The Key to Solving Face Distortion Lies in "Subject Size," Not "Resolution"
Have you ever experienced a situation where the facial features don't return correctly even after repeated high-resolution processing in image generation AI? It turns out the cause is not the resolution itself, but the "size of the subject" within the frame.
- AI tends to place subjects small and far away if left unattended
- When a subject is crushed into a few hundred pixels, it becomes difficult to repair in later stages
- It is important to explicitly control the size of the subject and eliminate downsizing tags
- A post-generation processing flow is effective for accurate size adjustments tailored to specific devices

Why High-Resolution Processing Alone Doesn't Solve It
When trying to fix "face distortion" in image generation, many people take an approach of increasing the resolution or increasing the number of high-resolution steps (upscaling: the process of enlarging a low-resolution image while the AI fills in the details).
However, in experiments at Kabegami Studio, it was confirmed that the impact of the "area" the subject occupies within the frame is far greater than adjusting the sampler (the computational procedure for generating images), resolution, or the number of upscaling steps.
This is because, without specific instructions, AI tends to place subjects "small" and "far away." If generation occurs in this state, the facial parts are crushed into an extremely small area of only a few hundred pixels. Once information is crushed like this, it is difficult to restore it as accurate features no matter how many times you repeat high-resolution processing later.
- If left alone, AI places the subject "small" and "far away"
- When the subject is crushed into a few hundred pixels, repair in later stages becomes difficult
- Controlling the area is the highest priority for preventing distortion
Solutions Implemented at Kabegami Studio
To solve this problem fundamentally, Kabegami Studio introduced a mechanism to "intentionally manipulate the size of the subject." Specifically, we allow users to explicitly choose the subject size from "Up (Close-up)," "Standard," or "In a Landscape."
Additionally, we designed the system to mechanically remove "downsizing" tags (keywords that emphasize long shots or wide angles) that cause the AI to automatically move the subject further away. This prevents the AI from arbitrarily crushing the face.
We have also cleared the challenges when specifying special compositions. For example, when giving instructions like "wide" or "panorama," some AI models may draw cinematic black bars (letterboxes) at the top and bottom. To prevent this, we built a mechanism that always inserts tags to cancel out the black bars as a set.
- Made it possible to select subject size (Up, Standard, In a Landscape)
- Added a process to mechanically remove downsizing tags on the system side
- Permanently included tags to cancel out "black bars" that occur during panorama etc. instructions
Resolution and Size Design Based on Actual Measurements
Next, we looked for the optimal solution in terms of concrete numerical values. When comparing base resolutions, it was found that the outline of the head would break at 1664x936. Therefore, we have adopted "1536x864" as the base resolution that can be depicted stably.
Extreme aspect ratios (height-to-width ratios), such as the vertical 9:19.5 for smartphones, present different challenges. In vertical formats, the subject is prone to being duplicated vertically, so we have implemented special additional measures only for vertical orientations.
Finally, we were particular about the accuracy of the output. By specification, image generation software tends to round sizes to multiples of 8. Therefore, there is a possibility of microscopic discrepancies with the device's resolution as is. To prevent this, we have incorporated a specification to perform "Zoom and Center Crop" right before saving to perfectly match the device's resolution.
- Adopted "1536x864" as the base, which can maintain the outline of the head
- Individual measures for the subject duplication phenomenon in vertical (9:19.5) formats
- Perfect fit for device resolution via "Zoom and Center Crop" right before saving
Summary
The biggest insight gained from this experiment is the importance of correctly controlling "where and at what size to place the subject" before commanding the AI to "draw in detail."
Problems that cannot be solved by simply increasing the resolution are dramatically improved by this area control. At Kabegami Studio, we aim to achieve more stable wallpaper generation by accumulating these trial-and-error values.
Next time, we will explain how to create a gallery where the operation doesn't get heavy even as the number of images increases.
- Controlling "area" takes priority over resolution
- Intentional tag removal and fixed sizing are shortcuts to stability
- Next theme: How to create a gallery that doesn't get heavy
- People struggling with distorted faces generated by AI
- People who feel that repeating high-resolution processing doesn't work
- People who want to know a more stable image generation mechanism
- Sampler
- The computational algorithm used when generating images from noise.
- High-resolution processing (Upscale)
- The process of enlarging a low-resolution image while the AI fills in the details.
- Aspect Ratio
- The ratio of the height to the width of an image.
FAQ
Why 1536x864 instead of 1664x936?
Official information (experimental results) confirmed that the outline of the head breaks at 1664x936, so we adopted 1536x864 which is more stable.
Why are downsizing tags removed?
Because AI tends to place subjects small if left alone, we remove them to prevent the subject from being crushed into a few hundred pixels.
To prevent face distortion in image generation AI, it is important to secure the "subject area" before increasing the resolution. Kabegami Studio solves this problem by allowing explicit selection of subject size and eliminating tags that encourage downsizing. Furthermore, by combining the optimization of base resolution with post-generation cropping processes, we have built a mechanism to generate high-quality images optimized for devices.