Mastering Eye Gaze Adjustment in Portraits with Nano Banana 2

Understanding the Goal of Subtle Gaze Shifting

Adjusting where a subject looks in a portrait is one of the most delicate tasks in AI image editing. The objective is not to change the person's identity or alter their facial structure, but simply to redirect their line of sight. When working with Nano Banana 2, which supports robust text-to-image and image-to-image workflows, achieving this requires a specific approach. A common pitfall is over-correcting the prompt, which leads to distorted features or an unsettling appearance often described as the uncanny valley.

The key to success lies in precision. You are asking the model to perform a micro-adjustment rather than a complete reconstruction. Nano Banana 2 allows users to leverage its prompt library for inspiration, but generic prompts rarely yield the subtle control needed for eye gaze. Instead, you must construct a prompt that explicitly defines the desired direction while strictly instructing the model to preserve existing geometry. Remember that prompt instructions describe desired outcomes; they do not guarantee identity, label, object, or typography preservation. Therefore, your language must be careful to emphasize stability alongside the requested change.

Constructing the Precise Prompt Structure

To achieve a natural-looking shift in eye direction, your prompt needs a structured format that separates the modification from the preservation constraints. Start by describing the current state of the image briefly to ground the model, then introduce the specific adjustment. For example, instead of saying "make them look left," a more effective instruction might be "shift the pupils slightly to the left while keeping the head angle and facial features identical."

Here is a usable prompt structure designed for this specific task. Note that these examples are untested scenarios intended to illustrate the syntax required for the tool:

Example Prompt: [Image Description], adjust the subject's gaze to look [direction] (e.g., slightly up and to the right), maintain original facial structure, keep lighting and skin texture unchanged, no distortion.

When using this structure, it is crucial to include negative constraints implicitly or explicitly. By stating "maintain original facial structure" and "no distortion," you signal to the model that the primary goal is a minor rotation of the eyes, not a re-rendering of the face. If the result feels too aggressive, try softening the directional cue. Use words like "subtly," "slightly," or "minimally" to reduce the intensity of the transformation. This helps prevent the model from hallucinating new facial features or altering the expression in unintended ways.

For those interested in exploring different models, Google documents Nano Banana 2 as Gemini 3.1 Flash Image. While Nano Banana Pro uses Gemini 3 Pro Image and Nano Banana 2 Lite uses Gemini 3.1 Flash Lite Image, the latter is focused on speed and cost and is not optimized for multiple reference inputs or multi-turn sequential editing. Consequently, for complex edits like gaze adjustment that require high fidelity, relying on the standard Nano Banana 2 workflow is generally safer than attempting it on the Lite version without understanding its limitations.

Judging Results and Fixing Common Issues

After generating the image, the first step in judging results is to compare the original and the output side-by-side. Look specifically at the iris position relative to the eyelids and the surrounding skin texture. If the eyes appear crossed, misaligned, or if the skin around the eyes has become blurry or warped, the prompt was likely too forceful. The uncanny valley effect often manifests when the eyes look correct but the rest of the face feels slightly "off" due to over-processing.

If the gaze did not move enough, you can refine the prompt by increasing the specificity of the direction. However, avoid simply repeating the same paragraph or doubling down on the command, as this often degrades quality. Instead, try adding context about the environment, such as "looking toward a light source on the left," which gives the model a logical reason for the shift. If the facial structure changed, you must reinforce the preservation constraint in the next attempt. Phrases like "strictly preserve facial geometry" or "keep the nose and mouth exactly as they are" can help anchor the generation.

It is important to remember that prompt instructions do not guarantee specific outcomes. Even with perfect phrasing, the AI may produce variations. If the results remain unsatisfactory after several iterations, consider that the initial input image might have had low resolution or ambiguous lighting, making the edit difficult. In such cases, starting with a clearer base image is often more effective than tweaking the prompt further.

For those ready to experiment with these techniques, you can access the necessary tools directly through the platform. Try Nano Banana to begin your own portrait adjustments. Always approach these edits with patience, testing small changes to find the sweet spot between correction and naturalism.