Choose Your AI Generation Engine

Flat-to-Model turns flat-lay product photos into photorealistic on-model imagery, and the AI engine doing the rendering shapes how the result looks. Some engines lean clean and editorial, others lean dramatic and creative. You can pick the one whose aesthetic fits your brand, or stay on Auto and let On-Model choose for you.

The Generation model selector lives in the Flat-to-Model review step, alongside the pose, background, and aspect ratio controls. The same choice is available on the API as an optional model field on the request's options object.

The engines

• Auto (default): picks the best model for your set of images and instructions.
• Nano Banana 2: fast, balanced workhorse with clean, photorealistic skin and strong garment fidelity.
• Nano Banana Pro: higher-fidelity option with the strongest garment detail; slower than Nano Banana 2.
• Seedream 4.5: more dramatic lighting and a more creative interpretation of the styling.
• Seedream 5.0 Pro: ByteDance's newest, scoring above 4.5 on both public leaderboards.
• GPT Image 2: distinct rendering style with a slightly warmer overall tone.

Every engine accepts 1K, 2K and 4K requests, but what arrives differs, and it is worth knowing where. GPT Image 2 caps total output at about 8.3 megapixels, so a wide 16:9 render reaches a true 4K while a 3:4 portrait tops out nearer 2450 × 3260. Seedream 5.0 Pro caps lower still, at about 4.6 megapixels, so a 4K request comes back around 2150 pixels square: order it at 2K unless you specifically need the larger file. Seedream 4.5 goes the other way, rendering a true 4K but not honouring the 1K tier at all. The reference-image budgets differ too, from six on Nano Banana Pro to fourteen on Nano Banana 2, with Seedream 5.0 Pro at ten, which matters on jobs that stack a garment, an identity and a pose in one call. Full specifications for every model on the market, with pricing and Elo ratings, are in our index of image generation models.

Pros and cons at a glance

Engine	Output quality	Speed	Best for
Nano Banana 2	High	Medium	Catalog work and the safe default
Nano Banana Pro	Maximum	Medium	Hero shots and premium garment detail
Seedream 4.5	Medium	Fast	Lingerie and intimate fashion
Seedream 5.0 Pro	High	Fast	Stronger scores than 4.5, best ordered at 2K
GPT Image 2	High	Slow	Editorial moods and distinct color tone

Same inputs, different engines, 2K

Here's the same flat-lay trio, the same identity, and the same pose, rendered at 2K. Seedream 5.0 Pro joined the lineup after this comparison was made, so it is not among the four below.

Same garments, same identity, same pose: seated on a wooden stool against a warm ecru plaster wall. Each engine reads the scene differently. Nano Banana 2 stays clean and frontal with the icy blue of the shirt held tight. Nano Banana Pro tightens the garment detail and the fall of the trousers. Seedream 4.5 nudges the model into a three-quarter turn, softens the tones, and pushes the backdrop warmer. GPT Image 2 leans into its own colour palette, with a slightly warmer overall cast. None is universally better. They are different aesthetic defaults, and picking the one that fits your brand catalogue is the point.

Which engine produced each output

Once a job finishes, you can see which engine actually rendered each result.

In the app, a small engine icon sits next to the processing time on each output card. Hover over it for the engine name.

In the API, each result in GET /jobs/<id>/results carries a model_used field:

{
  "job_id": "8c9f1a...",
  "status": "completed",
  "results": [
    {
      "image_index": 0,
      "group_index": 0,
      "model_used": "nano_banana_2",
      "output": { "full_size": "https://..." }
    },
    {
      "image_index": 1,
      "group_index": 0,
      "model_used": "gpt_image",
      "output": { "full_size": "https://..." }
    }
  ]
}

It's useful both for consistency checks and for tracking which engine performs best on which kinds of input over time.

Using it via the API

Pass model inside the options object on your Flat-to-Model request:

{
  "project_id": "...",
  "images": ["uuid1", "uuid2", "uuid3"],
  "identity_code": "default-pro-...",
  "instructions": [ /* ... */ ],
  "options": {
    "model": "nano_banana_pro"
  }
}

Valid values: "auto" (default), "nano_banana_2", "nano_banana_pro", "seedream" (4.5), "seedream_5_pro", or "gpt_image".

In our public Python integration, it's a single CLI flag:

python flat_to_model.py \
  --input-folder SKU/ \
  --identity-code <your-identity-code> \
  --model gpt_image

Full reference is in the Flat-to-Model API docs.

Create-Packshot uses the same lineup

The same six-option Generation model dropdown is available on Create-Packshot. The aesthetic differences described above transfer directly to packshot output: ghost-mannequin, flat-lay, marketing-ready, and white-cutout templates all benefit from engine choice in the same way. See the Create-Packshot guide for an end-to-end walkthrough.

Model-Swap gets a dropdown too

The Generation model selector is also available in the Model-Swap review step.

The default is Onda, PiktID's proprietary swap engine. The alternative is Nano Banana 2, available for teams who want to try a different aesthetic on their swaps. model_used works exactly the same way: every result in GET /jobs/<id>/results tells you which engine produced it.

Pass swap_options.model on a Model-Swap job, or --model onda | nano_banana_2 on the Python integration.

"Different engines are different tools. Brands and agencies who run hundreds of SKUs a week know their aesthetic, and they deserve the choice. Exposing engine selection turns On-Model from a black box into a controllable production pipeline, and model_used on every output makes the choice auditable."

— Nunzio Alexandro Letizia, Co-founder at PiktID and creator of On-Model

Comparing beyond the ones we ship

The engines in this dropdown are the ones that cleared our bar for fashion production: a public API, enough reference images to hold a garment and an identity at once, a resolution ceiling we can state honestly, and content filtering that tolerates ordinary catalogue work.

Plenty of strong models do not clear it. We keep a running comparison of every image generation model worth tracking, with Elo ratings from two independent leaderboards, published price per thousand images, resolution and reference-image limits, and whether each one is available inside On-Model. It is refreshed monthly.

Try it now

Open any Flat-to-Model or Model-Swap job and look for the Generation model dropdown in the review step. Pick an engine, generate, and compare the result against Auto.

New to On-Model? Start with the Flat-to-Model guide, explore presets, or learn about Create-Packshot.