Genmedia

genmedia workflow

Use this skill when the user wants to execute a fal.ai model — either by task description (e.g. "generate a video of a dog running") or by a specific endpoint_id (e.g. fal-ai/flux/dev). Load genmedia-ref alongside this skill for the full command reference.

Steps

1. Discover — If no endpoint_id is given, search for a suitable model:

   genmedia models "<task>" --json
   

2. Inspect — Get the model's input parameters:

   genmedia schema <endpoint_id> --json
   

   Read all required fields before running.

3. Upload files (if inputs include images/video/audio):

   genmedia upload <local_file_or_url> --json
   

   Use the returned url as the parameter value.

4. Run the model:

   • Fast model (completes in seconds):

     genmedia run <endpoint_id> --<param> <value> ... --json
     

   • Slow model (video generation, large jobs):

     genmedia run <endpoint_id> --<param> <value> ... --async --json
     genmedia status <endpoint_id> <request_id> --result --json
     

5. Save outputs — when the user expects files on disk, add --download to run or status. The CLI writes every media URL from the result to the filesystem and returns the local paths in downloaded_files[]. Do not curl the URLs yourself; use the flag.

   genmedia run fal-ai/flux/dev --prompt "a cat" --download --json                            # cwd, source file names
   genmedia run fal-ai/flux/dev --prompt "a cat" --num_images 3 --download "./out/{index}.{ext}" --json
   genmedia status <endpoint_id> <request_id> --download ./out/ --json                        # implies --result
   

   Use {index}, {name}, {ext}, {request_id} placeholders in the template when the model returns multiple files (images[], image_urls[], etc.) to avoid filename collisions. A trailing / or an existing directory path saves files under that directory using their source names.

6. Return the result to the user. If --download was used, reference the paths from downloaded_files[]; otherwise present the URLs from result clearly.

Assets library (genmedia assets ...)

The fal Assets API lets you persist generated media, organize it into collections, and define reusable characters.

Two ways to identify an asset

• vector_id — present for any media in the user's library or in a semantic-search result. Returned by assets browse / assets get / assets upload. Used both as a flag on mutations and as the positional argument for assets get and assets tags for-asset.
• request_id — the generation the user just ran. Returned by genmedia run and genmedia status.

Which to pass. Use --vector_id if you have it (anything in the library or a search result has one); otherwise use --request_id for a fresh generation.

Use these flags on every command that takes an asset target: assets favorite, assets unfavorite, assets tags assign / unassign / set, assets collections add / remove, assets characters create --reference_image <id>.

assets upload is for external media only

Use assets upload only for local files or non-fal URLs — media that isn't already in fal's system. For anything that came out of a genmedia run, use the request_id instead.

Cheatsheet

What you have	Use
vector_id (from assets browse / get / upload)	--vector_id <id>
request_id (from genmedia run / status)	--request_id <id>
Local file or non-fal URL	assets upload <path_or_url>

Example — character from a fresh generation

RUN=$(genmedia run fal-ai/flux/schnell --prompt "an elegant black cat" --json)
REQ=$(echo "$RUN" | jq -r .request_id)

genmedia assets characters create "Whiskers" \
  --description "An elegant black cat sitting on a moss-covered rock" \
  --reference_image "$REQ" \
  --identifier whiskers \
  --json

Same pattern for assets collections add --request_id <id> and assets favorite --request_id <id>. Newly-referenced media — whether by request_id or by a vector_id from a fresh upload — may take a moment to appear in assets browse / semantic search while the embedding finishes.

For the full command tree, run genmedia assets --help or any subcommand with --help.

Handling errors

When a command exits non-zero, it prints a JSON error object to stderr:

{
  "error": "Validation error — num_images: Input should be less than or equal to 4",
  "details": {
    "endpoint_id": "fal-ai/flux/schnell",
    "request_id": "019d...",
    "status": 422,
    "error_type": "ValidationError",
    "validation_errors": [
      { "field": "num_images", "message": "Input should be less than or equal to 4", "type": "less_than_equal", "input": 20 }
    ],
    "body": { "detail": [ ... raw FastAPI payload ... ] }
  }
}

• status: 422 / error_type: "ValidationError" means the inputs violated the model schema. Read details.validation_errors — each entry names the field, a human message, the validator type, and the input that was rejected. Re-run genmedia schema <endpoint_id> --json if you need the allowed values, fix the offending args, and retry.
• Other statuses (401 auth, 403 forbidden, 404 endpoint not found, 429 rate limited, 5xx upstream) surface the server message directly in error and do not contain validation_errors. Do not retry 4xx errors blindly — correct the request first.
• For in-flight model failures, details.logs contains the most recent model-side log lines.

Notes

• Always use --json so output is machine-readable.
• Run genmedia pricing <endpoint_id> --json first if cost is a concern.
• If unsure which model to pick, run genmedia docs "<task>" --json for guidance from fal.ai documentation.