Pick an image model family for the job

Connect the Scenario MCP

Connect your agent and sign in once at mcp.scenario.com. Then tell it which Scenario team and project to use, so every file lands there.

Give the agent the files

The tank hero made in Generate images with GPT Image 2.5 Sunburst: give the agent its asset ID. It's the image to compare against and the reference for the second job.

Shortlist with recommend

recommend ranks models for a task by measured quality, speed and cost, and is free. Krea 3 is too new for some lists, so the agent looks it up by name.

Use recommend to rank models for text to image, quality first, and give me the top three with what each is best at and its cost tier. Then find Krea 3 with search and give me its model ID. Don't run anything yet.

Brief the agent: same prompt, new model

Only the model changes: the prompt is read from the first image, and the size matches it at 16:9.

Read the prompt of the tank hero with asset_get. Run it word for word with Krea 3, one image, at 16:9 (1024 x 576), with its Berry (quality) engine and Reasoning Level low, no negative prompt and no seed. Read the parameter names with model_schema_get first. Quote the cost with a dry run first and wait for my go. Then show me the Krea 3 image next to the GPT Image 2.5 Sunburst one at the same size.

The second job: same character, new scene

An edit from an image needs a model that reads references and keeps a character. Accept when the armor, shield and pose are kept in the new scene; reject a changed character.

Use Nano Banana 2.1 (model_google-gemini-nano-banana-2-1), one image, with referenceImages set to an array holding only the tank hero's asset ID. Quote the cost with a dry run first and wait for my go.

Prompt, word for word:
The same character standing on a rocky canyon ridge at sunrise. Keep his armor, shield and pose.

File the test and report

Create a collection called "Model tests: tank hero" and add the original, the Krea 3 image and the Nano Banana 2.1 image to it. Report each image's model, settings, asset ID and cost, and the total cost spent.

Open the model library

Create › Images, click the model's name at the top of the panel, then View model library at the bottom of the list.

Answer the three questions

Tick the job under Capabilities (Text to Image), what you start from under Input kind (here only a prompt), and set Budget and Speed under Budget and speed.

Try a new model

Tick New · last 30 days under Shortlists, hover Krea 3 to read its rank, typical time, cost tier and samples, and click Use this model.

Generate with the same prompt and size

The prompt stays when you switch models. Krea 3: Model Berry (quality), Image Size 16:9 (1024x576) to match the first image, Reasoning Level Low, Negative Prompt and Seed empty. Click Generate.

Compare the two images

Select the Krea 3 image and the GPT Image 2.5 Sunburst one with the circle in each corner, click Compare Images, then Split and drag the line.

Find a model for another job

Clear the filters, tick Image under Input kind and Edit with Prompts under Capabilities, and use Nano Banana 2.1, which keeps a character as he is.

Add the reference and generate

Reference images › Select from Library: the character from the first lesson, Confirm. Replace the prompt with what changes and what must stay, then click Generate.

The same character standing on a rocky canyon ridge at sunrise. Keep his armor, shield and pose.