Dance Now

Generate five-second videos from batches of images using the open Wan 2.2 TI2V-5B model and a single cost-conscious AWS Batch Spot GPU job.

The default design deliberately processes a whole manifest sequentially on one g5.xlarge. Wan is downloaded and loaded once, then reused for every image in the batch. The Batch compute environment has zero minimum vCPUs and at most one GPU instance, so it scales back down when the work is finished.

What gets created

• A worker image containing the Wan inference code (not the model weights), built and published to a Docker Hub repository entirely via AWS CodeBuild — no local machine ever needs to build or upload it (see Build the worker image below). ECR is also supported as a local-build alternative.
• An AWS Batch Spot compute environment restricted to one g5.xlarge.
• A job queue and job definition.
• IAM roles scoped to a dance-now/ prefix in the S3 bucket supplied during deployment.
• A temporary 150 GB encrypted root disk, deleted with the Spot instance.

Model weights are downloaded from Hugging Face at the beginning of each cold job. This avoids paying every month for persistent model storage. Input images, the generated MP4 files, and JSON metadata live in your existing S3 bucket.

Prerequisites

• AWS CLI credentials with permission to use CloudFormation, IAM, CodeBuild, Secrets Manager, Batch, EC2, and the chosen S3 bucket.
• A Docker Hub account, if using the recommended CodeBuild build path (a free account is enough — see Build the worker image). Docker with buildx is only needed for the local-build alternative.
• A VPC and subnet with outbound internet access. A default VPC works.
• An EC2 Spot quota for G/VT instances — g5.xlarge needs 4 vCPUs of that quota. New accounts often start at 0 and need a one-time increase request (AWS Console → Service Quotas → Amazon EC2 → "All G and VT Spot Instance Requests") before the first GPU job can start; this can take anywhere from minutes to a couple of days for AWS to approve.
• Python 3.10 or newer for the local submission command.

The scripts use the normal AWS CLI region resolution. On this machine that is currently eu-west-2. Set AWS_REGION if the bucket and GPU resources should be somewhere else.

1. Build the worker image

The image (CUDA + PyTorch + Wan2.2 + this project's worker script) is several GB — building and uploading it from a laptop can take hours on a slow or metered connection, and never needs to happen more than once per code change. Recommended: build it entirely inside AWS via CodeBuild, which pushes to a Docker Hub repository you own; your own connection only ever carries the source code (tens of KB), never the image itself.

One-time setup, after creating a Docker Hub access token scoped to "Read & Write" (not your password):

aws secretsmanager create-secret \
  --name dance-now/dockerhub-credentials \
  --secret-string '{"username":"YOUR_DOCKERHUB_USER","password":"YOUR_ACCESS_TOKEN"}' \
  --region YOUR_REGION

chmod +x scripts/*.sh
./scripts/deploy_codebuild.sh dance-now-codebuild YOUR_BUCKET YOUR_DOCKERHUB_USER

Then, whenever the image needs (re)building:

./scripts/build_via_codebuild.sh dance-now-worker-build YOUR_BUCKET

This builds, smoke-tests (confirms Python/CUDA/Wan/worker imports cleanly before anything is published), and pushes the image, then prints it pinned by content digest, e.g. YOUR_DOCKERHUB_USER/dance-now-worker@sha256:... — pin to that exact digest rather than the mutable latest tag, so a later push can't silently change what you're running. Use that value as ContainerImage / --image-uri in the steps below.

Alternative: build locally and push to ECR

If you have fast, unmetered bandwidth and would rather not use Docker Hub:

./scripts/build_and_push.sh dance-now-worker v1

The last output line is the ECR image URI. The first build is large and may take a while because it contains CUDA, PyTorch, and Wan's dependencies.

2. Deploy the Batch stack

Find a VPC and suitable subnet IDs, then deploy, using the image URI from step 1:

./scripts/deploy_stack.sh \
  dance-now \
  YOUR_BUCKET \
  YOUR_DOCKERHUB_USER/dance-now-worker@sha256:... \
  vpc-0123456789abcdef0 \
  subnet-0123456789abcdef0,subnet-0123456789abcdef1

The script prints the resulting JobQueue and JobDefinition values. The stack uses Spot price/capacity optimization and will bid up to 70% of On-Demand. Change the SpotBidPercentage parameter if jobs remain RUNNABLE for too long.

3. Install the local submitter

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

Submit a local image directory

dance-now-submit \
  --input-dir ./images \
  --bucket YOUR_BUCKET \
  --prompt "The dancer moves naturally while the camera slowly pushes in" \
  --job-queue DANCE_NOW_JOB_QUEUE_ARN \
  --job-definition DANCE_NOW_JOB_DEFINITION_ARN

Supported image formats are JPEG, PNG, and WebP. Add --sidecar-prompts to use a prompt from photo.txt beside photo.jpg; images without a sidecar retain the default --prompt.

Submit images already in S3

dance-now-submit \
  --s3-input-prefix s3://YOUR_BUCKET/dance-now/source-images/ \
  --prompt "Subtle body movement and fabric motion, locked camera" \
  --job-queue DANCE_NOW_JOB_QUEUE_ARN \
  --job-definition DANCE_NOW_JOB_DEFINITION_ARN

The command prints the AWS job ID and manifest URI. Outputs default to:

s3://YOUR_BUCKET/dance-now/outputs/<job-name>/0000-<image-name>.mp4
s3://YOUR_BUCKET/dance-now/outputs/<job-name>/0000-<image-name>.json

Use --dry-run to upload inputs and create the manifest without submitting GPU work. Run dance-now-submit --help for output-prefix, job-name, seed, and region options.

