De-AI: Image Realism Refinement

The Problem: Fixing Generation Artifacts in AI Images

AI-generated images often exhibit visible "generation artifacts" that undermine their credibility and professional quality. These artifacts typically include:

• Material and Skin: Plastic-like skin, unnaturally smooth hair, and repetitive, mechanical textures.
• Anatomy: Deformed fingers, blurred or distorted facial features, and incorrect limb proportions.
• Lighting and Consistency: Contradictory light source directions, harsh shadows, and locally incoherent color grading.

This skill focuses on fixing the most impactful artifacts from a user's chosen image, rather than performing an aimless full redraw.

Core Capabilities and Workflow

The skill achieves refinement through minimal intervention. Its core workflow involves:

1. Diagnosis and Routing: The host Agent first examines the original image across dimensions like anatomy, skin, and lighting to pinpoint the most prominent set of artifacts. Based on the issue type, it selects the repair route:

   • For localized defects (e.g., one malformed hand), it uses beatra.images.edit with specified local region coordinates.
   • For image-wide issues (e.g., overall plastic skin), it also uses beatra.images.edit, but without specifying edit_regions.
   • It only resorts to beatra.images.transform when the user explicitly requests a new scene or composition.

2. Minimal Repair Principle: The skill adopts the "minimum viable repair path," addressing only necessary areas without altering user-approved text, logos, character features, or product shapes. After repair, it delivers the final assets, actual cost (billing.net_charged_credits), and task ID, alongside a visual comparison with the original.

3. Critical Control Points: The workflow pauses and requires user confirmation for any action involving paid API calls, altering the canvas ratio, or increasing output count. Once a task is submitted, it is polled only once and never resubmitted to ensure idempotency and traceability.

Application Boundaries and Notes

This skill is designed for existing portraits, product images, marketing graphics, or social media covers. Please note the following boundaries:

• Input Requirement: The host Agent must be able to view the original image directly; uploading a file alone does not constitute diagnosis.
• Repair Scope: By default, it fixes only one set of the most impactful artifacts per invocation, not all issues simultaneously.
• Non-Substitutability: It cannot be used to generate brand-new images or perform major compositional changes (unless explicitly requested by the user), nor does it provide content analysis (like Beatra image analysis).
• Task Resumption: If a task times out without completion, the task ID should be returned for later resumption. Creating a second, parallel task is strictly prohibited.