calesthio/generative-media-skills/amazon-nova-reel

Amazon Nova Reel production

Treat Nova Reel as an asynchronous production service, not a synchronous clip function. Plan the shots, validate every input, obtain an exact approval, submit once with an idempotency token, persist the invocation ARN, and promote only validated S3 artifacts.

All volatile facts in this skill were verified against first-party AWS sources on 2026-07-10. Recheck model access, Region, price, quota, and lifecycle immediately before a paid run.

Establish the current contract

Do not select Nova Reel as a new default. AWS's Bedrock lifecycle table marks both published Reel versions Legacy as of 2026-03-30, with EOL 2026-09-30:

New customers cannot use Legacy models, existing customers may lose access after 15 days of inactivity, and new Provisioned Throughput cannot be created. On or soon after EOL, calls fail unless the customer has a private continued-access arrangement. Migration is not automatic. AWS's table does not identify an active Nova Reel successor, so do not invent one: for work that must operate past EOL, select another currently active provider unless AWS publishes a successor or the account team confirms a private arrangement.

Where the account has confirmed Reel 1.1 access and the delivery can finish before EOL, Reel 1.1 supports:

• TEXT_VIDEO: one six-second text-to-video or image-to-video shot.
• MULTI_SHOT_AUTOMATED: one long prompt expanded into 12–120 seconds in six-second increments; no input image.
• MULTI_SHOT_MANUAL: an ordered list of six-second shots, each with a prompt and optional starting image; duration follows shot count, up to 120 seconds.
• English prompts, 1280×720 output, 24 fps, silent video, asynchronous StartAsyncInvoke, and S3 delivery.

The Nova generation guide still teaches Reel 1.1, but the Bedrock lifecycle page explicitly says its dates govern Bedrock usage. Treat the lifecycle table as authoritative when the pages differ. Before any run, confirm the account can invoke the exact model in the exact Region, confirm the job will complete before EOL, and record the exit route.

Sources: Bedrock model lifecycle, Nova Reel generation guide, Bedrock Nova Reel model page, regional compatibility.

Exact input limits

Always send fps: 24 and dimension: "1280x720". Seed is optional, defaults to 42, and must be 0–2,147,483,646. Reference images must be PNG or JPEG, 1280×720, 8-bit RGB; PNG alpha is allowed only when every pixel is fully opaque. AWS documents a 25 MB limit for images supplied from S3. Use an image as the starting keyframe, not as a general style-reference bank.

AWS's current manual-mode prose does not publish a separate minimum shot count. Use TEXT_VIDEO for one six-second shot and manual mode for a genuinely multi-shot list; do not turn the production convention “at least two” into a claimed service-schema minimum. Reel 1.0 is the legacy six-second model; long and multi-shot requests require Reel 1.1.

Source: video generation access and input schemas.

Choose the mode from the edit decision

• Choose TEXT_VIDEO for a self-contained six-second insert, a first-frame animation, or cheap concept exploration.
• Choose MULTI_SHOT_AUTOMATED when overall theme and stylistic flow matter more than exact shot boundaries. It is not an image-conditioned mode.
• Choose MULTI_SHOT_MANUAL when the edit has an approved shot list, specific starting frames, product angles, or controlled transitions.
• Generate separate six-second jobs when shots require independent selection, replacement, budgets, or delivery schedules. A failed multi-shot request does not yield a dependable partial master.

Nova Reel has no native audio. Budget and produce music, dialogue, ambience, and effects separately. Do not promise lip sync, precise typography, exact product geometry, physical accuracy, or deterministic reproduction from a seed. AWS specifically warns that coherent on-screen text and anatomically correct humans are not assured.

Direct motion rather than converse

Write a visual description, not a chat command. Put the highest-value subject and action first, then environment, lighting, look, and camera motion. Use one primary camera move per six-second shot.

Useful camera phrases documented by AWS include dolly in, dolly out, pan left, pan right, tilt up, tilt down, orbit, tracking shot, zoom in, zoom out, and static shot. Describe visible change over time: “steam curls upward as the camera slowly dollies in,” not “make it cinematic.”

For image-to-video, describe only the desired motion and changes. Re-describing the reference inconsistently can encourage drift. Preserve identity through stable nouns and attributes: repeat the same product name, material, color, wardrobe, time of day, and lens language in every manual shot.

Production heuristic: lock a continuity bible before a multi-shot run:

1. Subject identity: exact stable descriptors and prohibited changes.
2. World rules: place, weather, time, palette, texture, aspect and lens feel.
3. Motion grammar: camera height, speed, direction, and transition intent.
4. Shot purpose: what new information each six-second unit contributes.
5. Continuity anchors: entry/exit direction, screen position, action state, and lighting direction.

