Image Generation: a controlled second paid pass

Image Generation: a controlled second paid pass

An approved draft is useful only if the next agent can reproduce the setup. Keep the prompt, canvas, tier, and seed with the selected image. Then choose the next paid call based on the work that remains.

Pick the next call

• Test one prompt detail on the approved draft with image-generate at

$0.011. Keep the prompt, width, height, tier, and seed.

• Create a premium baseline with image-generate-pro at $0.06. Keep the

prompt, width, height, tier, and resolved model.

• Produce a flagship-model candidate with image-generate-ultra at $0.22.

Keep the prompt, width, height, tier, and resolved model.

image-generate is the repeat-pass default. It returns a hosted PNG URL with the resolved model, final dimensions, prompt, seed, and timing. Set tier: "fast" for routine drafts. creative, anime, and sd35 start a different visual baseline.

{
  "prompt": "the same red bicycle, now parked outside a corner bakery at dusk",
  "width": 1024,
  "height": 1024,
  "tier": "fast",
  "seed": 184729,
  "negative_prompt": "text, watermark, distorted hands"
}

Change the prompt detail under review and leave the canvas, tier, and seed in place. If the returned source is fal-ai, save that provenance with the image. A later Venice run starts a new comparison point.

Move to image-generate-pro when the brief needs a selected premium renderer. balanced starts with Flux 2 Pro; recraft, seedream, grok, and qwen select their respective model lanes. Treat the first premium image as a new candidate, then keep its tier and dimensions with the selected result.

Use the Synthforge MCP server with the same request body when the agent already has MCP configured. Keep the returned URL, model, canvas, tier, and seed beside the approved brief before asking for another paid pass.