AI Engineer World's Fair 2026
Evaling Video Slop
Loading slides…
Read the talk
Evaling Video Slop: Put Judgment Inside the Generation Loop
Video quality depends on what happens across time. Maor Bril walks through building a fast evaluator, discovering its shortcuts, and repairing the data behind its judgments.
From a talk by Maor Bril
Before you start: Basic familiarity with generative models, evaluation datasets, and model training will help; no prior video-processing experience is required.
Generating a video is easier than deciding it is good
How do you decide whether a generated video is good enough to use? Kling, Seedance, Veo, and Sora have made generation increasingly capable, but evaluation often still means watching the result and making a judgment by eye. Maor Bril, introducing this work after just over two years at Character, starts with that gap: improvements in generation have outpaced the tools for judging what was generated.
Compared with studio production, Bril characterizes video generation as almost free. Producing usable footage remains harder. A character gains an extra limb; a door appears to open and close simultaneously; someone hovers instead of moving under plausible physics. Human review catches these failures, but it also becomes part of the production cost. Long-form generated video typically emerges from many short generations and substantial editing, rather than one successful request.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Frame agreement does not establish a coherent story
The first evaluation tools come from image and text workflows. CLIPScore measures image–text compatibility, making it useful for asking whether a frame matches a prompt. LPIPS measures perceptual distance between images; applying it across frames can flag changes, although those changes may be legitimate motion rather than unwanted drift. These are useful signals, but neither establishes whether a sequence tells the intended story.
Video introduces requirements that only become meaningful across time:
| Requirement | Concrete question |
|---|---|
| Story and physics | Does the character walk downstairs, or hover? |
| Character continuity | Is this the same character across shots? |
| Pacing | Does moving between places take plausible time? |
| Audio synchronization | Does the door-slam sound coincide with the visible impact? |
A sequence can contain individually convincing frames while failing these checks. Video evaluation has to assess events and their relationships, not just the appearance of their component frames.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build a repeatable judge before making it fast
A multimodal foundation model can inspect a video and answer questions about it, but simply using an LLM as a judge introduces another source of variation. Different people write different prompts, and the same model can produce very different judgments. Asking whether a clip is consistent or follows its prompt also leaves the broader quality question underspecified. Bril's first iteration was therefore a repeatable evaluation harness rather than an isolated model call.
The harness combines frame metrics with a consistently prompted LLM judge. Humans annotate the reports it produces, and their feedback goes back into the judge's prompt to align its assessments with the annotators' preferences. This creates a rerunnable process with a calibration loop: generate a report, inspect its judgments, correct them, and use those corrections in subsequent judging. The remaining problem is operational. Running the full process is slow and expensive, especially when users need to generate and revise many clips during creation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Catch drift before it becomes a finished video
Evaluation becomes more useful when it moves closer to the user and earlier in generation. Consider two starting frames intended to seed different shots. If the character's identity has already drifted between those images, the defect is visible before either video is generated. Correcting the starting frame avoids paying to animate an inconsistency that was present from the beginning.
The same principle applies at the next level of assembly. A three-, four-, or five-minute video may contain many short clips. If a six-second generation drifts, evaluate and regenerate that clip before combining it with the rest. This does not eliminate later checks: story coherence, pacing, and sound alignment need temporal context. It places each check at the earliest point where the necessary evidence exists, so a local failure can receive a local correction.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Distill explanations into a small VLM
To make those checks practical near generation time, Bril distilled the committee of experts into a small vision-language model, or VLM. The desired output is more than a binary judgment that a video is low quality. It should explain the failure: an extra limb, a violation of physics, or audio that is out of sync. Those diagnoses give the creation process something specific to repair.
