Wan 2.2 TI2V-5B (Kelpie) Recipe

Last Updated: September 15, 2025

Overview

Wan 2.2 TI2V-5B is a text-to-video / image-to-video model that can produce cinematic 720p clips from a single prompt, or from a prompt plus a reference image. Model is open sources and can be found on Huggingface. Running Wan 2.2 TI2V-5B on SaladCloud is one of the most efficient and cost-effective ways to generate videos with open-source models today. Check our Benchmark for full performance and cost details. This recipe runs Wan as a Kelpie worker. You send jobs to the Kelpie queue; the worker pulls jobs, runs inference, and uploads results to S3-compatible storage of your choice. Image references can also be preloaded to s3 storage and synced with the worker. Kelpie provides a simple API for long-running jobs and data sync (download inputs before, upload outputs after). It also supports autoscaling by queue depth including scale-to-zero when idle.

Consumer Recipe (single-GPU nodes)

1 worker per node (one inference process per machine).
Targets consumer GPUs (≈24 GB+ VRAM), e.g. RTX 4090 / 5090.
Scaling: each node runs a single inference worker, so your maximum concurrency is exactly the number of nodes you have running (e.g., 10 nodes → 10 parallel jobs).

Prerequisites. Prepare Your Storage and Data

S3-compatible storage (Cloudflare R2 or AWS S3) for outputs and (optionally) reference images. You’ll provide your Access Key ID, Secret Access Key, Storage Region and Endpoint URL (R2).

Choose and create your S3-compatible bucket

Store:

Outputs (MP4s)
Optional inputs (reference images)

Popular options:

AWS S3: Get started with Amazon S3
Cloudflare R2: Get started with R2

See also the Kelpie storage guide: Cloud storage setup.

Plan your folder layout (prefixes)

We recommend job-scoped prefixes so multiple jobs never collide:

s3://<bucket>/
├─ assets/<job_id>/    # OPTIONAL: reference images for the job
└─ outputs/<job_id>/   # MP4s created by the worker

Paths used by the worker

/opt/outputs — final videos (.mp4) are written here
/opt/assets — optional reference image is placed here

Your Kelpie job must include a sync.after rule that uploads /opt/outputs/ to your bucket. If you provide a reference image, also include a sync.before rule to download it into /opt/assets/.

Getting Your Container Group ID

curl -X GET \
  --url "https://api.salad.com/api/public/organizations/<organization_name>/projects/<project_name>/containers/<container_group_name>" \
  --header 'Content-Type: application/json' \
  --header 'Salad-Api-Key: <api-key>'

This returns a JSON object for the container group; copy the value of its .id field for job submissions.

Parameters you can send (Kelpie job `arguments`)

Pass these flags directly to the worker script.

--prompt (string, required) — scene description
--size (string, default 1280*704) — output resolution
--frame-num (int, default 121) — total frames (~24 fps ⇒ ~5 s)
--sample-steps (int, default 50) — higher = better quality, slower
--base-seed (int, optional) — reproducibility
--sample-solver (unipc or dpm++, optional) — sampler choice
--sample-shift (float, optional) — advanced tweak
--sample-guide-scale (float, optional) — CFG strength
--output-filename (string, optional) — base name; worker appends .mp4 (defaults to job id)
Image input (optional; choose one):
- --image-b64 — raw base64 or data:image/...;base64, URI (highest priority)
- --image-url — public HTTPS URL
- --image-path — file already present in /opt/assets/
--extra-args-json (JSON) — pass-through for additional generate.py flags
--id (string, optional) — job id for logs/filenames (defaults to a UUID)

The worker fixes --task ti2v-5B, always enables --offload_model and --t5_cpu, and writes to /opt/outputs/<output-filename or job-id>.mp4.

Submit a Job to Kelpie (Text → Video)

Before running the command below, make sure you got all you need from the previous steps:

The container group ID. Check how to pull it in a step above.
An S3-compatible bucket and credentials set on the container (Access Key ID, Secret Access Key, and Endpoint URL or Region).
Your Salad API headers ready: SALAD_API_KEY, organization, and project.
Chosen generation parameters: at minimum a --prompt; optionally adjust the rest.

The request enqueues a text-video job. After the job completes, Kelpie uploads the MP4 from /opt/outputs/ to s3://<bucket>/outputs/job-001/.

export SALAD_API_KEY="<salad-api-key>"
export SALAD_ORGANIZATION="<organization>"
export SALAD_PROJECT="<project>"

curl -s -X POST "https://kelpie.saladexamples.com/jobs" \
  -H "Content-Type: application/json" \
  -H "Salad-Api-Key: $SALAD_API_KEY" \
  -H "Salad-Organization: $SALAD_ORGANIZATION" \
  -H "Salad-Project: $SALAD_PROJECT" \
  -d @- <<'JSON'
{
  "container_group_id": "<container_group_id>",
  "command": "python",
  "arguments": [
    "/opt/wan_worker.py",
    "--id", "job-001",
    "--prompt", "A moody cyberpunk street in the rain, neon reflections, shallow depth of field",
    "--size", "1280*704",
    "--frame-num", "121",
    "--sample-steps", "50"
  ],
  "sync": {
    "after": [
      { "bucket": "<bucket>", "prefix": "outputs/job-001/", "local_path": "/opt/outputs/", "direction": "upload" }
    ]
  }
}
JSON

