Architecture Knowledge HubArchitecture Knowledge Hub

Part 4 of 8: Four knobs that make Firefly output repeatable

By Pushparajan Ramar · September 23, 2026

Part 4 of 8: Four knobs that make Firefly output repeatable

Series: Firefly Services for developers. Checked against Adobe's documentation on September 23, 2026.

TL;DR

  • The parameter is seeds, and it's an array. Every output returns its seed. Store it.
  • negativePrompt is the field for exclusions.
  • Structure and style references each take a strength. Structure strength runs 1 to 100, default 50.
  • visualIntensity is a separate control.

Why this matters

Say marketing approves a look for an autumn campaign, and you change "wool sweater" to "cashmere sweater." Without a fixed seed, the whole scene can change with it: camera, light, background. Determinism controls let you change one thing at a time.

1. Seeds

A seed is a starting point for generation. Adobe's docs say using a seed lets you generate consistent images across calls. Its SDK reference gives the example of reusing a seed to get a similar image in a different style.

Details worth knowing:

  • The request field is seeds, an array of integers.
  • numVariations defaults to the number of seeds, or to 1 if you send none.
  • If you send both, the number of seeds must equal numVariations.
  • The seed for each output is returned in the finished job:
"outputs": [{"seed": 2142812600, "image": {"url": "https://..."}}]

A request that pins two variations (shape adapted from Adobe's custom-models tutorial):

{
  "prompt": "A wool sweater on a mannequin, autumn light",
  "numVariations": 2,
  "seeds": [66080, 82683],
  "size": {"width": 2048, "height": 2048},
  "contentClass": "photo"
}

flowchart LR
    A[prompt A + seed S] --> X[image A]
    B[prompt B + seed S] --> Y[image B: similar layout]

What to store next to each asset in your DAM: the seed, the full prompt and negative prompt, the model header value, any customModelId, and the request timestamp. Without the seed you can't rebuild a composition later.

2. negativePrompt

negativePrompt tells Firefly what to leave out. Writing "no snow" in the main prompt tends to pull toward snow, because the words in a prompt describe what you want to see. Put exclusions in the field built for them:

{
  "prompt": "A serene mountain valley with a clean river",
  "negativePrompt": "snow, winter, power lines"
}

Adobe's docs define the parameter and list it among the Image5 control fields. The explanation of why exclusions in the main prompt backfire is general behavior of text-to-image models, not something I found in Adobe's documentation.

3. Structure and style references

Both take an imageReference whose source is either an uploadId or a signed URL from an approved domain.

{
  "prompt": "a photo of a volcano",
  "structure": {
    "imageReference": {"source": {"uploadId": "<id>"}},
    "strength": 50
  },
  "style": {
    "imageReference": {"source": {"uploadId": "<id>"}},
    "presets": ["painting"],
    "strength": 50
  }
}
  • Structure reference: applies structural characteristics such as outline and depth. strength is 1 to 100; when omitted, the default is 50. Adobe's guide says it controls how closely the result resembles the reference.
  • Style reference: steers the look. It sits alongside style presets, and the strength value influences how much they apply.

Lower values let the prompt lead; higher values let the reference lead.

4. visualIntensity

visualIntensity is its own parameter. It appears in Adobe's request schema and tutorials (samples use values like 2 and 6). Keep it separate in your mental model from reference strength. I did not verify its allowed range.

Which knob for what

Goal Use
Rebuild a shot seeds
Leave things out negativePrompt
Keep the layout structure.strength
Match a look style.imageReference
Tune the overall look visualIntensity

Sources

Carousel

Download the PDF carousel