Source: Nova Reel prompting best practices.

Build valid requests

Example: single-shot image-to-video request

Intent: animate an approved packshot without introducing a new product design.

model_input = {
    "taskType": "TEXT_VIDEO",
    "textToVideoParams": {
        "text": (
            "Slow dolly in on the cobalt glass perfume bottle. "
            "A narrow highlight travels across the glass while faint mist drifts behind it. "
            "Dark slate studio, soft rim light, premium restrained motion."
        ),
        "images": [{
            "format": "png",
            "source": {"s3Location": {
                "uri": "s3://approved-inputs/campaign-42/bottle-1280x720.png",
                "bucketOwner": "123456789012"
            }}
        }]
    },
    "videoGenerationConfig": {
        "durationSeconds": 6,
        "fps": 24,
        "dimension": "1280x720",
        "seed": 1845
    }
}

Expected result: a silent six-second 720p clip beginning from the supplied frame. Likely failures include label deformation, changing reflections, excessive mist, or camera motion stronger than requested. Reject rather than repair a materially incorrect product depiction.

Example: automated 18-second concept

model_input = {
    "taskType": "MULTI_SHOT_AUTOMATED",
    "multiShotAutomatedParams": {
        "text": (
            "Cinematic dawn study of a coastal research station built from weathered steel. "
            "Begin with a wide aerial approach over dark water, transition to scientists crossing "
            "a glass walkway, and finish inside the warm laboratory looking back toward the sea. "
            "Maintain overcast blue exterior light, amber interiors, realistic architecture, slow "
            "controlled camera motion, and a quiet documentary tone throughout."
        )
    },
    "videoGenerationConfig": {
        "durationSeconds": 18,
        "fps": 24,
        "dimension": "1280x720",
        "seed": 9071
    }
}

Use this mode for ideation, not exact editorial timing. Review each delivered shot as well as output.mp4.

Example: manual three-shot storyboard

model_input = {
    "taskType": "MULTI_SHOT_MANUAL",
    "multiShotManualParams": {
        "shots": [
            {"text": "Wide static shot of a cobalt glass perfume bottle centered on dark slate, soft white rim light from camera left, faint haze, premium minimal studio."},
            {"text": "Macro tracking shot moving left to right across the same cobalt glass bottle, preserving its rectangular silhouette and silver cap; the camera-left rim light glides over the glass."},
            {"text": "Slow dolly out from the same cobalt glass bottle on dark slate as the haze clears, preserving the silver cap, camera-left rim light, and restrained premium studio palette."}
        ]
    },
    "videoGenerationConfig": {
        "fps": 24,
        "dimension": "1280x720",
        "seed": 1845
    }
}

The repeated anchors are deliberate. If exact product shape is mandatory, attach an approved starting image to each shot and still require human product QA.

Gate every paid submission

AWS prices Nova Reel per generated second. The Bedrock pricing page's interactive Nova Reel row showed USD 0.08 per generated second when verified on 2026-07-10. Treat that as a dated planning snapshot, not an execution constant: Legacy/public-extended-access pricing can change, and the lifecycle table currently shows no public-extended-access start date for Reel. Re-read the official price row or obtain an account quote immediately before approval and execution. Quote duration × verified rate, then add S3, KMS, logging, retrieval, and transfer costs separately. At the dated snapshot, 6 s = $0.48, 18 s = $1.44, and 120 s = $9.60.

Source: Amazon Bedrock pricing.

Require an approval record containing:

• canonical approval-envelope SHA-256, exact request, model ID, Region, S3 input and output URIs;
• generated seconds, verified unit rate, currency, maximum authorized amount, and price verification timestamp;
• lifecycle verification timestamp, confirmed Legacy access, EOL date, planned completion date, and exit route;
• rights/consent status for every reference and depicted person;
• retention classification, approver identity, and expiration;
• an explicit statement that the run is paid and asynchronous.

Fail closed on a missing or mismatched digest, expired approval, non-finite or insufficient maximum, stale price/lifecycle evidence, unconfirmed Legacy access, a job that can cross EOL, wrong Region, unapproved bucket/prefix, or changed input bytes. A human “looks good” message is not a reusable spending authorization.

Example: dry-run-first submission

This complete example is dry-run-first. It submits only when EXECUTE=1, current price and lifecycle evidence are supplied, Legacy access is explicitly confirmed, the approval digest exactly matches that evidence and the AWS request, and the approved maximum covers the estimate. It does not prove access without a call; the operator must verify access in the account before authorizing execution.

import hashlib
import json
import math
import os
from datetime import date, datetime, timezone
from decimal import Decimal, InvalidOperation
from pathlib import Path
from urllib.parse import urlparse