ScenarioExplorations Nine models, one brief

Click an image to zoom. Use + Compare on two results to put them side by side or under a slider.

Prompt and details

Prompt, sent to all nine

What we saw

Cost of each model

Credits and time, all 18 tests

One image per model per test. Credits are what Scenario charged for each job. Time runs from the moment Scenario started processing the job to the finished result, taken from each job's status history, so waiting in our own batch queue is left out. Click a column to sort.

#Model Credits per image Credits, all 18 tests Median time Fastest – slowest

18 tests, 162 images. Same prompt, same input, every time. Generated on Scenario with the Scenario MCP, 30 September 2026.

Scroll or pinch to zoom · drag to pan · double-click to toggle 100%← → models · Esc close

How it was made

One request, nine models, no cherry-picking.

A series of tests comparing nine image models on Scenario, organised in categories, with a 3×3 grid of results per test. Claude Opus 5.5 in Claude Code ran every generation through the Scenario MCP, then built the grids, the notes and this page from the job records.

  1. 1

    Pick the lineup

    Nine models that can both generate and edit, chosen from Scenario's rankings and newest releases.

    WithScenario MCP recommend, search and model_schema_get · dry runs for every price

  2. 2

    Write one checkable brief per test

    Every prompt has details that can be verified: exact text, counts, left and right sides, layouts. The character tests never list the accessories, so they test consistency, not prompt-following.

    With18 prompts in config.json · the same prompt for all nine models

  3. 3

    Run everything once, at defaults

    One output per model, default quality, no prompt rewriting. Only failed jobs were rerun.

    With207 model generations · 2 Grok retries after server timeouts

  4. 4

    Measure what can be measured

    Texture tiles are repeated 2×2 and given a seam score. Credits and times are read from each job's status history, from the start of processing to the finished result.

    Withseam_check.py · job_get status history for all 162 results

  5. 5

    Review, then redo what was weak

    The restyle got a stronger style, the relight moved to a portrait, the texture became a frozen koi pond, and the turnaround and expression sheets moved to wide canvases with exact layouts.

    Keptfirst attempts archived, not mixed into the page

What did not work

  • Square canvases squeezed four-figure turnarounds, and the first grid script cropped Grok's wide result to a square.
  • The stop-motion restyle was too subtle: three models barely changed the photo.
  • The narration filter refused lines that named models and tools, so those lines were rewritten without names.

Tip

For any multi-panel sheet, give the canvas shape and spell out the exact layout: rows, columns and order.

Credits

2,941 CU

2,669 CU for 207 image generations by the nine models, plus 272 CU for 17 narration clips.

Skill · ZIP

scenario-nine-models-one-brief

Download the skill ⤓

PDF

An illustrated making-of

Open the PDF ↗

Scenario

Make yours

Start free ↗