CoolFace
Apppublic

arno-05/aesthetic_engine

sourceHugging Faceupdated 5mo agoView on Hugging Face
0likes
App README

AestheticEngine

An intelligence platform that reduces the cost of content failure for creators.

Most marketers lose money on ads not because of bad strategy, but because of poor image quality. The human eye makes attention decisions in milliseconds**. AestheticEngine makes those decisions measurable.

What it does

Upload a photo. Get four things back: | Panel | Engine | Output | | Creator | Aesthetic Engine | Quality score (0–10) + improvement tips | | Marketer | Saliency Engine | Heatmap showing where the eye actually goes | | SEO | Semantic Engine | Top 5 strategic tags via CLIP cosine similarity | | Identity | Style Profiling Engine | Your artistic signature across your body of work |

The Problem

"Marketers can't predict what makes a user stop scrolling."

A $100 ad budget on an image where the viewer's eye goes to the background and not the product, is $100 lost. AestheticEngine gives you the science before you spend the money. And for creators, you've never been able to see your own style objectively. Not as a feeling. As data.

The Four Engines

Aesthetic Engine - Quality Scoring

Scores a photo on a 0-10 scale using a neural image assessment model trained on human aesthetic ratings (LAION-Aesthetics). Returns a grade + actionable tips to improve the image before publishing.

"This photo is a 8.2/10 - High Quality."

Saliency Engine - Attention Mapping

Generates a heatmap showing where the human eye is forced to look. Calculates the "center of gravity" of attention and tells you if it is on your product or somewhere else rather than on the product being sold.

"Viewers are looking at the bright coloured car in the background, not the product. Darken the background to redirect attention."

Semantic Engine - SEO Tagging

Uses OpenAI's CLIP to bridge pixels and search engines. Finds the top 5 keywords with the highest cosine similarity to your image's visual features, tags that aren't just descriptive, but strategically searchable.

"Top tags: minimal interior, warm lifestyle, editorial home, soft natural light, Scandinavian aesthetic."

Style Profiling Engine - Artistic Identity

Upload a batch of images. The engine extracts a style embedding from each using CLIP, then finds what is dominantly and consistently true across your entire body of work,not an average, but a signature. Scores your work across six dimensions:

  • —Tone - warm vs cool
  • —Contrast - dramatic vs flat
  • —Complexity - minimalist vs busy
  • —Subject Distance - intimate vs wide
  • —Mood - melancholic vs bright Then anchors your identity to a known visual tradition: Editorial, Street, Commercial, Fine Art, Documentary, Fashion.
"Your signature style is Warm Editorial Minimalism. Across your uploads, your work consistently shows high warmth, low visual complexity, and intimate subject distance. This places you closest to the editorial fashion photography tradition."

Architecture

AestheticEngine/
├── data/                  # Test images (failed vs success)
├── models/                # Pre-trained weights (CLIP, MSI-Net)
├── src/
│   ├── engines/
│   │   ├── aesthetic.py   # LAION-Aesthetics — quality scoring
│   │   ├── saliency.py    # MSI-Net — attention heatmap
│   │   ├── semantic.py    # CLIP — SEO tag generation
│   │   └── identity.py    # CLIP — style profiling & artistic identity
│   └── pipeline.py        # Visual Audit orchestrator
├── requirements.txt
└── .env

Build Phases

Phase 1 - ML Prototyping

  • —[ ] Initialize Python environment and project structure
  • —[ ] Aesthetic Engine - quality scoring
  • —[ ] Saliency Engine - heatmap generation
  • —[ ] Semantic Engine - CLIP SEO tagging
  • —[ ] Style Profiling Engine - artistic identity across batch
  • —[ ] Connect all engines in pipeline.py

Phase 2 - API Layer

  • —[ ] Set up FastAPI application
  • —[ ] Expose /audit endpoint (single image)
  • —[ ] Expose /audit-batch endpoint (multiple images + identity)

Phase 3 - Web Application

  • —[ ] Initialize Next.js project
  • —[ ] Build Creator Panel (aesthetic score + tips)
  • —[ ] Build Marketer Panel (saliency heatmap overlay)
  • —[ ] Build SEO Panel (top 5 tags)
  • —[ ] Build Identity Panel (style report + genre anchor)

Stack

  • —ML: PyTorch, HuggingFace Transformers, CLIP, MSI-Net, LAION-Aesthetics
  • —API: FastAPI (Phase 2)
  • —Frontend: Next.js (Phase 3)
  • —Cost: $0; entirely open-source.

Building in public

Zero-cost. Open-source. Built to make art measurable.