genmedia models [query] [--category <cat>] [--status active|deprecated|all] [--limit <n>] [--cursor <token>] [--endpoint_id <id,...>] [--expand openapi-3.0] [--json]

Free-text search across 600+ models. Use --endpoint_id to fetch specific models by ID. Use --expand openapi-3.0 to get the full OpenAPI schema. Use --cursor from a previous response to page through results.

schema — inspect model inputs/outputs

genmedia schema <endpoint_id> [--format compact|openapi] [--json]

Shows the input parameters and output shape for a model. --format openapi returns the full raw OpenAPI spec. Always run this before genmedia run to know what parameters the model accepts.

run — execute a model

genmedia run <endpoint_id> --<param> <value> ... [--async] [--logs] [--download [template]] [--json]

Pass any model input parameter as a --flag value pair. Waits for the result by default and returns { status, endpoint_id, request_id, result }.

Use --async for long-running models (video generation, etc.). Returns { status: "submitted", request_id, endpoint_id } immediately. Then poll with genmedia status.

Use --download to save every media URL in the result to disk — see "Downloading output files" below.

status — check or retrieve an async job

genmedia status <endpoint_id> <request_id> [--result] [--logs] [--cancel] [--download [template]] [--json]

Without flags: returns queue position and job status. --result: blocks until complete and returns the full result. --cancel: cancels a queued job. --download: implies --result. Writes every media URL from the result to disk — see below.

Downloading output files

Agents should not curl URLs out of the result payload. Pass --download to run or status instead; the CLI handles extraction, fetching, naming, and concurrency, and emits the saved paths in the JSON output under downloaded_files[].

genmedia run <endpoint_id> ... --download [template] --json
genmedia status <endpoint_id> <request_id> --download [template] --json

Behavior:

--download with no value → save to cwd using the source file name (from file_name, the URL path, or a content_type-derived extension).
--download <dir>/ (trailing /) or an existing directory → save inside that directory with source file names. Parent directories are created recursively.
--download <template> → substitute placeholders and write to the resulting path. Supported placeholders: {index} (0-based walk order), {name} (basename without extension), {ext} (extension without dot), {request_id}.
Plain filename with multiple outputs → first file keeps the name, later files get _1, _2, … suffixes on collision.

Detection rule: the CLI walks the result payload and downloads any object with a url: string starting with http(s)://. This matches every standard fal.ai media output shape (images[], image_urls[], video, audio, etc.).

Successful output (run with --download):

{
  "status": "completed",
  "endpoint_id": "fal-ai/flux/dev",
  "request_id": "...",
  "result": { "...": "untouched" },
  "downloaded_files": [
    { "url": "https://...", "path": "/abs/path/cat.png", "size_bytes": 204800, "json_path": "images[0]" }
  ]
}

If one or more URLs fail to download, the remaining files still land on disk and the failures appear under download_failures[]:

{
  "downloaded_files": [ ... ],
  "download_failures": [
    { "url": "https://...", "json_path": "images[1]", "error": "404 Not Found" }
  ]
}

The command's exit code reflects the model run itself, not individual download failures. When download_failures[] is non-empty, surface the errors to the user and/or retry those URLs.

Guidance:

Default to --download whenever the user expects files on disk. It replaces the old "parse result, then curl each URL" pattern.
For multi-output models (num_images > 1, image_urls[], etc.), use a template with {index} (e.g. ./out/{request_id}_{index}.{ext}) to guarantee unique paths.
--download on status implies --result — you do not need to pass both. Combining --download with --cancel is rejected.

upload — upload a file to fal.ai CDN

genmedia upload <file_path_or_url> [--json]

Accepts a local file path or a remote URL. Returns a CDN URL suitable for use as a model input parameter (e.g. --image_url).

pricing — check model cost

genmedia pricing <endpoint_id> [--json]

docs — search fal.ai documentation

genmedia docs <query> [--json]

Error output

Every command exits non-zero on failure and writes a JSON error object to stderr. The shape is stable across commands:

{
  "error": "<human-readable summary>",
  "details": {
    "endpoint_id": "fal-ai/flux/schnell",
    "request_id": "019d...",
    "status": 422,
    "error_type": "ValidationError" | "ApiError" | "Error",
    "validation_errors": [
      {
        "field": "num_images",
        "message": "Input should be less than or equal to 4",
        "type": "less_than_equal",
        "input": 20
      }
    ],
    "body": { "detail": [ ... raw server payload ... ] },
    "logs": [ { "level": "ERROR", "message": "...", "timestamp": "..." } ]
  }
}

Field meanings:

error — one-line summary you can show the user.
details.status — HTTP status from fal.ai (422 input validation, 401 unauthenticated, 403 forbidden, 404 endpoint not found, 429 rate limited, 5xx upstream). Missing for local errors (network, parsing).
details.error_type — ValidationError for 422 with FastAPI-style detail[]; ApiError for other HTTP failures; Error for local errors.
details.validation_errors — present only when the server returned a FastAPI validation payload. Each entry is { field, message, type, input } and uniquely identifies what the agent needs to fix. Dotted field paths like options.seed or images[0].url point at nested inputs.
details.body — the raw response body, preserved for agents that need the full FastAPI ctx/expected fields or vendor-specific payloads.
details.logs — recent model-side log lines when the failure happened during inference (not for pre-flight validation errors).
details.request_id — included when the request reached fal.ai; pass it to genmedia status <endpoint_id> <request_id> --logs --json for full history.

Agent guidance:

For status: 422, read validation_errors, correct the offending args, and retry. Re-run genmedia schema <endpoint_id> --json if you need allowed enum values or numeric bounds.
For 401/403, check credentials — do not retry.
For 404, verify the endpoint_id with genmedia models ... --json.
For 429 or 5xx, a short backoff-and-retry is acceptable; everything else should be surfaced to the user.

Common workflows

Synchronous inference (fast models)

genmedia schema fal-ai/flux/dev --json
genmedia run fal-ai/flux/dev --prompt "a cat on the moon" --json

Async inference (slow models — video, etc.)

genmedia run fal-ai/veo3.1 --prompt "a dog running" --async --json
# save the request_id from the response, then:
genmedia status fal-ai/veo3.1 <request_id> --result --json

Using a local image as input

genmedia upload ./photo.jpg --json
# use the returned url as input:
genmedia run fal-ai/some-model --image_url <cdn_url> --json

Discovering models for a task

genmedia models "text to video" --json
genmedia schema <endpoint_id> --json

Saving outputs directly to disk

# single file → cwd using source file name
genmedia run fal-ai/flux/dev --prompt "a cat" --download --json

# multiple images → templated paths
genmedia run fal-ai/flux/dev --prompt "a cat" --num_images 4 \
  --download "./out/{request_id}_{index}.{ext}" --json

# async job → download on completion (implies --result)
genmedia run fal-ai/veo3.1 --prompt "a dog running" --async --json
genmedia status fal-ai/veo3.1 <request_id> --download ./videos/ --json

genmedia-ref

genmedia CLI reference

Authentication

Commands

models — search and inspect models

More from fal-ai-community/genmedia-cli

genmedia