Generating media

Every model belongs to a mode (image / video / audio) and an input type that describes what goes in and what comes out. Understanding the input type tells you which inputs a model needs.

Input types

The two-letter codes you'll see across the catalog and in picsart_model_catalog:

Code	Meaning	Typical inputs
t2i	text → image	prompt
i2i	image → image	prompt + image
t2v	text → video	prompt
i2v	image → video	prompt + image
v2v	video → video	prompt + video
a2v	audio → video	prompt + audio
tts	text → speech	text + voice
sts	speech → speech	audio
sfx	sound effects	prompt
music	music	prompt

Providing inputs

CLI — pass files as local paths or URLs. Local files are uploaded for you before the request is submitted:

# image-to-video: animate a still
gen-ai generate -m wan-2.7-i2v -p "gentle parallax, drifting clouds" -i ./hero.png

# video-to-video: restyle a clip
gen-ai generate -m seedance-2.0-video-edit -p "claymation style" --video ./clip.mp4

MCP — pass imageUrls / videoUrl (must be reachable URLs):

{ "name": "picsart_generate",
  "arguments": { "model": "wan-2.7-i2v", "prompt": "gentle parallax", "imageUrls": ["https://example.com/hero.png"] } }

Model-specific parameters

Beyond the common params (aspectRatio, resolution, duration, count…), each model has its own surface — for example, Veo 3.1 accepts startFrame/endFrame keyframes and generateAudio, while ElevenLabs models accept voiceId and language.

Discover the exact surface:

gen-ai models info veo-3.1 --json          # full paramConfig
gen-ai validate -m veo-3.1 --schema        # the model's parameter schema as JSON

{ "name": "picsart_model_params", "arguments": { "model": "veo-3.1" } }

In the CLI, model-specific params are flags (--start-frame, --voice, …). Via MCP, pass them in the extra object.

Outputs

A generation returns one or more result URLs. Control delivery:

• CLI: downloaded to ./output by default (or --download <dir>) and saved to Picsart Drive by default (--no-save-to-drive to skip, --drive-folder to choose the folder). -q prints just the URL; --json prints { url, model, results, durationMs }.
• MCP: assets: [{ id, type, url, ... }] plus a resource_link per output. Media models run asynchronously by default: picsart_generate returns a job handle and the agent collects the result with picsart_job_status. Results are saved to Drive by default (saveToDrive: false to skip).

Some models return multiple results from one call (e.g. an Explore model returning several images) — the result items array simply has more than one entry.

Validate before you spend

echo '{"prompt":"test","duration":99}' | gen-ai validate -m seedance-2.0

{ "name": "picsart_preflight", "arguments": { "model": "seedance-2.0", "params": { "duration": 99 } } }

FAQ

How do I know which input type a model needs?

Run gen-ai models info <id> — it shows the input type code (t2i, i2v, etc.) and which inputs are required. Via MCP, call picsart_model_params.

Can I pass a local file as an input image via MCP?

MCP tools require URLs, not local file paths. Ask the agent to open the Picsart uploader (picsart_media_upload), or upload with picsart_drive (action: upload, returns a Drive URL), then pass that URL as imageUrls in picsart_generate. See Local files → URLs.

What aspect ratios are available?

Common values are 1:1, 4:3, 3:4, 16:9, 9:16. Not every model supports every ratio. Run gen-ai models info <id> to see the supported aspect ratios for a specific model.

Can I generate without a prompt?

Most models require a prompt. Some image-to-image models treat the prompt as optional — run gen-ai models info <id> to check. If you pass an empty prompt, the model may refuse or use a default.

What does enhancePrompt do?

When set to true, your prompt is rewritten by an LLM before being sent to the generation model. This can improve results for short or vague prompts, but it changes what is sent, so disable it if you need precise prompt control.

How many outputs can I request per call?

It depends on the model: many image models accept a count parameter (the CLI's -n takes 1–10, and each model narrows that), while others always return one output. Check picsart_model_params or gen-ai models info <id> for the specific model.