Runchat Docs

Image models are not interchangeable. One that turns a viewport screenshot into a convincing photograph can be hopeless at laying out a poster, and the cheapest model is the right first try on some tasks. This page tells you which model to start with for a task and which to step up to when the first result falls short. It covers the image models Runchat installs by default and was last updated on 11 October 2026. Credits per image are in brackets. Check the credits page for current prices.

Pick by task

“Start with” is the cheapest model that did the task well in our tests. “Step up to” is the one to try when the first result is not good enough.

Task	Start with	Step up to	What to know
Drafts and batches of options	Flux 2 Klein (4)	Muse Image (20)	Klein holds the camera and takes about 7 seconds. It cannot do text, icons, drawings or style transfer
Render a screenshot with one reference photo	Muse Image (20)	Nano Banana Pro (300)	Muse matched the reference while keeping the view
Render with several references	Muse Image (20)	Nano Banana Pro (300)	Nano Banana Pro keeps the geometry, material and style roles apart best. Nano Banana 2.1 places materials correctly but drops or borrows elements
Render from a written material and lighting spec	Muse Image (20)	GPT Image 2.5 (40)	Muse followed a JSON spec closely. GPT Image 2.5 assigned materials well but shifted furniture and framing
Place a design on a site photo	Muse Image (20)	Nano Banana 2.1 (120)	For Nano Banana 2.1, paint the lot solid rather than outlining it and keep the prompt short. GPT Image 2.5 struggles with scale and integration
Furnish a room with several products	Flux 2 Klein (4)	GPT Image 2.5 (40) or Nano Banana Pro (300)	The step-up models keep each product recognisable and light them consistently
Change one material and nothing else	Nano Banana 2.1 (120)	Ideogram 4.5 Edit with a mask (60)	Muse Image and GPT Image 2.5 also recolour neighbours in the same finish. Use the mask when nothing outside the area may change
Replace a marked part of an image	Nano Banana 2.1 (120)	Nano Banana Pro (300)	Both changed only the painted area and left no paint behind. Muse Image (20) stays in the area but follows the description less closely
Assign several material references to named parts	Muse Image (20)	Nano Banana 2.1 (120)	Nano Banana Pro let one material spill onto a neighbouring wall
An edit with several instructions at once	Nano Banana 2.1 (120)	Run it again	Models often do only part of the list. Check each instruction against the result before changing the prompt
Relight a render	Muse Image (20)	Nano Banana Pro (300)	Muse moved the sun convincingly on a one-line prompt
Enhance a render without changing the design	Nano Banana 2.1 (120) with a detailed prompt	Nano Banana Pro (300)	Name every element that must stay. On a one-line “enhance” prompt Nano Banana 2.1 swapped timber screens for glazing; Nano Banana Pro held the design best
Match a render to a reference photograph	Muse Image (20)	Nano Banana 2.1 (120)	Muse usually holds the design and camera, but changed the camera on one real request where Nano Banana 2.1 kept it. Nano Banana Pro moved the viewpoint or changed the cladding
Orthographic drawings of a design	Muse Image (20)	Nano Banana Pro (300)	Nano Banana Pro retains the most detail across the views
Merge two forms into a new object	Muse Image (20)	GPT Image 2.5 (40) or Nano Banana Pro (300)	Describe the topology of one and the massing of the other, in matte white, or models re-skin instead of morphing
Build a design at full scale in a reference material	Muse Image (20)	GPT Image 2.5 (40)	All three top models did this equally well
Icon sets from a brand kit	Muse Image (20)	Nano Banana Pro (300)	Use GPT Image 2.5 (40) for a full set of 24 in order. Describe how the logo is constructed rather than asking for “the style of” the kit
Posters with exact text	Muse Image (20)	Nano Banana Pro (300)	Read every line of the result. GPT Image 2.5 overlapped body text on a dense poster

The models

Model	Credits	Typical time	Reach for it when	Watch out for
Flux 2 Klein	4	7 s	You want many options fast, a simple view-preserving edit, or simple furnishing	Text, icons, drawings and style transfer. It copies a reference rather than reading it
Muse Image	20	30 s	Almost everything. It had the highest overall score in our benchmark and returns a larger image than most	Recolours same-finish neighbours on a single-material edit. Can change the camera or borrow shapes from a reference when asked to match one
GPT Image 2.5 Sunburst (low)	40	20 s	Furnishing, merging forms, full-scale material studies, complete icon sets	Its renders can read as renders. Recolours neighbouring joinery. Check poster text
Nano Banana 2.1	120	24 s	Replacing a marked area, changing one material, edits with several instructions, a building on a painted site	Redesigns a render on a vague “enhance” prompt. Results vary between runs, so run it again before rewriting the prompt. Six times the price of Muse Image, which does simple edits just as well
Nano Banana Pro (2K)	300	33 s	The finish matters: several references, orthographic drawings, icons, posters, enhancing a render	The price. It once left a site outline in the result and let a material spill onto a neighbouring wall

Runchat runs Nano Banana 2.1 with thinking set to high. In our tests that did not make a good result better, but it left fewer edits incomplete: 2 poor results in 20 requests, against 5 at the medium setting. Ideogram 4.5 Edit (60), Flux 3 Image (100 per megapixel) and Seedream 5.0 Pro (300) are also installed by default and have not been through these tests. Ideogram 4.5 Edit takes a mask and leaves everything outside it untouched, which is why it appears above for single-material changes.