Bril reports approximately three seconds to score a 15-second video with the trained small VLM. The talk does not specify the checkpoint, hardware, frame sampling, or timing boundaries for that result. A larger model performed better in his tests, but was significantly slower; the additional quality did not justify the delay for this workflow. The selection criterion was useful judgment at creation-time latency, rather than the strongest evaluator regardless of response time.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Train on comparisons, then inspect what they teach
Absolute quality scores introduce ambiguity before training even begins. Ask several people to rate a video's storytelling from one to ten, and the same clip might receive a four, five, six, or eight. Bril argues that asking which of two videos tells a better story is easier to agree on. The training data therefore used pairwise preferences—A versus B—rather than relying on a shared interpretation of an absolute scale. High-quality internet videos supplied positive examples; negatives came from corrupting good footage or generating poor footage.
The first version was confidently wrong. Bril reports that V1 awarded camera work 9.2 for a four-second shot in which the camera never moved. The example looked like a still image of the same character, yet the evaluator praised its cinematography. Other videos received positive physics judgments despite hovering ghosts and flying people.
The failures exposed what the dataset had actually taught. V1 learned surface coherence, artificial artifacts, and visual gloss instead of the intended quality axes. Pairwise training had not prevented shortcut learning: if the easiest difference between the preferred and rejected examples is their finish, a model can learn to reward that finish without learning storytelling or physics. The comparison format helps define a judgment; the data still determines which evidence the model uses to make it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Repair the dataset without building an AI detector
The next revision paired real footage with AI footage. Bril had initially avoided that approach because it introduces an obvious confound: if human footage always means good and AI footage always means bad, the evaluator can become an AI detector rather than a video-quality judge. Recognizing provenance would satisfy the training labels without solving the intended task.
He applied two safeguards:
- Consistent encoding: Put both sides of the comparison through consistent encoding so artificial differences in the files do not provide an easy cue.
- Consistent annotation: Use the same method to annotate every evaluation axis on both videos, rather than judging real and generated footage under different standards.
Bril reports that the revised dataset produced a substantial improvement, but supplies no numerical validation result. The repair targets the source of the earlier failure: the relationship between the examples, the labels, and the evidence the evaluator can exploit.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Give the generation agent a way to verify its work
With a usable evaluator, the creation system moved from a complex fixed pipeline to an agentic workflow. A pipeline can work well for a narrowly defined use case, but users introduce their own stories, characters, images, and voices. Those variations push the process beyond its original assumptions. Quality-validation tools let an agent inspect what it has produced, adapt to those differences, and repair problems as it goes.
The resulting design connects three decisions: compare alternatives, evaluate the axes that actually matter, and put those evaluations inside the generation loop. Storytelling, pacing, and physics need explicit supervision; they should not be expected to emerge from a generic quality score. The final criterion for the prepared talk is the intended story: the evaluator should help the system produce a video that communicates it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Audio alignment is narrower than understanding sound
The audio questions make the evaluator's boundaries more concrete. Bril describes a separate check for sound quality and intelligibility, followed by matching visual events to audio. The model can receive a video alone or the video together with its generation prompt. If the prompt describes a door slamming, it can guide the model toward the corresponding visual key frame.
The next operation is temporal correlation. In Bril's example, the door slams at frame six. That frame has a timestamp, and the evaluator looks for a spike in the audio at that time. Detecting a coincident spike does not establish that the sound is actually a door slam; it checks alignment without fully identifying the event acoustically.
Once visual-event and audio-peak timestamps are available, the alignment check can be expressed directly. This Python example uses an illustrative door-impact timestamp of 1.20 seconds, candidate peaks, and an application-selected tolerance of 0.08 seconds:
python
from dataclasses import dataclass
@dataclass(frozen=True)
class Alignment:
nearest_peak_s: float | None
offset_s: float | None
within_tolerance: bool
def check_alignment(
event_s: float,
audio_peaks_s: list[float],
tolerance_s: float,
) -> Alignment:
if tolerance_s < 0:
raise ValueError("tolerance_s must be nonnegative")
if not audio_peaks_s:
return Alignment(None, None, False)
nearest = min(audio_peaks_s, key=lambda peak: abs(peak - event_s))
offset = nearest - event_s
return Alignment(nearest, offset, abs(offset) <= tolerance_s)
door_impact_s = 1.20
audio_peaks_s = [0.40, 1.24, 2.10]
result = check_alignment(door_impact_s, audio_peaks_s, 0.08)
print(result)
The signed offset distinguishes an early peak from a late one. The function intentionally returns an alignment result, not a claim about which sound occurred.
