-form hooks, B-roll, product concept shots, and character-motion drafts. Generation is not always cheaper than stock or filming, and identity does not stay consistent automatically; compare real licensing and production costs and reuse approved references for each shot.

How do you write prompts for AI video?

Describe motion explicitly — that is the biggest difference from image prompts: what moves, how fast, and what the camera does (dolly, pan, locked off). A video prompt is a shot description, not an image caption.

Keep each prompt to one scene and one continuous action. Cramming a multi-scene storyline into a single clip almost always fails — generate the shots separately and cut them together in an editor.

A slow dolly-in on an astronaut walking through a sunflower field
at golden hour, petals drifting in the wind, warm backlight and
lens flare, cinematic color grading, smooth stabilized camera motion

How much does AI video generation cost?

Video generation consumes credits, and model, duration, resolution, reference route, sound, and other settings can change the charge. Pricing and capabilities can be updated, so use the value displayed before submission.

To control budget, validate the shot brief and still reference, then test motion with the smallest settings that answer the review question. Faster or higher-quality options do not map to one universal price tier; compare current models individually.

Before a batch, read the current parameters and run one short shot with explicit acceptance criteria. One sample cannot represent every prompt and source asset.

Can you generate videos through the API?

Yes. Create a key at /api-keys, read the current video-model parameters, POST the task, and poll task_uuid at a moderate interval until completed, failed, or cancelled.

For pipelines that render many shots, the API can record model inputs, task UUIDs, terminal status, and outputs systematically. Compare the operational effort and current credit estimate with the web workflow before choosing it.

curl -X POST "https://www.framedance.ai/api/v1/marketplace/run/<model_id>" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "prompt": "a slow dolly-in on an astronaut walking through a sunflower field at golden hour"
    }
  }'

# Video jobs take longer — poll the task until it completes
curl "https://www.framedance.ai/api/v1/marketplace/run/tasks/<task_uuid>" \
  -H "Authorization: Bearer YOUR_API_KEY"

Practice: compare a text-led and image-anchored shot

Plan one simple shot: “A travel mug rests on a cafe counter; steam rises gently while the camera makes a slow, short push-in.” For text-to-video, describe the mug, cafe, light, single steam motion, and camera move. For image-to-video, first approve a still with the desired opening composition, then use it in the model’s supported first-frame or reference slot and keep the motion instruction concise. Do not claim the two routes will match exactly.

Choose only duration, resolution, aspect ratio, audio, last-frame, or other controls actually shown for the selected model. Some image-to-video routes inherit the source ratio; do not force an unsupported setting. Review the displayed cost and make one short clip from each route. The goal is to learn which route better preserves composition and follows the single motion, not to infer a universal winner from one attempt.

Watch the whole clip before the next render

Play every frame at normal speed and inspect the opening, middle, and final frames separately. Accept only when the mug keeps a coherent shape, the steam moves without becoming a new object, the camera move stays smooth, the counter and background remain stable, and the final frame is still usable. If audio was intentionally enabled and supported, review it separately for relevance and artifacts. A strong opening frame does not excuse a broken ending.

If the composition drifts, use the approved still and a supported image-anchored route. If motion is chaotic, remove secondary actions and keep only steam plus the slow push-in, or reduce to one motion. If the subject changes near the end, test a shorter supported duration before revising everything. If the camera is static, make the camera instruction more explicit. Change one input per retry and begin with short renders; longer duration, higher settings, and repeated generations may increase credit use according to the selected model.