Fixing Lost Facial Features in Nano Banana 2 Extreme Portrait Crops

When using Nano Banana 2 to generate or edit images, users may encounter a specific issue where facial features such as eyes, mouths, or distinct identity markers disappear during aggressive cropping. This problem often manifests when forcing a tight vertical crop on a generated human subject. Instead of maintaining the integrity of the face within the new frame, the AI might blur, distort, or completely remove these critical details. The result is an image that lacks the intended expression or recognizable identity, rendering the portrait ineffective for its purpose.

This symptom is particularly noticeable when the aspect ratio changes drastically or when the user attempts to zoom in beyond the model's default comfort zone. It is important to distinguish this behavior from general image degradation; the issue is specifically tied to the spatial constraints applied to the human subject rather than a failure of the underlying generation engine itself.

Separating Plausible Causes from Known Facts

To resolve this issue effectively, it is necessary to separate what is known about the tool from plausible but unverified theories regarding its internal processing.

Known Facts: Nano Banana 2 refers to the AI image generation and editing tool available at /nanobanana2. It supports both text-to-image and image-to-image workflows. The system utilizes Google models, specifically identifying Nano Banana 2 as Gemini 3.1 Flash Image (gemini-3.1-flash-image). According to documentation, prompt instructions describe desired outcomes but do not guarantee the preservation of identity, labels, objects, or typography. This means that even with clear instructions, the model may prioritize composition over strict feature retention if the constraints are too severe.

Plausible Causes vs. Reality: A common assumption is that the AI simply "forgets" the face due to low resolution. However, the core issue often lies in how the prompt interprets spatial limits. Users frequently assume that adding words like "extreme close-up" will force the AI to keep the face visible. In reality, without proper weighting, the model may interpret "extreme crop" as a directive to fill the frame with texture or background elements, inadvertently pushing the facial features out of the focal area or smoothing them away to satisfy the geometric constraint.

Another misconception involves the model version. Some users might try to solve this by switching to Nano Banana 2 Lite. However, Google describes Nano Banana 2 Lite as focused on speed and cost. It is explicitly not optimized for multiple reference inputs or multi-turn sequential editing. Relying on the Lite version for complex tasks requiring precise feature preservation in edited crops is likely to yield inconsistent results. The standard Nano Banana 2 model is better suited for maintaining detail in these scenarios.

Diagnosing the Issue Through Prompt Weighting

The diagnosis for lost facial features usually points to insufficient emphasis on specific anatomical landmarks within the prompt structure. When the AI receives a command to create a tight vertical crop, it balances the request for "closeness" against the need for "completeness." If the prompt does not explicitly weight the importance of the eyes and mouth, the algorithm may sacrifice these details to achieve the requested framing density.

To diagnose this, review your current prompt. Does it focus heavily on the crop dimensions while neglecting the subject's specific features? For example, a prompt like "vertical crop of a woman" is often too vague. The AI fills the void with generic textures. A more diagnostic approach involves checking if the prompt includes weighted terms for the missing features. Without explicit instruction, the model treats the face as a single block rather than a collection of distinct parts.

It is also crucial to verify that you are not using the Lite version for this specific task. Since Nano Banana 2 Lite is designed for speed, it may lack the nuanced understanding required to hold onto fine details during aggressive edits. Ensure you are operating within the standard Nano Banana 2 environment to access the full capabilities of the Gemini 3.1 Flash Image model.

Fixing the Problem with Strategic Adjustments

To fix the loss of facial features, you must adjust your prompt weighting to prioritize identity preservation over pure framing. Start by explicitly listing the features you want to retain. Use descriptive language that assigns higher importance to the eyes and mouth. For instance, instead of just saying "portrait," try "extreme close-up portrait with sharp focus on eyes and detailed lips."

If you are working within the Nano Banana 2 interface, consider breaking down the request. First, generate the base image with the subject clearly defined. Then, use the image-to-image workflow to apply the crop, ensuring the prompt reiterates the need for feature clarity. Do not rely solely on the initial generation to handle the crop logic.

Remember that prompt instructions do not guarantee identity preservation. Therefore, iterative refinement is key. If the first attempt loses the eyes, refine the prompt to emphasize "clearly visible eyes" and "distinct mouth shape" before re-running the generation. You can also explore the prompt library provided on the website for examples that demonstrate successful feature retention, though these should be treated as starting points rather than guaranteed solutions.

For users seeking advanced capabilities, the Nano Banana Pro page at /nanobananapro offers access to the Gemini 3 Pro Image model, which may provide additional stability for complex edits. However, always verify the specific features available on the platform, as model names do not automatically equate to identical feature sets across all versions.

Verifying the Solution

After applying these adjustments, verify the output by checking the generated image against your original intent. Does the face remain intact? Are the eyes and mouth distinct? If the features are still missing, try reducing the aggressiveness of the crop description slightly or increasing the weight of the facial descriptors further.

Keep in mind that no method guarantees a perfect outcome every time. The nature of generative AI involves probabilistic outputs. However, by understanding the limitations of the prompt instructions and selecting the appropriate model version, you can significantly improve the likelihood of preserving facial details. For those ready to experiment with these techniques, Try Nano Banana to apply these troubleshooting steps directly in the generator.

By focusing on precise prompt engineering and understanding the specific strengths of the Nano Banana 2 model, users can overcome the challenge of losing facial features in extreme portrait crops.