01
Prepare the portrait
Your upload is prepared at the same 320×320 resolution used during training and validation.
Inside Echolens
From photo upload to a useful rating.
Echolens Ratings uses specialized computer-vision models to turn a portrait into one consistent comparison signal. Here is what happens to your photo, what the models learned from, and how we decide whether a prediction is useful.
Private by design. Photos are processed for rating and are not stored or reused for training.
100K+ Human-scored portraits
2 Specialized rating models
320×320 Model input resolution
Seconds Typical scoring time
Your Rating
What happens after you upload.
The product workflow is intentionally short. The technical work happens behind the scenes so the result is easy to use.
02
Run the model
A specialized ConvNeXt-based regression model reads the portrait and predicts one continuous score.
03
Compare the result
The output is mapped onto a familiar 0–10 scale so you can rank similar photo candidates consistently.
What We Optimize
Built to rank a shortlist, not flatten it.
The useful question is rarely whether a photo is simply “good.” It is whether one candidate creates a stronger visual signal than another.
Continuous scoring
The model predicts a continuous value instead of placing photos into broad good-or-bad buckets.
Held-out validation
More than a thousand photos per model are reserved for testing and never used as training examples.
Ranking first
Checkpoints are selected for Spearman rank correlation: how closely the model orders photos the way people do.
Under the Hood
How the model was built and tested.
The deeper technical story, explained without hiding the decisions that shape the final score.
Specialization 01
Two models, one job
Echolens Ratings runs two separate rating models: one trained exclusively on women's portrait photos and one trained exclusively on men's. They share the same architecture and training recipe, but they do not share fine-tuning data.
Specializing the models avoids averaging different visual patterns into a single prediction system and keeps each score focused on the photos that model learned to judge.
Foundation 02
Starting with general visual knowledge
Each model uses a ConvNeXt vision backbone pre-trained with CLIP on billions of image–text pairs from LAION-2B. That foundation already recognizes broad concepts such as faces, lighting, composition, and sharpness.
Fine-tuning then narrows that general visual knowledge into one task: predicting how a portrait photo is likely to be perceived.
Training data 03
Learning from 100,000+ rated portraits
The fine-tuning dataset contains more than 100,000 portrait-style photos with human-assigned scores. Those scores are normalized onto a common scale before training.
The objective is not to memorize individual faces. It is to learn recurring visual relationships across a large collection of scored portraits.
Training recipe 04
Teaching the model what should stay consistent
During training, photos are mirrored, gently shifted in color and lighting, slightly rotated, and zoomed. These variations encourage the model to learn durable visual signals instead of exact pixels.
A robust loss limits the influence of unusual labels, while the pre-trained backbone is updated more gently than the new scoring head so useful visual knowledge is refined rather than overwritten.
Evaluation 05
Choosing the model that ranks photos best
After each pass through the training data, the model is evaluated on held-out portraits it has never seen. The retained checkpoint is the one with the strongest rank correlation, not simply the lowest raw error.
That choice matters because Echolens Ratings is designed to compare a shortlist. A useful model must preserve meaningful differences instead of predicting an average score for everything.
Inference 06
Turning an upload into a private score
The selected model is exported to ONNX and deployed on our servers. Your photo is processed in memory at the model's expected resolution, then its raw output is mapped to the rating scale using training-set statistics.
The image is used only to return the rating. It is not stored, published, placed in a gallery, or reused for model training.
A clearer comparison.
- Ranking several strong photos meant for the same profile or audience.
- Finding the candidate with the clearest overall visual signal.
- Making repeated comparisons with one consistent scoring system.
A promise about outcomes.
- A guarantee of matches, interviews, engagement, or any other outcome.
- A universal judgment of attractiveness or personal worth.
- A replacement for context, audience fit, or your own final decision.
See It in Practice
Bring your strongest candidates.
Rate a few photos made for the same goal, compare the scores, and keep the result in context when you make the final choice.
Privacy notice: your photos are used only for rating and are not stored, published in public galleries, or used as marketing examples. Read the Privacy Policy