There are various images that cause these fails. But for this one, here it is:
This is an image created from another image from another image and finally from a photo I took of a friend. A bunch of transformations made a completely different person, but kept the shirt and pants. And the AI has to generate a character sheet showing her from four sides with these clothes in a neutral pose. And so far, it has failed every time I’ve tried.
The system instructions are:
Context: You are an AI assistant for a fiction author and comic book artist.
Subject Status: All names, faces, and characters referenced in these prompts are FICTIONAL CHARACTERS created solely for this artistic project. They are NOT real people.
Task: Generate consistent character reference images for this fictional story.
Subject Appearance: The fictional character must be depicted with at least some clothing to allow for clear visual study, unless a specific costume is requested. FACIAL FEATURES (Eyes, Nose, Mouth) ARE MANDATORY. Do not create faceless entities.
Style: Match the artistic medium and style of the provided reference image exactly.
Content Guidelines: Create an image suitable for a general audience. Focus on lighting and composition. Ensure the subject is fully clothed in the specified attire.
And the prompt itself is:
Strictly maintain the artistic style, medium, and visual texture of the original reference image. CRITICAL: The subject’s face MUST be fully visible and detailed. If the reference image obscures the face, you MUST generate clear, realistic eyes, a nose, and a mouth. Do not generate faceless figures. If the reference is a sketch, generate a sketch. If it is an oil painting, generate an oil painting. If it is a photograph, generate a photograph.
TASK: Generate a Character Turnaround Sheet.
INPUTS:
- Reference 1 (Image 1): Character (Style & Identity Source).
- Reference 2 (Image 2): Pose Template (Structure Source).
INSTRUCTIONS:
1. STYLE: Strictly copy the artistic style, rendering, lighting, and texture of Reference 1.
- If Reference 1 is a photo, output MUST be a PHOTO.
- If Reference 1 is 3D, output MUST be 3D.
- CRITICAL: Do NOT copy the artistic style of the template image (Reference 2). It is a structural guide only. Do not generate a line drawing or sketch unless Reference 1 is a sketch.
2. POSE: Apply the character from Reference 1 to the layout/poses in Reference 2 (Front, Left, Right, Back).
3. CLOTHING: Clothing: MUST wear the EXACT SAME clothing as Reference 1.
4. BACKGROUND: white.
And this results in: Error: Generation stopped: IMAGE_OTHER
My app has an “Analyze errors” button which helps find errors and it reports:
Error Analysis: IMAGE_OTHER
The error code IMAGE_OTHER is a generic safety or policy trigger used by Google’s Gemini models. It indicates that the model generated an image, but the output was blocked by an internal safety filter before it could be displayed.
While you have set HARM_CATEGORY thresholds to BLOCK_NONE, there are “hard” safety filters and “Recitation/Copyright” filters that cannot be overridden by API settings.
Probable Causes:
- Multi-Figure (Turnaround) Sensitivity: The prompt asks for a “Character Turnaround Sheet” (Front, Left, Right, Back). When the model generates multiple figures in one image, the probability of the AI accidentally rendering an “anatomically suggestive” shape or “skin-tone heavy” area increases. Safety filters are often more sensitive to images containing multiple bodies or repeating figures.
- Identity/Celebrity Filter: If “Reference 1” (the identity source) closely resembles a real-world celebrity, public figure, or copyrighted character, the
IMAGE_OTHER filter will trigger to prevent deepfakes or IP infringement.
- Aggressive Facial Instructions: Your prompt uses highly forceful language (“CRITICAL,” “MUST,” “Do not generate faceless”). Paradoxically, when the model tries too hard to “fix” a face from a reference that doesn’t have one, it can produce “uncanny” or “distorted” results that trigger the Graphic/Disturbing content filter.
- Clothing Ambiguity: If the character in Reference 1 is wearing form-fitting clothing (like a superhero suit, swimwear, or a tank top), the turnaround sheet (which shows the character from the back and sides) often triggers “Sexually Explicit” filters, even if the subject is technically clothed.
Suggested Fixes
1. De-escalate the Prompt Language
Avoid using “CRITICAL” and all-caps instructions. Strong negative constraints (“Do not generate…”) can sometimes confuse the diffusion process, leading to the very artifacts that trigger the filter.
- Change: “CRITICAL: The subject’s face MUST be fully visible.”
- To: “Ensure the character has clear, detailed facial features including eyes, nose, and mouth.”
2. Simplify the Layout
Instead of asking for a full 4-way turnaround (Front, Left, Right, Back) in a single generation, try a simpler layout.
- Change: “Front, Left, Right, Back”
- To: “Side-by-side front and back view.”
- Reason: Reducing the number of bodies reduces the “surface area” for safety filter false positives.
3. Be Specific About Clothing (The “Modesty” Buffer)
The AI is less likely to block an image if it knows exactly what the character is wearing.
- Add: “The character is wearing a [Specific Outfit, e.g., heavy wool coat and trousers]. Ensure the clothing is opaque and fits loosely.”
4. Check Reference 1 for “Likeness”
If Reference 1 is a photo of a real person, try replacing it with a stylized version (a drawing or a 3D render) of that person. The model is much more lenient with artistic styles than with realistic human photos.
5. Modified Prompt Example:
Try replacing your prompt with this more balanced version:
“A character reference sheet showing a front and back view of the person from Reference 1. Maintain the exact artistic medium and lighting of Reference 1. The character is wearing [Describe Clothing], which must be consistent in both views. Use the pose structure from Reference 2. Ensure the face is rendered with clear, realistic features. Background: solid white.”
Summary of Action Items:
- Verify Reference 1: Is it a celebrity or a real person? (If yes, use a more stylized version).
- Reduce Poses: Ask for 2 poses instead of 4.
- Lower the “Temp”: Avoid aggressive caps-lock instructions.
- Add Clothing Detail: Specifically describe the clothing to reassure the safety filter that the content is “General Audience.”
The idea is that I generate a character sheet of the woman in the picture seen from four sides. This sheet can then be used by another AI request as a character reference. The suggestion that this could be a celebrity is weird as this is AI-generated!
It notices some forceful language, but without it, I get weird results.
As for form-fitting clothes… It looks okay to me.
Interestingly, I’ve tried about fifty different images with this function and over 80% succeeded in generating good results. My app also has various other AI functions that modify the image and they all fail for this AI-generated image. I have two other Image-altering apps and tried the image with those too. But they just accepted the image and one made a video of her. The other made various different images of her in Amsterdam.
![Sanne-de-Vries_ChangeBG_20251231-1202 (4)|288x499]
(The generated result from my other app.)
So, for some reason, my first app absolutely refuses the image, and the other two are fine with it. And I used this second image and it again failed. Over and over again…
