Fixing Distorted Sleeve Volume in 18th-Century Puff Sleeves with Nano Banana 2

When generating historical fashion imagery, particularly 18th-century garments, the AI often struggles with complex volumetric structures. A frequent symptom reported by users is the distortion of sleeve volume. Instead of the intended fullness, the model may collapse the fabric into a flat shape or, conversely, over-expand it until the proportions look unnatural and disconnected from the body. This issue specifically affects the silhouette of puff sleeves, where the balance between the gathered cuff and the inflated upper arm is critical for historical accuracy.

The visual result is often a garment that lacks the structural integrity of the era. The fabric appears to melt away or swell like an inflated balloon rather than hanging with the weight and drape of period-appropriate textiles. This distortion breaks the immersion of the image and requires specific intervention within the generation workflow to restore the correct form.

Separating Plausible Causes from Known Facts

To effectively resolve this issue, it is essential to distinguish between what is known about the tool's capabilities and plausible reasons for the error. It is a known fact that Nano Banana 2 supports text-to-image and image-to-image workflows, allowing users to guide the generation process through detailed instructions. However, prompt instructions describe desired outcomes; they do not guarantee identity, label, object, or typography preservation. This means that simply stating "puff sleeves" does not ensure the AI will render the specific volume required without further refinement.

Plausible causes for the distortion include the model's tendency to prioritize smooth textures over complex geometry when the prompt lacks sufficient directional constraints. In some cases, the algorithm may interpret "puff" as a generic texture rather than a structural element. Another factor could be the default weighting of keywords, where the emphasis on the sleeve style inadvertently conflicts with the overall composition logic of the model. While these are logical hypotheses based on how generative models operate, there are no specific statistics or first-hand tests confirming the exact internal mechanism causing the collapse in every instance. Users must rely on iterative testing to find the right balance for their specific needs.

Diagnosing the Silhouette Proportions

Diagnosis begins with analyzing the generated output against the target historical standard. If the sleeve volume is missing, the diagnosis points to insufficient emphasis on the spatial requirements of the garment. The model likely prioritized the face or the bodice, treating the sleeves as secondary background elements. Conversely, if the sleeves are over-expanded, the prompt may have been too aggressive in requesting volume without defining the boundaries of the shoulder or the wrist.

It is important to note that Google documents Nano Banana 2 as Gemini 3.1 Flash Image (gemini-3.1-flash-image). While powerful, this model operates within specific parameters regarding how it interprets spatial relationships. If the prompt relies solely on positive descriptions like "large sleeves," the AI might lack the context to understand the limits of that largeness. The diagnosis should also consider whether the user is attempting to use multiple reference inputs. Google describes Nano Banana 2 Lite as focused on speed and cost, noting it is not optimized for multiple reference inputs or multi-turn sequential editing. Therefore, if a user is trying to fix a sleeve issue by uploading several reference images simultaneously, they may be encountering limitations inherent to the Lite version, even if the main product page suggests broader capabilities.

Fixing the Issue with Negative Prompts and Parameters

The most effective strategy to maintain correct silhouette proportions involves a combination of negative prompting and precise parameter adjustments. To prevent the collapse of the sleeve, you should explicitly instruct the model to avoid flattening or smoothing the fabric. Use negative prompts such as "flat sleeves," "collapsed fabric," "smooth arms," or "lack of volume." These instructions help the model understand what the final image should not contain, reinforcing the structural requirements of the 18th-century style.

For over-expansion, the fix lies in adding constraints to the prompt. Instead of just asking for "huge puffs," specify the relationship between the sleeve and the shoulder line. Try phrases like "structured puff sleeves," "defined shoulder seam," or "fabric weight consistent with cotton." This guides the AI to generate volume that feels grounded rather than floating. Additionally, adjusting the guidance scale can help. Higher values often enforce the prompt more strictly, which can be useful for maintaining the rigid structure of historical clothing, but they must be balanced to avoid artifacts.

Users can explore example prompts available in the Nano Banana 2 prompt library to see how others have structured similar requests. These examples serve as starting points but should be adapted to your specific vision. Remember that prompt instructions do not guarantee results, so experimentation is key. For those needing a quick iteration, the Try Nano Banana link provides direct access to the generator where these strategies can be applied immediately.

Verifying the Corrected Output

Verification is the final step to ensure the troubleshooting was successful. After applying the negative prompts and parameter changes, review the new generation for consistency. Check if the sleeve volume remains stable across different angles or variations. Does the fabric appear to have weight? Are the gathers at the cuff distinct and realistic? If the distortion persists, try refining the negative prompts further or adjusting the descriptive adjectives for the fabric type.

It is crucial to remember that while Nano Banana 2 is a robust tool, it does not promise guaranteed outcomes. Each generation is unique, and the interplay between the prompt and the model's interpretation can vary. By systematically addressing the symptoms, understanding the limitations of the model, and applying targeted corrections, users can significantly improve the historical accuracy of their generated fashion imagery. Whether working with the standard Nano Banana 2 or exploring other versions, the focus remains on clear communication with the AI to achieve the desired 18th-century aesthetic.