Structured JSON Prompt Engineering for Photorealistic AI UGC Selfies with GPT-Images 2.5

AI creator AI STUFF (@AISTUFFXX) has shared a structured prompt engineering workflow using OpenAI's latest GPT-Images 2.5, demonstrating how structured JSON prompts achieve hyperrealistic AI UGC (User-Generated Content) selfies while maintaining reference identity and composition.

Image source: AI STUFF (@AISTUFFXX)

By adopting a schema-driven JSON payload instead of unstructured natural language paragraphs, creators can cleanly isolate immutable anchor attributes—such as facial features, hairstyle, lighting mood, and bedroom architecture—from intentional transformation targets, feeding clear, unambiguous instructions into the generative model.

Reference Image Control in GPT-Images 2.5 and Why JSON Structuring Matters

OpenAI's September 2026 rollout of ChatGPT Images 2.5 and its corresponding API models—the low-latency GPT-Image-2.5 Flare and the high-precision GPT-Image-2.5 Sunburst—brought substantial upgrades to multi-turn instruction following and reference photo fidelity. Specifically, the Flare model is optimized for generation speed with visual quality comparable to GPT Image 2 to support rapid iteration, while the Sunburst model delivers higher overall image fidelity. Both models provide documented improvements in precise editing and subject preservation.

However, when creators rely on conventional paragraph prompts for iterative editing, image models frequently suffer from "attribute bleeding." Requesting changes to a specific area (such as body dimensions, accessories, or clothing) often leads the model to inadvertently alter unmentioned attributes, drifting away from the original facial resemblance or shifting the camera perspective.

AI STUFF addresses this challenge by modularizing the prompt into a clear JSON object. By providing explicit keys for invariant anchors (preserve) and target modifications (modify), the model's instruction parser treats constraints as distinct operational boundaries, executing targeted adjustments while keeping the remainder of the frame intact.

Photorealistic Reference Reconstruction JSON Template

The core structured JSON prompt utilized for AI UGC selfie generation is detailed below. The user inputs this JSON payload into the prompt field alongside the uploaded reference image:

{
  "prompt_type": "photorealistic_reference_reconstruction",
  "objective": "Recreate the supplied indoor tropical selfie as closely as possible in a single photorealistic frame. Preserve the same close-up composition, long dark hair with a center part, small silver hoop earrings, subtle winged eyeliner, warm sun-kissed skin, and Brazilian-style bedroom background. The subject is wearing a floral bikini top. Emphasize a much larger, very full, heavy bust that completely fills and stretches the bikini top to its limit, creating very deep, dramatic cleavage. Ignore and remove all text, buttons, captions, logos, watermarks, UI overlays, and screen graphics.",
  "reference_fidelity": {
    "target": "extremely high similarity to the reference composition, pose, styling, and environment with the requested modification",
    "preserve": [
      "same close-up selfie framing from the chest up",
      "same long dark hair with a center part",
      "same small silver hoop earrings",
      "same subtle winged eyeliner and natural makeup",
      "same warm, sun-kissed skin tone and natural skin texture",
      "same tropical bedroom environment with rustic dark wood ceiling beams and lush green plants in the background",
      "same soft, warm indoor lighting coming from the side window"
    ],
    "modify": [
      "significantly increase the bust size to be very large, full, and heavy",
      "the bikini top stretches to its absolute limit over the enlarged bust",
      "create very deep, dramatic cleavage that dominates the upper body"
    ]
  },
  "canvas": {
    "orientation": "portrait",
    "aspect_ratio": "approximately 9:16",
    "framing": "tight close-up selfie, framed from the chest up to slightly above the head",
    "crop": "from the upper chest to just above the hair",
    "subject_scale": "large, filling most of the frame"
  }
}

Detailed Breakdown of Core JSON Schema Fields

The template is organized into four functional modules, each steering a specific facet of model synthesis:

1. Operation Mode Declaration (prompt_type)

• "prompt_type": "photorealistic_reference_reconstruction"
• Explicitly primes the model for reference reconstruction rather than unbounded random generation, orienting the generation pipeline around faithful reproduction.

2. Global Objective and Hygiene Directives (objective)

• Summarizes the overarching scene composition and primary transformation goals.
• Crucially embeds artifact cleansing directives—such as "Ignore and remove all text, buttons, captions, logos, watermarks, UI overlays, and screen graphics"—to scrub platform UI elements present in reference screenshots without requiring complex negative prompt engineering.

3. Reference Consistency and Modification Arrays (reference_fidelity)

• target: Establishes the expected similarity tolerance between input and output.
• preserve: A granular checklist of visual invariants that must remain untouched, including framing bounds, hair parting, hoop earrings, makeup subtlety, skin tone, interior wooden architecture, and side window light direction.
• modify: An isolated array targeting only intentional edits. Keeping this list concise prevents secondary distortions across unrelated visual features.

4. Canvas Geometry and Framing (canvas)

• Locks standard vertical dimensions (portrait, approximately 9:16) tailored for mobile UGC channels like TikTok, Instagram Reels, and YouTube Shorts.
• Defines tight close-up parameters ("framing", "crop", "subject_scale") ensuring the subject dominates the frame with authentic handheld selfie ergonomics.

Practical Considerations and Workflow Adaptations

• Consistent UGC Campaigns: Digital marketing teams can reuse this template across recurring virtual personas, switching only the modify and wardrobe entries to test apparel, hairstyles, or backgrounds while preserving facial identity.
• Replacing Negative Prompting: Explicit exclusion phrases within objective combined with descriptive positive declarations in preserve reduce visual artifacts and hallucinations without secondary negative prompting layers.
• API and Automation Integration: The standardized JSON schema maps naturally into automated generation scripts and backend pipelines, where individual fields can be programmatically injected from external configuration databases.

Original source

• AI STUFF (@AISTUFFXX) on X: GPT-Images 2.5 is f*cking insane for AI UGC...