Trending & LM Arena Rankings

The Browse panel in modelBridge includes ranking data from the LM Arena leaderboard, a community-driven benchmark where real users compare AI model outputs head-to-head. This gives you a signal beyond marketing claims — which models are actually performing well according to people using them.

How it works

modelBridge’s backend fetches the latest LM Arena dataset from HuggingFace nightly. The dataset includes category-specific rankings — image-to-video, text-to-video, text-to-image, and more. These rankings are matched against fal.ai models available in modelBridge and surfaced as metadata in the Browse panel.

When you browse models, ranking data appears alongside each model’s card — giving you context about how the model compares to alternatives in its category.

What the rankings mean

LM Arena rankings are based on blind head-to-head comparisons. Users are shown outputs from two anonymous models and asked to pick the better result. Over thousands of comparisons, models earn an Elo-style rating that reflects real-world preference — not synthetic benchmarks or self-reported scores.

This means:

A model ranked #1 in image-to-video has been preferred over other models by real users in blind tests
Rankings shift as new models launch and existing models improve
Rankings are category-specific — a model can rank highly for video but not for image generation

Categories covered

Rankings are available for the model categories that LM Arena evaluates:

Image to Video — animating still images
Text to Video — generating video from text prompts
Text to Image — generating images from text

Not all categories have LM Arena coverage. Categories without rankings (audio, speech, video-to-video) rely on other signals in the Browse panel — like the NEW badge for recently added models and model descriptions.

How rankings stay current

The ranking data updates automatically every night via a scheduled job. You don’t need to update the plugin or take any action. When you open the Browse panel, you’re always seeing recent ranking data.

The matching between LM Arena model names and fal.ai endpoints is maintained via an alias table — so even when leaderboard names differ from fal.ai endpoint IDs (which is common), the correct models are matched.

Using rankings in your workflow

Rankings are one signal among many when choosing a model. Consider:

Top-ranked models are a strong starting point when you’re new to a category
Rankings don’t account for speed or cost — a #1 model may be slower or more expensive than a #3 model that’s nearly as good
Your footage matters — a model that ranks well on general benchmarks may not be the best fit for your specific content. Use Dual Mode to compare top-ranked models on your actual clips