# Compare AI models side by side with one prompt

> Feed one Nodespell Prompt node into several image model nodes, run them together, compare results in the Design panel, and keep the best model.

Source: https://docs.nodespell.com/recipes/compare-models-side-by-side/ · Last updated: 2026-11-10

Comparing models in Nodespell means sending the same prompt to several model nodes at once and judging the outputs next to each other. One Prompt node feeds every model, so the only difference between results is the model. You run them together, compare pairs as A and B in the Design panel, and keep the winner connected to the rest of your graph.

**What you need**

- A prompt you want to test.
- A shortlist of two to four models. Use the [model picker](https://docs.nodespell.com/creative-tools/choose-a-model/) to find candidates.
- Credits for one run per model.

**Graph at a glance:** Prompt → Model A, Model B, Model C (in parallel) → Design panel compare → winner → next step

> **Where credits are spent:** Each model node is a separate paid run, so comparing four models costs four runs. On paid plans, hover the run button before you start: it shows the estimate for everything selected and which nodes will generate.

## Stage 1: Write the prompt once

1. Open the Project in **Graph** mode.
2. Press `Tab` and add a **Prompt** node.
3. Double-click it and write the prompt.

See [Prompts and saved text](https://docs.nodespell.com/graph/prompts-and-saved-text/).

## Stage 2: Add the model nodes and connect them

1. Press `Tab` and type a model name, for example `Nano Banana 2` or `GPT Image`, and add the model node. Repeat for each model on your shortlist.
2. Connect the Prompt output to the **Prompt** input of every model node. One output can feed any number of inputs.
3. Select each model node and match the settings that affect the look, such as aspect ratio or image size, so the comparison is fair. Leave model-specific options at their defaults unless you are testing them.

You can also start from a Creative Tool, such as **Image Generator**, and switch its model; but a separate model node per candidate keeps every result on the canvas at once. See [Add nodes](https://docs.nodespell.com/graph/add-nodes/) and [Choose a model](https://docs.nodespell.com/creative-tools/choose-a-model/).

## Stage 3: Run every model at once

1. Select all the model nodes.
2. Hover the run button to see what will happen: each model node is marked to generate, and the Prompt node needs no run.
3. Press `Cmd/Ctrl+Shift+R` or select the run button.

Progress for each node appears in the [Tasks panel](https://docs.nodespell.com/graph/tasks/). To get more than one result from each model, open **Run options** and choose a batch such as **Generate 2 takes**.

## Stage 4: Compare in the Design panel

1. Select exactly two of the model nodes. The Design panel shows them as **A** and **B** in a comparison view.
2. Open the comparison mode menu and choose how to view them: **Horizontal swipe**, **Vertical swipe**, **Side by side**, **Stacked**, **Flip**, **Dissolve**, or **Difference**.
3. Press `X` to swap A and B.
4. Zoom in on fine detail such as text, hands, and edges.
5. Select a different pair to continue. The winner of each pair goes on to the next round.

The comparison view handles two images at a time. For a quick overview of all results, look at the node previews on the canvas. See [Viewers](https://docs.nodespell.com/edit/viewers/).

## Stage 5: Keep the winner

1. Connect the winning node's output to the next step, such as Relight, Image Upscaler, or Image to Video.
2. If the winning node has several takes, open its **Takes**, select **Use Take** on the best one, and choose **Keep this take**.
3. Disconnect or delete the losing model nodes so they are not run again by accident.

## Check the result

- Confirm every model received the same prompt: each model node's **Prompt** input should be connected to the one Prompt node.
- Compare at the same zoom level and in the same comparison mode for every pair.
- Note the run estimate per model if cost matters in your decision.

## Variations

- **Compare video models.** Connect the Prompt and a start-frame image to several image-to-video model nodes. The Design panel compare works for images; open each video take with **View Full Screen** to judge motion.
- **Compare prompts on one model.** Do the reverse: several Prompt nodes, each feeding a copy of the same model node.
- **Compare in Create.** In Create, use **Previous model** and **Next model** in the model picker to try models one after another on one node. Each change on a node that has already run creates a branch, so earlier results stay. See [Revise without starting over](https://docs.nodespell.com/recipes/revise-without-starting-over/).
- **Add references to the test.** Connect the same reference image to every model that accepts one, to compare how well each keeps a subject; see [Keep a character or product consistent](https://docs.nodespell.com/recipes/consistent-character/).

## Frequently asked questions

### How many models can I compare at once?

You can run as many model nodes as you select. The Design panel compares two selected images at a time, so judge them in pairs.

### Why does one model ignore my prompt?

Check that its **Prompt** input is connected to the Prompt node. Some models also have required inputs, such as a source image, and cannot run without them.

### Does rerunning the comparison regenerate every model?

Running a selection makes new takes for the selected model nodes. To refresh only one model, select only that node before you run.

### Can I compare two takes from the same model?

Yes. Open the node's **Takes** to preview each take in turn, or use **Pin to a branch** on one take so it appears as its own node, then select both nodes to compare them.

## Related

- [Choose a model](https://docs.nodespell.com/creative-tools/choose-a-model/) — find and switch models
- [Viewers](https://docs.nodespell.com/edit/viewers/) — Design panel comparison modes and zoom
- [Run nodes](https://docs.nodespell.com/graph/run-nodes/) — run previews, estimates, and batches
- [Takes](https://docs.nodespell.com/graph/takes/) — keep the winning result