Examples

Four tasks where one model stood clearly apart. Each strip shows the inputs on the top row and the outputs below, ordered by score. The strips come from the September round, so they also show two models that are no longer defaults, and Flux 2 Klein appears under its earlier name, Runchat Image.

Several references: Nano Banana Pro

A line drawing for geometry, a brick building for materials and a 1960s Kodachrome slide for photographic style, with the prompt Render image 1 as a photograph. Materials from image 2, photographic style from image 3. Nano Banana Pro kept the three roles apart. GPT Image 2.5 lowered the viewpoint and underplayed the style.

Courtyard housing drawing rendered with a material reference and a style reference by six models

A building on a site photo: Muse Image

A Revit apartment block and an aerial photo with the lot outlined in red, with the prompt Place the building in image 1 inside the red outline in image 2 and remove the outline. Muse Image put the building in the outline and removed it. Most of the others pasted the render onto the photo without a shadow.

Apartment block placed on an outlined lot in an aerial photo by six models

Furnishing with several products: GPT Image 2.5

An empty sunlit room and three product photos: a sofa, a rug and a side table with a lamp. GPT Image 2.5 and Nano Banana Pro kept all three recognisable at a believable scale and lit them from the window; GPT Image 2.5 does it for 40 credits.

Sunlit room furnished with a sofa, rug and lamp from product photos by six models

Icons from a brand kit: Nano Banana Pro

A brand kit whose logo is drawn from nested parallel lines, and a prompt asking for six icons constructed the same way. Nano Banana Pro built every icon from nested lines with the kit’s swirl worked in. GPT Image 2.5 used heavier strokes and added containers; Muse Image added labels.

Prompting that changes the result

These held across models.

Say what stays the same, not only what changes. Name the elements.
Refer to images by position: “the building in the first image”, “the brick in the second image”.
Mark the area on the image for a local change, then ask for the marks to be removed: Replace the area painted in red with … Keep the rest unchanged and remove the red paint.
Run it again before rewriting the prompt. A second run often fixes a result that ignored part of the instruction.
Describe construction, not style. “In the style of this brand kit” gets generic icons. Describing how the logo is drawn gets icons built the same way.
Neutralise the material when you want form. Asking for matte white stops a model from re-skinning one object in the other’s texture.

How this was tested

Method, dates and limits

Three rounds of testing feed this page.

September 2026, 23 design tasks. Six models, one short and one detailed prompt per capability, reviewed blind from lettered sheets and scored out of 10. GPT Image 2.5 Sunburst was added on 21 September and reviewed with model names visible.
11 October 2026, Nano Banana 2.1. Twelve tests in four jobs (site placement, enhancing a render, replacing part of an image, rendering with several inputs) against Muse Image, Nano Banana Pro and Nano Banana 2, at 1K.
11 October 2026, 20 real edit requests. Nano Banana 2.1 at medium and high thinking, reviewed blind, then compared with the Muse Image result for the same request.

Every test is one generation per model, so single results are indicative. Scores from different rounds were given by different reviewers and are not comparable with each other.Mean scores from the 23-task round, for the models still installed by default:

Model	Credits	Score /10
Muse Image	20	8.1
Nano Banana Pro (2K)	300	8.0
GPT Image 2.5 Sunburst (low)	40	7.5
Flux 2 Klein	4	5.0

Nano Banana 2.1 scored 7.1 on its twelve tests, against 7.4 for Nano Banana Pro and 6.5 for Muse Image on the same tests. On the real edit requests it scored 7.3 at medium thinking and 7.9 at high, against 7.3 for Muse Image.Nano Banana 1, Nano Banana 2 and Flux 2 Pro were part of these tests and are no longer default models. They appear in the image and data below but not in the recommendations above.

Capability matrix showing each model's output and score for each of the 23 tests — The 23-task round at a glance. Rows are tests, columns are models, the number is the score out of 10.

Download the per-test scores, prompts and review notes for the 23-task round.

Next steps

Credits: what each model costs per image
Models: installing and selecting models for media nodes
Rendering from Rhino screenshots: a worked example of rendering a screenshot with a reference

Get Started

Plugins

Clipchat

Core Concepts

Node Reference

Examples

Choosing a model

Pick by task

The models