Nano Banana 2.1: Model Capabilities, Quality & Performance

Nano Banana 2.1 is an image generation and conversational editing model. It updates Nano Banana 2 with improved realism, instruction following, text rendering and character consistency across edits. It combines image creation with understanding of text and visual references. Official model overview.

Use it to turn a brief into a visual, revise an existing photograph or bring several references into one composition. The most useful evaluation is whether it preserves your requirements while producing an image that works at its final viewing size.

Describe the subject, setting, action, style and composition together. Add concrete lighting and framing instructions when they determine the result: a soft light from the left, a full-body subject, or clear space for a headline. The official prompt guide recommends these elements for controlling the image. Image prompting guide.

The documented improvements include text rendering and infographic layout accuracy. Official capability documentation.

For posters and social graphics, quote the exact words and describe their placement and hierarchy. Keep essential copy short. Proofread dates, URLs, numbers and punctuation at full size, then check readability at the size the audience will actually see. For informational graphics, verify the facts before generating the design.

Use an existing image as the starting point, name the change and list the details that must stay. A background replacement, lighting adjustment or color variation is easier to evaluate when it has a clear preservation checklist. Work from the original reference when repeated edits begin to drift.

For a person, track face shape, hair, glasses and distinctive features. For a product, track geometry, hardware, labels and material. Judge consistency across several scenes rather than accepting one attractive output as proof of reliable identity preservation.

The model documentation lists up to 14 reference images, with consistency for up to four characters and fidelity for up to ten objects. Reference-image capabilities.

Assign each reference one purpose: image 1 supplies the product, image 2 the environment, image 3 the style. Start with two inputs and add references after checking role separation. Inspect object counts, relative scale and shared lighting; uploading more references does not by itself improve the result.

Output options include 1K, 2K and 4K. The documented update fixes tiling artifacts on extreme wide and tall ratios—1:4, 4:1, 1:8 and 8:1—at 2K and 4K. Resolution and format updates.

Choose the composition first, then inspect a larger export. Look for repeated textures across panoramic canvases and inconsistent fine detail in skin, fabric and lettering. More pixels help only when the image remains coherent.

These are human-preference Elo results, not accuracy percentages. Scores compare models within the same evaluation; different rows are different tasks. The settings shown do not measure the exact Pixwit workflow. Official model card and methodology.

The model supports configurable reasoning and search grounding. These capabilities depend on the interface; Pixwit’s current workspace exposes text and reference-image inputs, with no search or reasoning controls. Official capabilities.

Evaluate performance using several comparable runs, tracking both the wait for an image and the revisions needed to finish it. A single fast generation is a weak guide to everyday latency. For factual diagrams, check the underlying information separately from the visual polish.

The model card identifies weaknesses in small or lengthy text, imperfect identity preservation, partial compliance in masked edits, left/right placement and advanced spatial or factual reasoning. It also notes occasional slow responses and timeouts. Documented limitations.

Make review specific to the task: compare labels character by character, inspect preserved parts against the source, and check spatial relationships across the whole image. Keep your original files so you can restart a revision from a clean reference.

Try the image generator and editor or choose a starting brief from the prompt library. For model selection, read the comparisons with Nano Banana 2 and Nano Banana Pro.