Lip synchronization remains unresolved in Bril's system at the time of the talk. Human talking heads at least offer identifiable lips for analysis. Animated characters are harder: their mouth movements may have no direct correspondence with human speech articulation. The team is still working through that problem rather than treating event-level audio alignment as a lip-sync solution.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep human taste in the calibration loop
Human evaluation introduces its own question: whose taste should the judge learn? Bril's answer returns to JudgeJudy, the report-and-annotation harness. People can inspect its reports and annotate the judgments. The team periodically spends 10–15 minutes annotating videos across multiple axes. Rather than asking everyone to evaluate one video on ten dimensions, Bril randomizes the annotation tasks.
Those annotations calibrate the AI judges, whose resulting judgments then become training data for the next version of the model. This makes alignment an ongoing process rather than a one-time label collection exercise. It takes time, and it does not immediately dissolve disagreements over taste: something Bril considers excellent may leave another reviewer unconvinced. Human feedback continually adjusts the evaluator as those preferences become visible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose distillation for the economics of the workflow
The small VLM came from the Qwen family. Bril had tried other models, but the team already had good experience post-training Qwen for other use cases. The available setup worked, and its performance was good enough. He does not identify the other candidates or the exact Qwen variant; the choice is presented as a practical fit rather than a comprehensive model ranking.
An audience member then asks whether the approach makes sense for a domain with hundreds or roughly a thousand videos. Bril separates serving capacity from the reason to distill. He describes deployment ranging from one model instance on one GPU to a hundred instances, but his primary motivation was speeding up creation. A smaller operation can still use the committee of experts—CPU and GPU metrics combined with frontier-model judges—without training its own distilled evaluator.
| Approach | Main investment | Operational trade-off |
|---|---|---|
| Expert committee | Integrate metrics and judges | Slower evaluation; potentially stronger judgments |
| Distilled small VLM | Curate data and train a model | Faster evaluation, plus model-serving costs |
The decision balances dataset curation and training effort against serving cost, latency, and evaluator quality. Bril speculates that a newly available model might judge better, but offers no measured comparison. Expensive judging may be acceptable for one or two videos; at thousands or tens of thousands of evaluations per day, its costs accumulate. There is no fixed video-count threshold in the talk: the relevant question is whether faster creation and repeated use repay the investment in the smaller evaluator.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A reusable harness, with tracing still requested
The final question concerns OpenTelemetry traces from the LLM judges. Bril confirms that the repository does not export those traces at the time of the talk. He describes the repository as a harness that can connect different agents and LLMs; internally, the team also runs a service version with an agentic harness and the metrics it needs.
Bril accepts the request and commits to adding OTel telemetry. The public README inspected on August 28, 2026 documents OpenTelemetry metrics export, which does not establish delivery of the requested judge traces. The closing integration question matters because evaluation is now part of creation itself: the harness must fit the surrounding agent system, and its judgments need to be observable there.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Multimodal evaluation harness combining automated metrics, AI judges, human annotations and calibration, with runnable examples.
Original paper on evaluating image–caption compatibility without reference captions.
Official implementation and examples for measuring perceptual distance between images.
Further reading
Speech mean-opinion-score prediction system with pretrained inference examples from the VoiceMOS Challenge 2022 work.
Official vision-language model repository with model variants, video-understanding examples and fine-tuning guidance.
Full transcript
Loading transcript…