Future Vision Image Generation

Problem to solve

When integrating image generation with Future Vision, the harder part is not sending one request, but passing prompt, model, size, reference-image order, and task state to the backend reliably. A successful task creation only means the task entered the queue; treating it as finished too early can produce incorrect output.

How the skill works

• Input normalization: requests are organized around prompt, modelType, and size; display names such as FE-image, FE-banana, and doubao-5.0 map to normal, 2k, and 4k.
• Task creation: the skill calls the Future Vision custom image generation API to create a custom_image_task, where the backend deducts points and queues the job; creation success only means enqueued.
• Reference images: supports a single image or ordered inputImages; the skill preserves order and avoids collapsing multiple references into one URL or reordering them.
• Status tracking: final state is checked with custom-image get --task-id; custom-image queue is only for queue diagnostics. If not finished, [SORA_TRACK] is returned so frontends can keep polling.

Boundaries

• Use it only for image generation or image editing, not video or image-to-video; route video needs to future-vision-video-generation.
• Resolution and size options are controlled by backend model configuration; if the backend rejects a value, show the error and ask the user to choose a supported option.
• When a required parameter is missing, ask for only one field at a time, rather than asking for model and size together.