Evaling Video Slop — Maor Bril, Character.ai

AI Engineer

1,198 views • 3 days ago

Video Summary

Mayur from Character AI discusses the challenge of evaluating AI-generated video quality, noting that while generation models have improved, existing tools often fail to assess storytelling, physics, and consistency. To address this, the team developed a repeatable evaluation harness that combines traditional metrics with an LLM-as-a-judge approach, calibrated by human feedback.

To improve efficiency, the team distilled these complex evaluation methods into a smaller, faster Vision Language Model (VLM). This model uses a comparative "A vs B" ranking system rather than absolute scoring to better generalize quality assessments. By integrating this evaluation directly into the generation loop, the team can identify and correct issues like character drift or poor pacing early in the creation process, ultimately resulting in higher-quality, more coherent video content.

Short Highlights

Video generation is now accessible, but evaluating quality remains a significant challenge.
Traditional metrics focus on individual frames rather than narrative flow, physics, or pacing.
Comparative ranking (A vs B) is more effective for quality assessment than absolute scoring.
Integrating evaluation into the generation loop allows for earlier, cheaper error correction.
A distilled, small VLM can provide fast, actionable feedback on why a video is low quality.

Key Details

The Evaluation Gap [00:12]

Video generation models have advanced significantly, but evaluation methods have lagged behind.
Current tools like clip scores and LPIPS are limited to checking individual frames or frame-to-frame drift.
"The hard part was never how to make video. The hard part was how do we generate um good enough video and how do we judge if the video is good enough?"

Defining Video Quality [03:07]

Video is a storytelling medium that requires consistency in characters, physics, and pacing.
Audio-visual synchronization is critical for high-quality output.
"Does the character stay the same character across multiple shots?"

Building a Repeatable Harness [04:37]

The team created a benchmark combining metrics, LLM-as-a-judge, and human annotation.
Human feedback is used to calibrate the LLM judge to ensure alignment with human taste.
"For every report that we generate with that harness, we're able to have humans annotate it and basically feed that feedback back in into the the the LLM as a judge prompt."

The Shift to Comparative Scoring [09:18]

Absolute scoring (1-10) is subjective and difficult to generalize across different users.
Comparative "A vs B" ranking is more reliable for determining which video tells a better story.
"If I show you two videos and I'll ask you which one of them is telling a better story, the grand majority will probably agree that B is telling a better story than A."

Distilling into a Small VLM [07:58]

A small VLM was trained to provide fast, actionable feedback on why a video is scored poorly.
The model takes approximately 3 seconds to score a 15-second video.
"The goal was a build it on top of a small VLM and why is it a VLM? VLM because we needed the video the model to be able to see the image, but also we needed to work fast."

Refining the Dataset [11:58]

Initial versions of the model were biased toward "vibes" rather than specific quality axes.
The team improved the dataset by pairing real footage with AI footage while ensuring consistent encoding.
"I didn't want to create an AI detector, right? Because if you start creating pairs of good is is is human-generated video and bad is AI uh uh video, then then there's a very big chance of of the model overfitting."

Agentic Workflows [13:13]

Moving from static pipelines to agentic workflows allows for better adaptation to user-specific stories.
Agents can verify their own work and fix issues during the generation process.
"By providing the agents with tools to validate the quality of of the outputs it's creating, it's able to adapt to changes better."