Submit a Job to Kelpie (Image → Video)

First, upload your reference image to s3://<bucket>/assets/job-002/input.jpg.

curl -s -X POST "https://kelpie.saladexamples.com/jobs" \
  -H "Content-Type: application/json" \
  -H "Salad-Api-Key: $SALAD_API_KEY" \
  -H "Salad-Organization: $SALAD_ORGANIZATION" \
  -H "Salad-Project: $SALAD_PROJECT" \
  -d @- <<'JSON'
{
  "container_group_id": "<container_group_id>",
  "command": "python",
  "arguments": [
    "/opt/wan_worker.py",
    "--id", "job-002",
    "--prompt", "Dreamy golden-hour dolly shot through wildflowers, shallow depth of field",
    "--size", "1280*704",
    "--frame-num", "121",
    "--sample-steps", "50",
    "--image-path", "/opt/assets/job-002/input.jpg"
  ],
  "sync": {
    "before": [
      { "bucket": "<bucket>", "prefix": "assets/job-002/", "local_path": "/opt/assets/", "direction": "download" }
    ],
    "after": [
      { "bucket": "<bucket>", "prefix": "outputs/job-002/", "local_path": "/opt/outputs/", "direction": "upload" }
    ]
  }
}
JSON

Alternative image inputs (no S3 download)

--image-url "https://example.com/my-image.jpg"
--image-b64 "<base64...>" (raw or data:image/...;base64, URI)

Monitor a Job

Use the Kelpie API to check the status of a specific job. You can find job IDs in the response when you submit a job.:

curl -X GET \
  --url "https://kelpie.saladexamples.com/jobs/<kelpie-job-id>" \
  --header 'Content-Type: application/json' \
  --header 'Salad-Api-Key: <salad-api-key>' \
  --header 'Salad-Organization: <organization-name>' \
  --header 'Salad-Project: <project-name>'

Replace kelpie-job-id and the header placeholders with your values. The response includes the job status, the last machine that worked on it, and recent heartbeat info. You can also extend your submission script to trigger a webhook when a job completes.

Autoscaling

Kelpie has an optional autoscaling feature that automates adjusting replica count based on the number of queued jobs, including scale-to-zero when the queue is empty. This feature works through the Salad API, and requires adding the Kelpie user to your Salad Organization to grant the required API access. Currently that is me (shawn.rushefsky@salad.com). To enable autoscaling, you can submit a request to the Kelpie API to establish scaling rules for your container group.

curl -X POST \
  --url https://kelpie.saladexamples.com/scaling-rules \
  --header 'Content-Type: application/json' \
  --header 'Salad-Api-Key: <salad-api-key>' \
  --header 'Salad-Organization: <organization-name>' \
  --header 'Salad-Project: <project-name>' \
  --data '{
    "min_replicas": 0,
    "max_replicas": 10,
    "container_group_id": "<container-group-id>"
}'

Make sure to replace the placeholders with your Salad API key, organization name, project name, and the ID of the container group you created earlier. This will create a scaling rule that will scale the container group to a minimum of 0 replicas and a maximum of 10 replicas. The scaling rules will be applied to all jobs submitted to the container group. The Kelpie scaling algorithm works as follows:

Every 5 minutes, all scaling rules are evaluated.
The number of replicas in a container group is set to equal the number of queued or running jobs, up to the maximum number of replicas, and down to the minimum number of replicas.
If the desired number of replicas is 0, the container group will be stopped.
If the desired number of replicas is greater than 0 and the container group is not currently running, the container group will be started.

Explanation

Tutorials

How-to Guides

Storage

Reference

Wan 2.2 TI2V-5B (Kelpie) Recipe

Overview

Consumer Recipe (single-GPU nodes)

Prerequisites. Prepare Your Storage and Data

Choose and create your S3-compatible bucket

Plan your folder layout (prefixes)

Paths used by the worker

Getting Your Container Group ID

Parameters you can send (Kelpie job `arguments`)

Submit a Job to Kelpie (Text → Video)

Submit a Job to Kelpie (Image → Video)

Alternative image inputs (no S3 download)

Monitor a Job

Autoscaling

Resources

Explanation

Tutorials

How-to Guides

Storage

Reference

​Overview

​Consumer Recipe (single-GPU nodes)

​Prerequisites. Prepare Your Storage and Data

​Choose and create your S3-compatible bucket

​Plan your folder layout (prefixes)

​Paths used by the worker

​Getting Your Container Group ID

​Parameters you can send (Kelpie job arguments)

​Submit a Job to Kelpie (Text → Video)

​Submit a Job to Kelpie (Image → Video)

​Alternative image inputs (no S3 download)

​Monitor a Job

​Autoscaling

​Resources

Overview

Consumer Recipe (single-GPU nodes)

Prerequisites. Prepare Your Storage and Data

Choose and create your S3-compatible bucket

Plan your folder layout (prefixes)

Paths used by the worker

Getting Your Container Group ID

Parameters you can send (Kelpie job `arguments`)

Submit a Job to Kelpie (Text → Video)

Submit a Job to Kelpie (Image → Video)

Alternative image inputs (no S3 download)

Monitor a Job

Autoscaling

Resources