Choose the Studio image-to-video model for MCP

Choose image-to-video only when the approved still must remain the starting frame

The current Studio catalog gives a narrow category answer: it lists one image-to-video workflow, Kling 2.6 Pro Motion, at 42 credits. That makes it the catalog entry to investigate when a rights-cleared still is a required input and the requested change is motion over time. It does not prove current runtime availability, an exact Studio ID, accepted ratios, selectable duration, or preservation of the supplied frame.

That workflow distinction should come before model-family preference. If the composition, product placement, character pose, crop, or art direction has already been approved as a still, preserving that artifact as the submitted source makes image-to-video the relevant path. If no still is approved, or if the desired shot needs a new composition rather than movement from an existing frame, move to the text-to-video guide instead. This is a task-fit decision, not a claim that one workflow produces better-looking video.

Treat the 42-credit figure as a current catalog field, not a permanent quote. The catalog itself says its estimates are operating inputs rather than supplier guarantees. Immediately before spending, call the read-only `list_models` tool to confirm that an image-to-video model is still returned, then call `get_credits` to see the current account balance.

Read the Studio model record as a workflow contract

The catalog record supplies three useful screening fields: model name, workflow, and published credit cost. For this page, the decisive public field is `image-to-video`; “Image animation” is the catalog’s stated fit, and 42 credits is the published charge. Do not expand those fields into current availability, an executable ID, a ratio or duration promise, quality ranking, reliability rate, turnaround promise, or source-detail preservation claim.

The underlying fal endpoint is more configurable than the public Studio MCP schema. fal currently documents a required prompt and `start_image_url`, five- or ten-second duration options with five seconds as the default, optional `end_image_url`, a `generate_audio` control that defaults to true, and an MP4 file record in the response. Those provider fields explain the endpoint behind the workflow, but they are not automatically valid MCP arguments. The public Studio `generate` schema exposes model ID, prompt, aspect ratio, and wait; its `upload_input` schema exposes the reserved generation ID, image input, content type, aspect ratio, and wait. Neither schema publishes duration, end-image, negative-prompt, or audio controls.

Current Studio route

Kling 2.6 Pro MotionA public catalog snapshot, not an authenticated runtime observation.

Provider envelope

fal Kling 2.6 Pro image-to-videoThe provider documents start-image, duration, optional end-image, audio, and MP4 output fields.

MCP boundary

Runtime configuration unresolvedRequire an exact returned Studio ID and model-specific compatibility before a paid call.

Use four gates before reserving the generation

First, verify that the still is actually an input, not merely a reference someone expects the model to reinterpret. Second, confirm that the requested result is one continuous motion idea: for example, a locked product shot with drifting steam and a slow push-in. Third, check the live model and credit record. Fourth, obtain explicit approval for one paid attempt and preserve the exact still and prompt used.

A useful decision record is short: `source approved`, `motion brief`, `model returned by list_models`, `published credits`, `aspect ratio`, `one-run approval`, and `stop condition`. The stop condition matters because an output that changes product geometry or composition may call for a revised still or prompt, not an immediate repeat. This record makes the next decision inspectable without pretending that one generation measures model reliability.

Source gate

An approved still must be the required source inputThe workflow category does not guarantee that the output preserves it.

Motion gate

The brief describes change over timeKeep a single dominant camera, subject, or environmental movement.

Spend gate

Recheck model and credits before the runCatalog values are current snapshots, not permanent guarantees.

Approval gate

Authorize one attempt with a stop conditionReview the unedited output before approving another spend.

Compare image-to-video with text-to-video on inputs, not hype

Image-to-video has an extra operational step because Studio must receive the source image. The public package's stdio implementation says `generate` reserves an image-to-video job when upload is required, and `upload_input` uploads the image and submits the reserved generation. Text-to-video queues from the prompt without that source-upload handoff.

Keep the transport boundary explicit. OfflineCreator's installation guide defines `OFFLINECREATOR_UPLOAD_ROOT` only under the local stdio fallback: it bounds local `upload_input` paths and defaults to that process's working directory. The public package schema labels `filePath` as a local path relative to that root and offers `imageBase64` separately. By contrast, remote OAuth setup is URL-only, and the repository README says the hosted backend is not included in the public package. These sources do not establish that the hosted server can read a path on the MCP client's workstation. For a remote connection, inspect the connected tool schema and use a content-transfer method it actually supports; do not send a client-local `filePath` based on the stdio documentation.

The current catalog creates a factual comparison set: Kling 2.6 Pro Motion is image-to-video at 42 credits, Kling 2.6 Pro is text-to-video at 42 credits, and Veo 3.1 Fast is text-to-video at 96 credits. Those fields support workflow and budget screening only. They do not support a cross-model benchmark, a conclusion about visual quality, or an assurance that an image-to-video result will match the source more closely than a text-to-video result matches a prompt.

Choose the Motion route when losing the approved first frame would invalidate the concept. Choose a text-to-video route when the concept is still described primarily in language and a newly composed opening frame is acceptable. If the brief mixes both goals, settle the first frame as an image asset before paying for motion.

Avoid the provider-schema mismatch failure

One configuration mistake is to read the fal endpoint page and send provider-level fields directly to the Studio MCP tool. The fal documentation describes duration, end image, negative prompt, and audio controls, while Studio’s public package schemas do not expose those fields. An agent should not invent them, assume provider defaults are Studio behavior, or report a requested duration or ratio as accepted without model-specific Studio evidence.

Use the schema returned by the connected MCP server as the execution contract. In the documented local stdio fallback, `filePath` is relative to the stdio process's upload root, while `imageBase64` is the package schema's separate in-band image representation. The remote configuration supplies a Streamable HTTP URL rather than launching that local process, and the public repository does not contain the hosted backend. Therefore, a workstation path documented for stdio is not evidence of remote file access. On remote OAuth, use only the image representation exposed and supported by the connected server and client.

For the public stdio package, a future run would use an ID actually returned by `list_models`, a prompt, and only arguments supported for that exact entry. If the response says upload is required, `upload_input` accepts one documented source representation. The package's upload schema validates PNG, JPEG, WebP, or GIF content types and a shared ratio enum, but this draft does not establish which ratio Kling Motion accepts. The lifecycle tools document polling and completed-output retrieval at package level; no named-model completion was observed.

Record the mismatch instead of silently dropping a requested control. For example: “The provider documents ten seconds, but the connected Studio tool does not expose duration; this run will use only the Studio schema.” That statement keeps provider capability separate from product behavior.

Take the next route that matches the unresolved decision

Open the model-guide hub when the task is still model discovery. Use the text-to-video page when there is no approved source frame or when the opening composition may be generated from the prompt. Return here only when the first-frame artifact and the motion brief are both defined; that keeps this page focused on the single current Studio image-to-video choice rather than repeating the broader catalog.

Keep the recommendation inside the available evidence

This page owns category selection: determine whether a rights-cleared source image is a required input, identify the single public catalog entry to investigate, and distinguish it from text-to-video. The model-specific `/models/kling-motion-mcp-guide` page owns the documented reserve-and-upload implementation mechanics. Neither page has observed an authenticated ID, model-specific ratio, Studio duration, successful completion, or output quality.

Do not turn missing or irrelevant community material into consensus, and do not use engagement totals as a substitute for task-matched evidence. Before publishing a broader recommendation, require an attributable source or a preserved source-image, prompt, model record, unedited output, and review decision. Until then, choose the workflow from the approved input, verify the live catalog, obey the connected tool schema, state the local-versus-remote upload boundary, and inspect the result.