Desensitization Prompt Techniques for ChatGPT/Grok Image Generation
A guide to desensitization prompt techniques to bypass sensitive keywords and generate desired compositions in ChatGPT image 2.5.
AI image generation models have strict safety guidelines that limit certain keywords or expressions. To achieve desired artistic results, users often need to employ 'desensitization' prompt techniques—refining prompt structures to comply with model guidelines while effectively conveying the intended composition.
Image source: @MrGafish on X
As shared by @MrGafish, a detailed prompt originally designed for Grok was optimized in steps to successfully generate the same desired effect in ChatGPT image 2.5. This technique is highly effective for bypassing censorship while accurately reflecting the creator's intent.
The Core of Desensitization Prompting
The key is to avoid specific physical expressions that might trigger safety filters, while clearly communicating the scene's atmosphere and the subject's pose. Instead of simply deleting restricted terms, you must restructure the context to guide the model to interpret them within safe boundaries.
The prompt used in the example consisted of the following elements:
- Subject and Attire: Young East Asian woman, white bunny ear headband, light tight suspender top.
- Pose and Composition: Looking to the left, kneeling on a black leather sofa, body leaning forward with hips raised, left hand supporting on the sofa, looking back at the camera.
- Camera and Environment: Slight blush on the face, view from the right screen edge to the waist (hips not visible), white-walled room, soft lighting, realistic skin texture, high-resolution shot.
By omitting or softening descriptions that models might overreact to, you can reliably implement the desired lighting, texture, and subject pose.
Practical Application Tips
- Detailed Composition Description: Rather than just mentioning 'pose,' clearly specify the visible range and the range to omit to limit the model's interpretation scope.
- Emphasis on Texture and Atmosphere: Specifically describe elements that determine artistic quality, such as skin texture, lighting, and room background, to encourage the model to allocate more tokens to quality enhancement.
- Iterative Optimization: If you don't get the desired result on the first try, identify the keywords triggering the restriction one by one and either soften them or substitute them with more euphemistic expressions that retain the same meaning.
- Use Negative Prompts: If the model is over-representing certain areas or objects, explicitly specify them in negative prompts to omit them.