Fixing Vertical Shaft Distortion in Nano Banana 2: Restoring Perspective

When generating architectural structures like deep vertical shafts, mineshafts, or towering spires with Nano Banana 2, users may encounter a specific visual artifact where the vertical lines fail to remain parallel. Instead of converging naturally toward a vanishing point or maintaining strict orthogonality, the walls might appear to bow inward or outward, breaking the intended isometric illusion. This symptom often manifests as a warped tunnel effect that makes the structure look unstable or physically impossible rather than a precise engineering drawing. The primary goal when addressing this issue is to restore accurate spatial relationships so that the depth feels real without compromising the geometric integrity of the vertical elements.

Distinguishing Symptoms from Known Model Behaviors

It is crucial to separate the observed visual symptoms from the known capabilities and limitations of the underlying models. The symptom here is strictly the failure of vertical lines to maintain their intended trajectory, resulting in a distorted view of the shaft. This is distinct from general blurriness, loss of detail, or incorrect object placement. While some AI image generation models struggle with complex geometry, the specific behavior of Nano Banana 2 varies based on the model variant selected. Google documents Nano Banana 2 as Gemini 3.1 Flash Image (gemini-3.1-flash-image), while Nano Banana Pro corresponds to Gemini 3 Pro Image (gemini-3-pro-image). These are distinct Google image models with different optimization targets.

Known facts indicate that prompt instructions describe desired outcomes but do not guarantee identity, label, object, or typography preservation. Therefore, if the vertical alignment fails, it is not necessarily a bug in the rendering engine but rather a limitation in how the model interprets spatial constraints without sufficient guidance. Additionally, users must be aware that Nano Banana 2 Lite is focused on speed and cost. It is not optimized for multiple reference inputs or multi-turn sequential editing. Attempting to fix complex perspective issues in Nano Banana 2 Lite without understanding these limitations may lead to inconsistent results, as the model prioritizes rapid generation over fine-grained control over structural geometry.

Diagnosing the Root Cause of Perspective Breakdown

The root cause of distorted views in vertical shafts usually stems from ambiguous prompting regarding the camera angle and the convergence of lines. When a user requests a "deep shaft" without specifying the projection type, the model may default to a standard perspective view where vertical lines converge at a distant vanishing point, which can look unnatural if an isometric or orthographic look was intended. Alternatively, the model might interpret the request as a fisheye or wide-angle shot, causing the straight walls to curve.

Another diagnostic factor is the lack of explicit negative constraints. If the prompt does not explicitly state what not to do, such as avoiding curved walls or ensuring parallel vertical edges, the model fills in the gaps with its training data's most common interpretations of tunnels, which often involve natural curvature. Furthermore, relying on vague terms like "tall tower" without defining the viewing angle (e.g., "looking straight up," "bird's eye view," or "orthographic projection") leaves too much room for the model to hallucinate a perspective that breaks the isometric illusion. The issue is rarely the tool itself failing to render, but rather the prompt failing to constrain the geometric rules of the scene.

Implementing Fixes Through Explicit Constraints

To resolve these distortions, the most effective strategy is to reprompt with explicit perspective constraints. Users should rewrite their prompts to include specific directives about line behavior. For instance, adding phrases like "strictly parallel vertical lines," "orthographic projection," or "isometric view" forces the model to adhere to a specific geometric framework. If the goal is a realistic deep shaft, specifying "vanishing point at the center top" helps align the convergence correctly.

Users can leverage the prompt library available on the website to find example prompts that successfully handle similar architectural challenges. These examples serve as templates for structuring your own requests. Remember that prompt instructions describe desired outcomes; they do not guarantee identity, label, object or typography preservation, so you may need to iterate on the wording until the geometry holds. If the initial attempt still shows warping, try refining the description of the camera position. For example, changing "a view down a shaft" to "a direct overhead view looking straight down a perfectly vertical shaft" can significantly alter the output.

For users requiring higher fidelity in complex spatial edits, Nano Banana Pro (Gemini 3 Pro Image) may offer better adherence to detailed constraints compared to the standard Nano Banana 2 or the Lite version. However, always verify the specific capabilities required for your workflow, as Nano Banana 2 Lite is not optimized for multi-turn sequential editing. You can explore the features further by visiting Try Nano Banana.

Verifying the Correction and Final Output

Once the revised prompt has been generated, verification involves a close inspection of the vertical lines relative to the frame edges. In a corrected image, the walls of the shaft should either remain perfectly parallel (in an orthographic view) or converge smoothly and symmetrically toward a single, logical vanishing point without any bowing or twisting. Check the corners where the floor meets the walls; they should form consistent angles throughout the depth of the image. If the lines still appear wavy or misaligned, the prompt likely lacked sufficient specificity regarding the projection type, and another iteration with stronger geometric keywords is necessary.

By systematically adjusting the prompt to enforce strict perspective rules, users can effectively eliminate the distortion that breaks the isometric illusion. This process transforms a warped, unrealistic image into a coherent representation of a deep vertical space. Always remember that while the tool provides powerful generation capabilities, the precision of the result relies heavily on the clarity and specificity of the text instructions provided by the user.