Compare image models by the file you still have to ship

Compare image models by the file a brief needs. A useful round uses one prompt, one aspect ratio every candidate can select, the same reference order, and a pass rule you can check by looking. This page is the plan. Scores belong in your notes after you run it.

The named candidates are GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, Nano Banana 2, the Seedream 5.0 variants, plus FLUX 2 Pro and Qwen-Image-2.0 when you are testing outside this catalog. Flare and Sunburst, the GPT Image 2.5 editing guide, GPT Image 2.5 beside GPT Image 2, and Seedream 5.0 Pro beside Nano Banana 2 explain those products. They are not a substitute for a round you run yourself.

Name each candidate by its model id

On OmniArt, GPT Image 2.5 Flare (gpt-image-2.5-flare) is the default image model, and GPT Image 2.5 Sunburst (gpt-image-2.5-sunburst) sits beside it. Both list 1080p, 1440p, and 2160p, Quality settings from Low through Max, up to 16 image inputs, and the ratios 1:1, 16:9, 9:16, 3:2, and 2:3. The catalog describes Flare as fast everyday generation and Sunburst as precise generation and editing. Those are model-card labels. The round checks whether a brief matches them.

Nano Banana 2 is gemini-3.1-flash: 512p, 1080p, 1440p, and 2160p, up to 9 image inputs. Nano Banana (gemini-2.5-flash), Nano Banana 2 Lite (gemini-3.1-flash-lite, 1080p, up to 14 inputs), and Nano Banana Pro (gemini-3.0) are other registrations. Seedream 5.0 Pro lists 1080p and 1440p with up to 10 references. Seedream 5.0 Lite lists 1440p, 1800p, and 2160p with up to 6. Seedream 4.5 and Seedream 4.0 are older ids. "Seedream" or "Nano Banana" alone does not name the run.

Keep outside names in their own rows

FLUX 2 Pro does not appear in this public catalog. Record an outside test with the product name and the date, in a row of its own. The registered model qwen-image is Qwen Image: 720p and 1080p, standard ratios, seed, and up to three image inputs. That picker entry is Qwen Image. Qwen-Image-2.0 is a different name, and this catalog does not register it.

A provider route is separate from the picker. Nano Banana 2 uses fal-ai/nano-banana-2. Qwen Image's reviewed fal-ai/qwen-image route is a narrower text-to-image mapping than the picker card. Neither route name is a version.

Hold fixed only settings every model lists

Group candidates by settings they can actually share. If there is no shared resolution, run separate groups and record the real output dimensions rather than pretending the settings were identical. GPT Image 2.5 does not list 4:3 or 21:9, so a 4:3 test that includes Flare is already two briefs. Use a shared ratio such as 1:1 or 16:9, one image count, and the same prompt text. Stop a reference round at the smallest input cap in the group: 16 for Flare and Sunburst, 9 for Nano Banana 2, 10 for Seedream 5.0 Pro, 6 for Seedream 5.0 Lite, and 3 for Qwen Image. Ten references against three is a different assignment.

Five briefs, with the pass rule written first

Read each file against its pass rule

Run the label card when the deliverable has words. Run the product photo when identity matters more than mood. Run the reference set when more than one input has to survive, using the cap from the section above. Output suitability is the frame. A 16:9 still misses a 9:16 placement even when the subject looks right. The version line passes when the note matches the picker id, including when a filename says "2.0" or "Pro" and the id says otherwise.

Getting started on OmniArt

Open the image workspace, pick one brief from the table, and duplicate it once per model id you can select. Write the id beside the file before you judge it. Add FLUX 2 Pro or Qwen-Image-2.0 only on a row run outside this catalog. Leave it blank until that run exists.