Use your cast in Nova Reel
Nova Reel uses one reference image as the starting key frame of a 6 second shot and, in manual multi-shot mode, lets you give every shot its own starting image, which is the practical way to use several views from the pack in one video. It has no reference-image or last-frame mode and produces silent 1280x720 video, so consistency depends entirely on the start image of each shot and the bible text.
Which file goes where
- hero.pngStarting image of the first shot. Must be re-sized or letterboxed to exactly 1280x720 before upload.
- three_quarter.pngStarting image of a later shot in manual multi-shot mode (also 1280x720).
- profile.pngStarting image of a profile shot in manual multi-shot mode.
- back.pngStarting image of a walk-away shot in manual multi-shot mode.
- closeup.pngStarting image of a close-up shot; pad the 1:1 face to 1280x720 rather than stretching it.
- contact_sheet.pngNot used; Nova Reel has no reference slots.
- character.jsonBible text, trimmed, inside every shot prompt (prompts are short, so keep the essentials).
Capabilities
- First frame (image-to-video)Yes
- Last frameNo
- Reference imagesNone
- Native extendNo
- Max clip120 s at 1280x720 at 24 fps (6 second shots; multi-shot up to 2 minutes)
- AudioNo
Steps
- Enable model access for Amazon Nova Reel in the Bedrock console in US East (N. Virginia) and create an S3 bucket in the same region for output.
- Resize each pack image you will use to exactly 1280x720 PNG or JPEG: letterbox the 2:3 views on a plain background rather than cropping the face.
- Single shot: in the Bedrock playground or API, choose Nova Reel, upload hero.png as the input image and write a prompt that starts with a short bible line and ends with the action.
- Multi-shot (Nova Reel 1.1): choose manual multi-shot and define one shot per view, each with its own 1280x720 image and a prompt of at most 512 characters; up to 20 shots of 6 seconds.
- Keep every shot prompt starting with the same bible line so the model does not reinvent the outfit between shots.
- Submit; the job runs asynchronously and writes output.mp4 to your S3 prefix. Poll the invocation until it completes, then download.
- Add audio in your editor; Nova Reel outputs video only.
Prompt template
{{bible}} Same character as the input image: identical face, hair, skin tone and outfit. {{action}} Smooth realistic motion, steady camera, cinematic lighting. (Keep each prompt under 512 characters; trim the bible to the lines about face, hair and clothing.)
Tips
- The input image must be exactly 1280x720, so prepare a letterboxed set of the pack views once and reuse them for every shot.
- Use manual multi-shot as a poor-man reference mode: one shot per view, each seeded with the matching pack image.
- Short prompts work best; put the two or three most identifying bible lines first and the action last.
- The seed in the request only affects the model noise; it does not reproduce the character sheet seed. Fix it anyway for repeatable retries.
- For continuity between shots, describe the handoff (same room, same light, character turning left) in both adjacent shot prompts.
- Expect no audio; plan a voice and music pass in your editor.
Caveats
- Official docs state Nova Reel 1.1 is available only in US East (N. Virginia); the general Nova overview lists Europe (Ireland) and Asia Pacific (Tokyo) for Nova Reel, so check the current region table for the 1.1 model ID.
- Two model IDs exist: amazon.nova-reel-v1:0 (the original) and amazon.nova-reel-v1:1 (1.1 with multi-shot). Examples in AWS docs use v1:1; confirm which your account can call.
- Automated multi-shot takes a prompt only and no image; only manual multi-shot accepts a starting image per shot.
- The 512 character per-shot prompt limit and the 20-shot cap come from AWS blog and third-party write-ups rather than the API reference page checked here.
- The seed range (0 to 2,147,483,646) is reported by a reseller page and was not confirmed from AWS docs.
- No audio output was confirmed from the model card (output modality: video); treat audio as unsupported.
API notes
Call Bedrock Runtime StartAsyncInvoke (boto3 start_async_invoke) with modelId amazon.nova-reel-v1:1 and a modelInput of taskType TEXT_VIDEO, textToVideoParams containing text and an images list with format (png or jpeg) and a source of base64 bytes or an S3 location, plus videoGenerationConfig with durationSeconds 6, fps 24, dimension 1280x720 and a seed; multi-shot uses taskType MULTI_SHOT_MANUAL with a shots array. Pass an outputDataConfig pointing at your S3 bucket, then poll GetAsyncInvoke with the returned invocationArn until status is Completed and read output.mp4 from the S3 prefix.