Media Creation

When to use

• Generate or edit images.
• Generate or edit videos.
• Generate audio or speech.
• Generate or transform 3D assets.

Workflow

1. Prefer the direct Media AI tools when the distribution exposes them: search_media_models, estimate_media_cost, and run_media_model. If they are absent, use a provider the user has configured or explain what configuration is required.
2. Choose the endpoint and schema.
   • For image generation, default to GPT Image 2 unless the user needs cheaper drafts, a specialized style model, or a non-OpenAI model: openai/gpt-image-2.
   • For image editing, default to GPT Image 2 unless a specialist model is clearly better: openai/gpt-image-2/edit.
   • Use search_media_models when the endpoint or required fields are uncertain.
3. Estimate before spending.
   • Call estimate_media_cost for the chosen endpoint and output count before run_media_model.
   • Tell the user the expected provider cost and any distribution-specific balance charge when both are available.
4. Check balance when cost can climb.
   • For video, 3D, multi-output, high-quality image batches, or budget-sensitive work, inspect the configured provider's limits or balance when that capability is exposed.
   • Mention when retries could compound the cost quickly.
5. Run through the configured tool or provider.
   • Never embed provider credentials in prompts, source files, or generated artifacts.
   • A distribution-provided run_media_model may upload local inputs and charge a hosted balance; describe that behavior before using it.
6. Do not add a hard approval gate on your own.
   • Follow the approval policy of the active runtime and provider.
   • Make the price clear before spending and ask when the active policy requires confirmation.
7. Save outputs into the workspace unless the user explicitly wants remote-only results.
8. If these instructions are already loaded in context, do not spend a shell turn re-reading this SKILL.md from disk.

GPT Image 2

• Use openai/gpt-image-2 for text-to-image work that benefits from high quality, fine typography, product scenes, UI mockups, signage, posters, or detailed realistic images.
• Use openai/gpt-image-2/edit for edits such as replacing objects, changing clothing, cleaning backgrounds, relighting, compositing references, or preserving identity/layout while changing a specific element.
• fal describes GPT Image 2 as OpenAI's latest image model, with text-to-image support for detailed images and fine typography, and edit support for fine-grained changes to images.
• GPT Image 2 supports streaming, commercial use, flexible resolutions up to 4K, and multiple output qualities.
• Text-to-image core input:

{
  "prompt": "Describe the scene, subject, important details, use case, and constraints.",
  "image_size": "landscape_4_3",
  "quality": "high",
  "num_images": 1,
  "output_format": "png"
}

• Edit core input:

{
  "prompt": "Change: exactly what changes. Preserve: identity, pose, lighting, framing, background, geometry, text, and layout. Constraints: no extra objects, no redesign, no watermark.",
  "image_urls": ["./source.png"],
  "image_size": "auto",
  "quality": "high",
  "num_images": 1,
  "output_format": "png"
}

• For masks, add mask_url when the edit should be constrained to a region.
• For multiple references, label the roles in the prompt, for example: "Image 1 is the base scene to preserve. Image 2 is the jacket reference."
• For text inside images, specify exact copy, placement, typography, and constraints such as "Render the text verbatim", "No extra words", and "No duplicate text."
• For edits, keep the prompt narrow: one sentence for what changes, one for what stays locked, and one for physical realism.

Current model examples

• Image examples on fal pages as of April 24, 2026:
• GPT Image 2 text-to-image: openai/gpt-image-2; good default for premium image generation, typography, and realistic detailed scenes.
• GPT Image 2 edit: openai/gpt-image-2/edit; good default for high-quality image edits with clear preserve/change constraints.
• FLUX Kontext Pro: about $0.04/image
• Nano Banana 2: about $0.08/image at 1K
• Nano Banana Pro: about $0.15/image
• Video examples on fal pages as of April 21, 2026:
• Hailuo 02 Standard: about $0.045/second
• Kling 2.5 Turbo Pro: about $0.07/second
• Seedance 2.0 Fast: about $0.2419/second
• Veo 3.1: from about $0.10-$0.20/second on fast/standard 720p/1080p tiers, with higher-cost variants for audio or 4K
• 3D examples on fal pages as of April 21, 2026:
• SAM 3D Objects: about $0.02/generation
• Hunyuan3D v2: about $0.16/generation
• Tripo3D v2.5 image-to-3d: about $0.20-$0.40/generation
• Hunyuan 3D v3.1 Rapid / Pro: about $0.225 / $0.375 plus add-ons

Decision rules

• If the user names a model, still estimate the cost before running it.
• If the user does not name a model, search for candidates and pick one based on quality, latency, editability, and price.
• For image generation or image editing where quality matters more than the cheapest possible draft, start with GPT Image 2.
• For quick ideation, bias toward cheaper image or fast video models.
• For text-heavy graphics, precise editing, or high prompt adherence, favor reasoning-heavy image models after warning about higher cost.
• For long clips, high resolution, audio, or multiple variations, assume the estimate can climb quickly and say so explicitly.

CLI fallback

If the direct Media AI tools are not exposed but the Interpreter CLI is available:

interpreter-app tools builtin-media-ai search_media_models --json '{"query":"GPT Image 2 image editing","limit":5}'
interpreter-app tools builtin-media-ai estimate_media_cost --json '{"endpoint_id":"openai/gpt-image-2/edit","num_outputs":1}'
interpreter-app tools builtin-interpreter interpreter_usage_get --json '{}'
interpreter-app tools builtin-media-ai run_media_model --json-file media-run.json