Memorable Image Generator

Science-backed image generation agent that scores and optimizes images for memorability using ResMem (Brain Bridge Lab, University of Chicago) before returni...

Security Scan

The skill's code, runtime instructions, and required GEMINI_API_KEY are consistent with its stated purpose (generate images via Gemini and score them with ResMem); nothing in the files indicates covert data exfiltration or unrelated credential requests.

Purpose & Capability

Name/description, SKILL.md, and the included script all describe using Google Gemini for generation and ResMem for scoring. The single required env var (GEMINI_API_KEY) and the listed Python packages (resmem, torch, torchvision, pillow, requests) are appropriate for that purpose.

Instruction Scope

Runtime instructions and the script only perform image generation (HTTP POST to Google's generativelanguage endpoint), local memorability scoring (ResMem), temporary file handling, and writing the final image to the requested output path. The script reads the API key from CLI/env/config and does not access other system files or unrelated environment variables.

Install Mechanism

There is no automated install spec (instruction-only). The SKILL.md correctly instructs users to pip-install the listed Python packages. This is low-risk compared to arbitrary binary downloads; the dependency list (including torch) does imply heavy local installation but is expected for ResMem.

Credentials

Only GEMINI_API_KEY is requested. That key is necessary to call the Gemini API. No other credentials or secrets are required. (Minor metadata inconsistency: registry metadata showed Primary credential: none even though GEMINI_API_KEY is required, but this is an administrative bookkeeping mismatch rather than a functional incoherence.)

Persistence & Privilege

The skill does not request persistent or elevated platform privileges (always:false). It does not modify other skills or system-wide configs; it runs as a normal script when invoked.

Assessment

This skill appears to do what it says: it sends prompts to Google Gemini and scores returned images locally with ResMem. Before installing, be aware that: (1) you must provide a GEMINI_API_KEY (the key will be sent to Google's API and is included in the request URL), so treat it like any API secret; (2) installing/resmem and torch can be large and may require a GPU or significant disk space; (3) ResMem is under a non-commercial license—confirm licensing if you plan commercial use; (4) generated prompts and images are transmitted to Google's service, so avoid sending sensitive or private image content. If these tradeoffs are acceptable, the skill's requirements and behavior are proportionate to its stated purpose.

Science-backed image generation that optimizes for memorability — not just aesthetics.

Most image generators stop when the image looks good. This one keeps going until the image is remembered. Powered by Google Gemini for generation and ResMem (Brain Bridge Lab, University of Chicago) for memorability scoring, it iterates until your image clears a science-validated memorability threshold.

---

Prerequisites

API Key:

• GEMINI_API_KEY environment variable, OR
• --api-key CLI flag, OR
• ~/.config/gemini/api_key file

Python 3.8+ with these packages:

pip install resmem torch torchvision pillow requests

---

Quick Start

python scripts/generate_memorable_image.py \
  --prompt "a lone astronaut standing on a red desert planet at dusk" \
  --output hero.png \
  --threshold 0.7 \
  --verbose

---

How It Works

1. Generate — Calls the Gemini REST API (gemini-2.0-flash-exp) with your prompt
2. Score — Runs the image through ResMem to get a memorability score (0–1)
3. Threshold check — If score ≥ threshold (default: 0.7), saves and returns the image
4. Regenerate — If below threshold and attempts remain, enhances the prompt with composition cues and tries again
5. Max attempts — After 3 attempts (default), saves the best result regardless

Each failed attempt appends increasingly strong composition cues to the prompt:

• Attempt 2: ", striking composition"
• Attempt 3: ", vivid colors, memorable focal point"

---

Script Usage

# Basic usage
python scripts/generate_memorable_image.py --prompt "your image description"

# Full options
python scripts/generate_memorable_image.py \
  --prompt "your image description" \
  --output path/to/output.png \
  --threshold 0.75 \
  --max-attempts 3 \
  --api-key YOUR_KEY \
  --verbose

# With verbose scoring output
python scripts/generate_memorable_image.py \
  --prompt "a surreal clock melting over a desert canyon" \
  --verbose

CLI Arguments:

Argument	Default	Description
--prompt	(required)	Image description
--output	memorable-image.png	Output file path
--threshold	0.7	Memorability threshold (0–1)
--max-attempts	3	Max regeneration attempts
--api-key	(env/file)	Gemini API key
--verbose	(flag)	Show memorability scores per attempt

---

Prompt Tips for High-Memorability Images

Research shows these compositional elements consistently score higher with ResMem:

• Faces and eyes — Human faces, especially with direct gaze, are inherently memorable
• Unusual juxtapositions — Unexpected scale, context, or combination of objects ("a whale floating through a city skyline")
• Strong focal point — One clear subject against a contrasting background
• High contrast — Bold color separations between subject and background
• Emotional resonance — Images that imply narrative or emotion
• Unusual lighting — Dramatic shadows, golden hour, bioluminescence, neon
• Unexpected scale — Macro details of normally-seen-large things, or vice versa

Prompts that tend to score low: generic landscapes, symmetrical compositions, neutral palettes, cluttered scenes with no clear focal point.

---

Memorability Science

ResMem — Brain Bridge Lab, University of Chicago © 2021 The University of Chicago. Non-commercial use license. https://github.com/Brain-Bridge-Lab/resmem

ResMem is a deep learning model trained to predict image memorability scores — how likely a person is to remember having seen an image after a brief exposure. Scores range from 0 (instantly forgotten) to 1 (highly memorable).

License: ResMem Non-commercial License — redistribution permitted for non-commercial purposes with attribution. For commercial licensing: [e-mail]

---

Built for Claude Code. Requires a Gemini API key and a local Python environment with resmem installed.