Choosing a model

Choose the operation first: write a scene, create an image, animate a first frame, generate with references, edit an existing clip, or create a separate audio track. Then choose a model whose connected mode accepts those inputs. A famous model name or a higher quality tier does not establish that it fits your task.

Choose and quote through MCP

For live video or image selection, use describe_studio → model_advice → price_quote before a paid run. describe_studio provides the current catalog and controls. model_advice explains the models and their scoped observations; select video or image and optionally a model ID. Then use price_quote for the exact model, mode, quality and inputs you intend to run.

For an individual operation, quote the same parameters you will submit. For run_node, first inspect the saved node’s model, mode, quality and inputs, then quote its node ID and run mode; the quote includes required upstream work. Review the estimate and warnings, agree a max_cost_usd ceiling, and quote again if settings or inputs change. Follow Run a node through MCP or individual media operations for the execution steps.

Find the current model contract

The tables on each model page are generated from Vini’s current model and selector contracts. They distinguish a new-node default from a fallback for older saved nodes. Model-author documentation describes a wider product; it does not add controls to Vini.

Family	Model pages
Video	Kling, Seedance 2.0, Seedance 2.5, Grok Video 1.5, Gemini Omni Flash, MiniMax H3, HappyHorse 1.1, FLUX 3, Wan, DreamActor 2, OmniHuman 2
Image	Seedream 4.5, Nano Banana 2.1, Nano Banana Pro, GPT Image 2.5, FLUX.2, FLUX.1 Kontext, Seedream 5.0 Pro, Seedream 5.0 Lite, Grok Image 2
Text	Text and authoring models
Music	Music models
Speech	Voice models
Mouth animation	Lip-sync models
Processing	Upscaling, transcription

Read availability and outcomes separately

Label	Meaning
Documented	There is guidance explaining the task and its boundaries. Documentation alone is not a successful render.
Supported	The stated Studio input, mode or authoring route accepts the request. For a camera technique, this may only mean it can be requested in prose.
Not connected	The model author may offer the capability, but this Studio path does not expose it.
Unsupported	The stated Studio contract excludes that input or operation. This is narrower than a claim that the model can never perform a visual effect.
Unknown	The contract or provenance is not sufficient to decide.
Empirically tested	A dated render was inspected for the stated criterion. Its outcome may be success, partial, failed or refused.
Untested	No matching outcome evidence is claimed. Absence of a test is not unsupported behavior.

Keep model/version, mode, quality, input role and date with a result. A Fast result does not certify Full or Draft. A prior Nano Banana or GPT Image version does not certify its replacement. An input refusal is different from a completed clip that failed the visual criterion. Automated prompt checks establish the request contract, not the rendered behavior.

A practical selection example

For a recognizable gardener opening a greenhouse, first decide whether the opening composition must exactly match an approved image. If so, use a first-frame workflow. If appearance matters but the model may compose the scene, consider references. If the clip already exists and only the coat must change, use an edit workflow. Read the relevant model’s example and failure notes before the first short generation.

For camera direction, start with one static shot or push-in. For speech, plan dialogue or separate narration explicitly. Adding more techniques at once makes a failure harder to diagnose.

Older names and saved projects

Compatibility aliases keep some saved model selections loadable; they are not extra independently available models. The generated model facts list relevant aliases. Stable Audio 2.5 remains offered as a previous default. Claude Sonnet 4.6 is saved-node legacy, not a new-node option. Removed models and the archived Telegram/n8n pipeline are not current Studio choices.

See the model–technique applicability matrix for connected operations, dated outcomes and configurations that remain untested.

Reviewed 15 October 2026. This catalog accounts for current selectors and fixed processing services. It does not promise provider uptime, exact output reproducibility, or a successful result for every model–technique combination. Use the MCP support guide before assuming a UI operation can run through an agent.