Skip to main content
Take a screenshot from your 3D model, write one strong prompt, and produce a styled render. Then branch the prompt into variations and apply the locked-in style across other views of the same building. Time: 12 minutes You’ll need: A 3D model in Rhino (or any 3D tool), or just a screenshot to start with
A Rhino linework capture of a gabled house beside the photograph rendered from it

Open the example workflow

A linework capture from our demo Rhino model and the Nano Banana Pro render made from it. Run the node marked (optional) Rhino to capture your own view instead.

​
1. Get a strong baseline render

Start by feeding two things into a Nano Banana Pro node:
  • A clean line drawing screenshot from Rhino (pen linework, not shaded). Colours and shading in your screenshot muddy the output.
  • A prompt that names three things: lighting, material, and a named reference.
Example prompt:
Steal this shape for your own prompts. Swap the lighting, material, and named reference and you have a new look without starting from scratch.

​
What controls the output

  • Aspect ratio: set the output aspect ratio to match your screenshot with More (…) > Size on the node toolbar. Don’t leave it on Auto when your references are a different shape: the model picks up the reference’s shape, then reframes your building or returns the reference instead.
  • Image order depends on the model: Nano Banana Pro and GPT Image are most influenced by the last image, so put the screenshot last. Muse Image is the opposite: put the screenshot first, as in the Model to Render — Architecture example in the example browser.
  • Similar views confuse the model: if a reference shows a view close to your screenshot (another render of your building, or an earlier output), the model may copy it instead of rendering your screenshot. Use the thumbnail technique below on your line drawing.
  • What works depends on the model: see Choosing a model for what each model handles well, such as site placement on Muse Image.
  • Switch to Pro for client work: Nano Banana Pro handles ambiguity better, especially when references and screenshots have different camera angles. The cost is worth it for ship-quality output.

​
2. When the model returns the wrong thing

Two failure modes are common when you’re rendering against a reference: The model returns the reference instead of editing the screenshot. Use the thumbnail technique: paste your screenshot as a small thumbnail inside the reference image. Select the reference, click Edit on its toolbar, paste the screenshot, scale it down into a corner, and click Merge. Now you have a single image with both, and the model pays attention to the main frame while staying influenced by the thumbnail. The model just returns the screenshot unchanged. This happens when the reference and screenshot are very similar. The thumbnail technique fixes this too. The building changes shape or the camera moves. Check the aspect ratio first, then the image order for your model, then whether a reference shows a similar view. A longer prompt with more “do not change” rules rarely helps: change one of those settings and run a single test image. A slider or setting that only rewrites the prompt isn’t a geometry lock, because none of these models has a geometry-strength control.

​
3. Write a multi-step prompt for control

When you need fine control over photographic qualities (film grain, vignetting, light balance), write the prompt as a sequence of steps rather than one sentence:
Why this works: the model internally generates and checks against your prompt. A multi-step prompt gives it something to judge against at each step. When matching photographic qualities, name them specifically:
  • Film grain
  • Vignetting
  • Light and shadow balance
  • Colour palette
  • Activity and people
  • Greenery or context

​
4. Branch the prompt with an agent

Once your baseline render works, generate stylistic variations by feeding the prompt through an Agent node:
A text input holding the baseline prompt, connected to an Agent node that returns a list of four prompt variations
  1. Add an Agent node
  2. Open More (…) on its toolbar and set Output format to List
  3. Connect your original prompt as context
  4. Prompt the agent: Generate 4 variations of this prompt that explore different photographic styles.
  5. Connect the agent’s list output to the prompt of a new image node
  6. Each list item becomes a separate render
Four stylistic variations of the same building, rendered from one line drawing
You now have a series of stylistic options to pick from.

​
5. Carry the style across other views

Once one variation works, lock it in and apply it to other angles of the same building:
  1. Crop the finished render to just the building (remove surroundings) and copy it.
  2. Select a new screenshot from a different angle and click Edit on its toolbar. Paste the render, drag it into the top-left, and click Merge.
  3. Prompt Nano Banana Pro: “Apply the style from the top left to the main image. The screengrab shows the same building from a different angle. The vegetation and surroundings must match the new camera position.”
Keep this second prompt simpler than the first. The hard work was done in the baseline.
Rotating camera angles is one of the hardest things for any image model. Style-transferring many views in one batch usually fails. Work through them one at a time, tweaking prompts per view.

​
Quick tips

  • Type @ to inline reference any node’s output into a prompt. See Connecting Nodes.
  • Alt-drag to duplicate a node.
  • Crop references before collaging. Less clutter, more focus.
  • Square aspect ratios for cropped references. Wild aspect ratios produce worse results.
  • Workshop long prompts with the agent when one isn’t producing what you want.

​
Next steps