What you can and can't configure

Wan 2.2 TI2V-5B's pipeline only takes a start image, a text prompt, a fixed set of resolution presets, and a frame count — so that's what's exposed:

Setting	Configurable?	How
Prompt (motion/camera direction)	Yes, per image	--prompt, or per-image via --sidecar-prompts (photo.txt beside photo.jpg)
Duration	Yes	--seconds. Rounded to the nearest frame count Wan supports (frames must be 4n+1); default is 121 frames ≈ 5.04s
Resolution / aspect ratio	Yes, auto by default	Picked per image from Wan's 8 supported presets (1280*704, 704*1280, 1280*720, 720*1280, 1024*704, 704*1024, 832*480, 480*832) using whichever preset's aspect ratio is closest to that image's own. Pass --video-size to force one preset for every image in the batch instead
Seed	Yes	--seed (base seed; each image in a batch gets seed + index)
Output format	No	Always MP4
End frame / last-frame conditioning	No	Wan's TI2V-5B pipeline takes a single start image only — no last-frame argument exists in this model. Wan does have a separate first-last-frame model (FLF2V-14B), but it isn't wired into this repo (larger weights, more VRAM, different pipeline)
Sampling steps, guidance scale, shift, fps	No	Fixed to Wan's own TI2V-5B defaults (50 steps, 5.0 guidance, 5.0 shift, 24 fps) — not currently exposed as flags

Fully automated run (build → deploy → submit → wait → teardown)

Steps 1–3 above are still useful for iterating, but for a normal batch you can run everything in one shot with scripts/run_job.sh. It builds/pushes the image if needed, deploys the CloudFormation stack, submits the job, polls until it finishes, prints the outputs (or the failure logs), and then deletes the stack it created — so nothing is left running or sitting in your account when it exits, including on Ctrl-C or a failed build.

./scripts/run_job.sh \
  --stack dance-now \
  --bucket YOUR_BUCKET \
  --vpc vpc-0123456789abcdef0 \
  --subnets subnet-0123456789abcdef0,subnet-0123456789abcdef1 \
  --input-dir ./images \
  --prompt "The dancer moves naturally while the camera slowly pushes in" \
  --image-uri YOUR_DOCKERHUB_USER/dance-now-worker@sha256:...

Requires the venv from step 3 to be active (dance-now-submit on PATH). Run ./scripts/run_job.sh --help for the full option list, including --s3-input-prefix, --seed, --seconds, --video-size, --job-name, and --poll-seconds.

Pass --codebuild-project dance-now-worker-build instead of --image-uri to have it build fresh via CodeBuild first (see step 1) — useful when iterating on the Dockerfile/worker code itself; skip both to build locally into ECR.

On success it also downloads the resulting .mp4/.json pair from S3 into videos/<job-name>/ (a sibling of images/) so you don't have to go fetch them yourself. By default the S3 copies (manifest, uploaded inputs, outputs) are left in the bucket; pass --delete-outputs to delete just this job's objects once the local download is confirmed — the bucket itself is never deleted, so --bucket can safely point at a fixed bucket you keep around permanently rather than a throwaway one.

With --image-uri (a Docker Hub image, as recommended), the image is never deleted on exit — that's the point of publishing it once and reusing it. With a local ECR build, by default it deletes the ECR repository/image on exit instead, so a rerun rebuilds the multi-GB CUDA image from scratch — safe, but slow; pass --keep-image to skip that rebuild next time. --keep-stack similarly skips deleting the CloudFormation stack (Batch scales it to zero vCPUs between jobs, so leaving it costs nothing but the IAM/queue bookkeeping).

Monitor a job

aws batch describe-jobs --jobs JOB_ID

The AWS Batch console links each running child to its CloudWatch log stream. A Spot interruption is retried once. On retry, the worker checks S3 and skips MP4 files that were already completed, avoiding duplicate inference.

Cost controls

• MinvCpus: 0 prevents an always-on GPU.
• MaxvCpus: 4 permits only one g5.xlarge at a time.
• One job handles all images sequentially and loads the model once.
• The model disk is deleted when the instance terminates.
• Existing output files are skipped during retries.

AWS Batch can take several minutes to scale an idle compute environment back to zero. Check the EC2 console after the first test and set an AWS Budget alert. For an initial smoke test, submit exactly one image before starting a larger batch.

Alternatives

This project self-hosts the open Wan 2.2 TI2V-5B model on your own AWS Batch GPU job, which keeps costs low but means you manage the infrastructure. If you would rather use a hosted, pay-as-you-go generation service instead, some alternatives include:

• MiniMax H3 (768P) — ~$0.047/sec, ~2250s video generation
• Hailuo 2.0 / 2.3 — unlimited plans, ~3857s video generation
• Nano Banana Pro — 4K, unlimited
• Nano Banana 2 — 4K, unlimited
• Seedream 5.0 Lite — 3K, unlimited
• Seedream 4.5 — 4K, unlimited
• GPT Image 1.5 — unlimited
• Light Studio — unlimited
• AnyAngle — unlimited
• Veo 3.1 — via unlimited-plan access
• Sora 2 — via unlimited-plan access

These are commercial, hosted offerings and pricing/limits change frequently; check each provider's current terms before relying on them.

Tests

The lightweight tests do not require CUDA or AWS access:

PYTHONPATH=src python3 -m unittest discover -s tests -v