Nano Banana 2: Preventing Identity Drift During Costume Changes

Users frequently encounter a frustrating issue where Nano Banana 2 successfully generates a new costume or outfit but inadvertently changes the subject's facial features in the process. This phenomenon, known as identity drift, occurs when the AI prioritizes the visual details of the clothing over the specific characteristics of the person depicted. Instead of seeing your character in a new jacket or dress, you might see a stranger with similar hair color but different eyes, nose shape, or jawline. This symptom is particularly common when the prompt instructions heavily emphasize the new attire while providing insufficient constraints on the original facial structure.

It is important to distinguish between plausible causes and verified facts regarding this behavior. While users often suspect that the image resolution or the complexity of the costume design are the primary culprits, the core issue usually lies in how the model interprets the relationship between the reference input and the text prompt. The AI does not inherently "know" which parts of an image must remain static unless explicitly guided. Furthermore, it is a known fact that prompt instructions describe desired outcomes but do not guarantee identity, label, object, or typography preservation. Therefore, assuming that simply uploading a reference image will lock the face in place without additional textual reinforcement is a misconception that leads to inconsistent results.

Diagnosing the Root Causes

To effectively address identity drift, one must first diagnose whether the issue stems from the prompt strategy or the selected model capabilities. A primary cause is the lack of explicit negative constraints in the prompt. If the instruction focuses solely on adding a "cyberpunk suit" without mentioning "keep original face," the model may treat the face as just another element to be optimized for the new aesthetic.

Another critical factor involves the choice of model within the Nano Banana ecosystem. Google documents Nano Banana 2 as Gemini 3.1 Flash Image, Nano Banana Pro as Gemini 3 Pro Image, and Nano Banana 2 Lite as Gemini 3.1 Flash Lite Image. These are distinct models with different strengths. Specifically, Google describes Nano Banana 2 Lite as focused on speed and cost. It is not optimized for multiple reference inputs or multi-turn sequential editing. Consequently, using Nano Banana 2 Lite for complex tasks requiring strict identity retention across multiple costume iterations is likely to result in higher rates of drift compared to the standard Nano Banana 2 or Pro versions. Users attempting to maintain identity across a sequence of images should be aware that Lite is not designed for these workflows without understanding its specific limitations.

Additionally, the nature of the costume change itself plays a role. Highly detailed or transformative costumes can sometimes overwhelm the model's attention mechanism, causing it to shift focus away from the facial region. This is not a bug but a feature of how generative models balance competing visual elements. The model attempts to satisfy the most prominent part of the prompt, which often becomes the new costume if described vividly.

Corrective Prompt Strategies for Consistency

Fixing identity drift requires a deliberate approach to prompt engineering. Since prompt instructions do not guarantee preservation, users must employ specific phrasing to anchor the facial identity. Start by clearly defining the subject's immutable features before introducing the new costume. For example, instead of saying "Generate a warrior in armor," try "A portrait of [Subject Name] with [specific eye color], [nose shape], and [skin tone] wearing a futuristic warrior suit."

Incorporating reference-based language is also essential. When using the image-to-image workflow, ensure the reference image is high quality and clearly shows the face. Pair this with a prompt that explicitly states the goal of retaining the original likeness. You might add phrases like "maintain exact facial features" or "preserve original face geometry" to the instruction set. While these terms do not offer a mathematical guarantee, they provide the strongest signal available to the model to prioritize the source identity.

If you find that the standard prompts are still resulting in drift, consider adjusting the weight of the reference image relative to the text prompt. In many cases, reducing the emphasis on the costume description slightly and increasing the focus on the subject's attributes can yield better results. Remember that Nano Banana refers to the AI image generation/editing tool; it is not a skincare brand or physical product, so no external cosmetic factors influence the output. The control lies entirely within the digital interaction.

For users who require high-fidelity consistency across multiple variations, upgrading to a model better suited for complex editing may be necessary. As noted, Nano Banana 2 Lite is not optimized for multi-turn sequential editing. If your workflow involves changing costumes repeatedly while keeping the face identical, the standard Nano Banana 2 or Nano Banana Pro models are more appropriate choices. These models generally handle the nuance of identity retention better than the Lite version.

Verifying Results and Next Steps

After applying these strategies, verification is the final step. Generate a batch of images with slight variations in the costume description while keeping the facial descriptors constant. Compare the outputs against the original reference. Look specifically for subtle shifts in eye shape, ear placement, or skin texture. If the face remains consistent across the batch, the prompt strategy is effective. If drift persists, refine the prompt further by adding more specific anatomical descriptors.

It is crucial to manage expectations regarding guaranteed outcomes. No generative AI tool can promise 100% identity preservation in every single generation, especially when significant visual changes like costume swaps are requested. However, by understanding the limitations of the specific model being used and crafting precise prompts, you can significantly reduce the frequency of identity drift.

For those ready to experiment with these techniques, you can access the generator directly to test your new prompt structures. Try Nano Banana. By combining careful prompt construction with the right model selection, you can achieve professional-quality results where your characters look great in any outfit without losing their unique identity.

Sources: Google Gemini image generation documentation (https://ai.google.dev/gemini-api/docs/image-generation)