·3 min read
How to Create a Visual Concept with an AI Agent and a Reference Image

You can start a visual concept in TokAtlas by adding a reference image and describing the scene you want to explore. In one test, we uploaded an illustrated capybara and asked the Agent for a square image set in a warm café. The Agent quoted the image request before generating, and a result appeared alongside the reference in the project canvas. Review the output carefully: a reference guides the image but does not guarantee an exact copy.
Prepare an image and a focused brief
Choose a clear image you have permission to use. Decide what should remain recognizable and what you want to change. For our test, the input was an existing capybara illustration. We asked for one square visual that kept the capybara character and moved the setting to a warm café.
Add the reference and ask for an estimate
Open TokAtlas Agent, choose the add-reference control and upload the image. Write your brief in the composer. In this test we explicitly asked the Agent to quote the model and estimated cost and wait for confirmation before generating.
The Agent proposed gpt-image-2.6-sunburst at 1024 × 1024 and estimated US$0.035712 for one image. It said reference-image input costs were not included, so the final charge could differ. This was an estimate for this request, not a fixed price or a guarantee that every request will receive a quote. Check the interface and current pricing before confirming.
Generate and review the result
After confirmation, the project showed the request in a queue. The finished image appeared in the conversation and on the canvas, next to the uploaded reference. The result depicted a capybara in a café with a cup. The image request log recorded US$0.036634 for this generation; Agent language-model calls were billed separately. We did not measure a general average cost or latency.
The Agent message still described the task as queued even after the image appeared. Check the media card and canvas before resubmitting: a status message may not be updated in the same moment as the result.
Ask for one change at a time
To continue the same project, use the generated image’s Add to conversation control and describe the desired change. We requested a light-blue cup while retaining the character and café, and asked for an estimate before proceeding. The Agent estimated US$0.035300 for another image, excluding reference-input charges. We confirmed one edit. The returned image visibly had a light-blue cup and retained the capybara and café setting; this visual check does not establish pixel-perfect preservation. The image request log recorded US$0.006744 for the edit. We have not verified why that recorded charge differs from the estimate. For this test, the two image requests totalled US$0.043378 and 12 associated Agent language-model requests totalled US$0.003830, or US$0.047208 in recorded request costs. Actual costs vary by request, settings and usage.
Questions to check
Does a reference guarantee the exact character, logo or text? No. Inspect the generated image against your source, especially identity, text and small details.
Is the estimate the final charge? Not necessarily. In this test the Agent noted that reference-image input costs were excluded from its estimate; the image log recorded a different amount. Agent calls may have separate costs.
Can I use this for storyboards or video? The interface offers those directions, but this test covered one reference-image request only. Check the relevant workflow and its current requirements separately.
Related reading: Packaging concepts from a product reference and a storyboard from one reference image. These guides describe separate plays, not the Agent workflow tested here.