Generate product-scene concepts with MCP

Set a concept-only input contract

Use this workflow to structure a product-scene concept, not to reproduce approved packaging. The generate tool in the current published MCP package, version 0.1.3, accepts a model ID, an optional prompt of up to 2,000 characters, an optional aspect ratio from a shared six-value list, option values declared by list_models, an optional locked quote token, and an optional wait flag. It has no image field of its own; files attach through separate upload tools to input slots a model declares. The text-to-image models compared below list a prompt as their only input, and the shared ratio list does not establish support for a named model, so a prompt-only request cannot anchor exact product geometry to a supplied pack shot.

Before prompting, separate the fixed facts from the creative variables. Record the product category, approximate silhouette, material, palette, viewing angle, background, surface, lighting, and intended crop. Put protected elements such as a real logo, legal copy, ingredients, certification marks, and exact label layout on a do-not-invent list. The generated image is then a scene-direction proposal for review, not a product record or production-ready pack shot.

Put human approval at the fidelity boundary

Approve the concept only after checking silhouette, proportions, closure or hardware, label placement, text, logo shape, color, reflections, contact shadows, and crop. Recraft documents that generated text can be repositioned or restyled, very small text may be distorted or misspelled, and text generated inside a frame is not editable afterward. Those limitations make a visual inspection mandatory whenever a scene contains packaging copy.

If the concept is approved but the product identity is not, move to compositing or a mockup workflow. Recraft documents a separate mockup process that can place artwork, a logo, or a design onto product photography and can convert existing photorealistic imagery into a mockup frame. That is a downstream editing option, not a capability exposed by OfflineCreator's current MCP generate tool, so keep the handoff explicit.

Choose the image model by the question you need answered

OfflineCreator Studio's model catalog lists more than ten text-to-image models. Three with distinct catalog labels frame the choice for this workflow: FLUX Schnell at 1 credit for fast iteration, FLUX 1.1 Pro Ultra at 8 credits for photoreal detail, and Recraft V3 at 5 credits for brand graphics, each listing a prompt as its only input. Those are catalog positions and prices, not comparative quality guarantees or an observed recommendation for a particular product scene.

Use the labels only to frame a human selection question, then require an exact ID and model-specific compatibility from the connected Studio service before execution. A future proof should change one declared scene variable at a time, but no controlled output set was created here and no model was shown to improve the decision.

Read the proof samples without overclaiming

No task-matched product-scene sample set supports this workflow. The catalog entries cited above are labels, prices, and declared inputs rather than outputs, so they cannot show repeatability, comparative quality, or preservation of a particular package. Treat example media found elsewhere, including community posts, as heterogeneous reference material unless it discloses the prompt, model, settings, and review protocol behind it.

Black Forest Labs documents that FLUX 1.1 Pro Ultra can generate up to 4-megapixel images at varying aspect ratios and can use an image prompt with adjustable influence in its own API. OfflineCreator's catalog lists that model with a prompt as its only input, and the published MCP generate tool has no image field, so do not infer that every upstream model parameter is available through MCP. Check the model's declared inputs and options from list_models before a paid run, and judge the returned pixels, not the upstream feature list.

Write a scene prompt that can be reviewed

Build the prompt in a stable order: product category and generic form; camera position and crop; support surface; environment; light source, quality, direction, and temperature; palette; material cues; then exclusions expressed as positive constraints. Black Forest Labs advises describing photographic choices such as camera, lens, and lighting rather than asking only for a professional photo. Its guidance also says quoted text should be short because long strings are harder to render accurately.

For example: “Unbranded amber glass skincare bottle with a blank matte label, three-quarter view on pale limestone, soft window light from camera left, restrained warm-gray bathroom background, realistic glass reflections, square product photograph, no readable text, no logo, no certification marks.” Keep the product generic when exact identity matters. After choosing the scene, composite the approved pack shot and controlled typography in a deterministic design tool instead of treating generated lettering as final.

Choose the next action by what failed

Return to the creative workflow directory when the deliverable is still undecided. Move to the CLI variants guide when you have an approved direction and need a controlled batch. Use neither route to claim exact packaging fidelity: preserve the approved product artwork outside generation and document the final compositing handoff.

What this article covers

This article explains the operator procedure for “mcp product scene concepts”: define a generic scene contract, use catalog labels to frame model discovery, draft a reviewable text prompt, and stop before unsupported runtime or fidelity claims. `/use-cases/ecommerce-teams` instead owns team roles, canonical product records, approval, and production handoff. The public gallery is not a task-matched sample set, so keep this workflow draft and consolidate actionable guidance into the parent until its real-sample requirement can be met.