Tinder Reveals How Its AI Photo Ranking System Functions

Tinder has shared details on how its Smart Photos feature uses artificial intelligence to rank users’ profile photos, aiming to show the most effective images first.

In a technical blog post, Tinder’s machine learning engineers described the evolution of Smart Photos. Previously based on convolutional neural networks (CNNs), the system now relies on a Vision-Language Model (VLM). This allows the AI to understand not just visual elements like lighting and composition, but also higher-level context such as vibe, social cues, and intent.

The current system treats photo ranking as a pairwise comparison problem. For a user with multiple photos, the model evaluates every possible pair and determines which photo is likely to perform better based on engagement signals like likes and impressions. These comparisons are then aggregated into win counts and average scores to produce the final ranking order.

Training data comes from real user interactions collected through a Multi-Armed Bandit (MAB) approach, which shows different photos in the top position over time to measure performance. The team focused on high-confidence pairs to reduce noise in the training data.

Tinder reported that the VLM-based system achieved approximately 76% pairwise accuracy in offline tests, up from 68% with the previous model. Online A/B testing showed improvements in likes, matches, and conversations for users.

The company also experimented with factors such as LoRA configuration, batch size, image resolution, and whether to train the vision encoder. Larger models (7B parameters) and higher image resolutions generally performed better, though they increased computational cost. Tinder emphasized that users can still manually order their photos or opt out of Smart Photos entirely. The goal is to reduce guesswork for users while helping them present their most effective profile.