AMD Creative Factory

Turn a description or an image into a short video with sound, or straight into a still image. All of it on AMD GPUs.

Video and Image are separate

Video and Image in the top bar are two separate pages. Each has its own model list, its own work list and its own Explore feed. Whichever you pick, you only see that kind of output — no videos mixed into the image page.

They are split because they are not the same thing: video has a duration, has sound and comes out as MP4; an image has none of those and comes out as PNG. The URLs are separate too — bookmark or share /video and /image directly.

The site root redirects to Video, so old bookmarks keep working.

What Video does

• Text to video — write a description and get a shot from nothing.
• Image to video — supply an image as the first frame and describe how it moves.
• First & last frame — give a start and an end image; the model fills in between.

Output is MP4 at 24 fps, 768px on the short edge, with a generated stereo soundtrack. Aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 — portrait and landscape are native, not cropped.

What Image does

• Text to image — write a description, get a PNG.

Output is PNG, with the same choice of aspect ratios; exact pixel sizes depend on the model. Text to image is the only task type here for now — image upload, duration and sound do not appear on the Image page at all. They are not hidden; they genuinely do not exist for images.

Image models run on AMD Radeon™; video runs on AMD Instinct™. It makes no difference to you — same API, same credits, same retention.

Six steps

1. Sign in with your AMD account. First sign-in grants 100 credits: about 16 five-second videos, or 100 images.
2. Pick Video or Image in the top bar. This decides which models you can pick and whether a duration control exists at all.
3. Write a prompt in the bar at the bottom. Up to 8,000 characters.
4. Adjust by tapping + for model and aspect ratio — plus duration and input type on the video side. The defaults suit most cases.
5. Generate. Five seconds of video takes about 25 seconds; a 1024×1024 image about 50. Queue position and elapsed time update live.
6. Publish — optional. Share it to Explore for everyone to see.

Writing prompts

Say how the camera moves. If you don't, the model decides, and the result is often nearly static. A structure that works:

subject + specific action → camera move → light and mood → medium

Example:
  A lighthouse on a black volcanic shore, surf rolling in and
  dissolving into foam;
  the camera drifts slowly sideways from above;
  low dawn light rakes across the water from the right, turning
  the spray amber;
  photorealistic, shallow haze, natural colour grading.

For video you can describe sound as well as picture — surf, rain, room tone. For images only the picture half applies. Everything in Explore has a copy-prompt button.

Duration

Frame count lands on a 17n+5 lattice at a fixed 24 fps, so asking for 5 seconds delivers 5.167. The ticks on the duration slider are those positions; nothing exists between them. Between 4 and 15 seconds, 8.000s is the only whole second. Billing follows the delivered length.

This section is about video only. Image models have no duration: the control disappears from the interface, and passing duration to the API changes nothing.

Credits

Video is billed by delivered length at roughly one credit per second, so five seconds costs 6. An image at default steps costs 1. Credits are frozen on submit, charged on success, and refunded in full on failure.

The Account page shows the full ledger: every grant, hold, charge and refund, traceable to the task it belongs to.

Keeping your work

Results are kept for 7 days; each card shows how long is left. Press Keep to hold one indefinitely, up to 20 at a time. Anything published to Explore never expires.

Download links are signed and valid for 24 hours. They need no login, so browsers, ffmpeg and ComfyUI can fetch them directly — and so can anyone else holding the link.

Publishing to Explore

Finished work has a Publish button, for video and images alike. Publishing puts it in a review queue; once an admin approves it, it appears in the Explore feed of its own tab — video in Video, images in Image. Your nickname and prompt become public too — the confirmation says so. You can unpublish at any time, with no review and no delay.

Review exists because the site runs on an AMD domain, so anything public carries that brand.

Limits

Using the API

Everything in the interface is available over HTTP, wire-compatible with the MiniMax video generation API. Video and images share one API: you change the model and task fields, not the endpoints. Request and response shapes, status codes and error types are in the API reference; runnable samples are on the home page under the API tab.