Pixazo Research · Generative-AI Benchmarks

The AI Model Leaderboard

Independent benchmarks of the leading generative models across image, video, audio and 3D. Every model is run by our research team and scored by automated ensembles under one fixed rubric — a controlled, reproducible complement to crowd-voted human-preference arenas.

4Categories

48Models ranked

10Benchmark tracks

EloBradley–Terry ranked

Explore a benchmark

Four independent leaderboards. Open any category for its full interactive ranking, all tracks, and the per-category method.

How decisive is each race?

The Elo gap between the #1 and #3 model on each board. A ~100-point gap means the leader wins about 64% of head-to-head matches — so a short bar is a photo-finish and a long bar is a clear front-runner. Gaps are comparable across benchmarks even though the raw scores are not.

AI 3D Model Generation162Elo AI Video Generation121Elo AI Audio Generation51Elo AI Image Generation6Elo

How the models are ranked

The same method underpins every board — automated, reproducible, and calibrated within each benchmark.

Run by our research team

Every model generates the same prompt groups. Pixazo's team curates the prompts, runs each model, and reviews the outputs — the benchmark is re-run whenever a major model launches.

Scored by automated ensembles

Each output is judged by a fixed ensemble of open scorers, not by public voting — so a re-run gives the same result and no single rater's bias can tilt a rank.

Per-group Elo + Bradley–Terry

Outputs are matched within each prompt group, converted to Elo, and cross-checked with a Bradley–Terry fit. Overlapping intervals are ties — read the tier, not the exact order.

Calibrated within a benchmark

An image Elo and an audio Elo are not comparable — each benchmark has its own scale and judges. Compare models inside a category, a controlled complement to crowd-voted arenas.

Category	Scoring method	Models	Judge ensemble
AI Image Generation	6-model vision-language ensemble	15	PickScore · HPSv2 · ImageReward · VQAScore · VLM · CLIP
AI Video Generation	Motion-aware VLM panel	10	Qwen2.5-VL on the clips · RAFT optical flow
AI Audio Generation	Audio scorer ensembles	14	MuQ-Eval · CLAP · UTMOS · Whisper WER · Audiobox
AI 3D Model Generation	Multi-view VLM rubric	7	Qwen2.5-VL-32B rubric · trimesh/pymeshlab mesh hygiene

Who leads across the board

Models benchmarked by provider, counted straight from each board's own ranking — 37 models across the four boards' primary tracks.

Google9

Alibaba3

Black Forest Labs2

ByteDance2

Deemos2

Kuaishou2

Microsoft2

OpenAI2

Current track leaders

Benchmark track	Current leader	Elo
Image · Text-to-Image	GPT Image 2.5 OpenAI	1330
Image · Image Editing	GPT Image 2.5 OpenAI	1330
Video · Text-to-Video	Gemini Omni Flash Google	1212
Video · Image-to-Video	Gemini Omni Flash Google	1183
Video · Video-to-Video	Happy Horse 1.0 Alibaba	1200
Audio · Text-to-Music	Lyria 3 Google	1036
Audio · Text-to-Song	ACE-Step XL ACE Studio	1029
Audio · Text-to-Speech	ElevenLabs v3 ElevenLabs	1078
3D · Text-to-3D	Hunyuan 3D Pro Tencent	1451
3D · Image-to-3D	Hunyuan 3D Pro Tencent	1451

Frequently asked questions

How are AI models ranked on this leaderboard?

Every model is run by Pixazo's research team on the same prompt groups, then scored by a fixed ensemble of open scorers under one rubric — a vision-language ensemble for image, a motion-aware panel for video, audio ensembles for music/song/speech, and a multi-view rubric plus mesh-hygiene checks for 3D. Ensemble wins become per-group Elo with a Bradley–Terry cross-check.

What does the Elo score mean?

Elo is a relative rating from pairwise comparisons. ~1000 is the per-benchmark baseline and a 100-point gap means the higher model wins about 64% of head-to-head matches within that benchmark. Scores are calibrated within a category, so compare models inside a category — not across them.

Are these rankings independent?

Yes. The scoring ensembles are automated and identical for every model, and the benchmark is re-run whenever a major model launches. Pixazo's own models compete under the same rubric — Tracks sits second on the text-to-song board and Pixal 3D mid-table — shown at their real positions with no adjustment.

Which AI image generator is best in 2026?

OpenAI's GPT Image 2.5 leads text-to-image at 1330 mean Elo, with gpt-image-2 (medium) and Microsoft's MAI-Image 2.6 within a few points — the top of the board is a tight pack rather than a runaway.

Which AI video model is best in 2026?

Google's Gemini Omni Flash leads text-to-video at 1212 mean Elo, followed by Alibaba's Wan 3.0 and Black Forest Labs' flux-3-video.

Which AI music generator is best?

Google's Lyria 3 leads the text-to-music track at 1036 mean Elo, ahead of Lyria 2 and Lyria 3 Pro. Song and speech are ranked on their own tracks.

Which AI 3D model generator is best in 2026?

Tencent's Hunyuan 3D Pro leads at 1451 mean Elo, followed by Deemos' Hyper3D Rodin 2 and Meshy 6 Preview, across the text-to-3D and image-to-3D tracks.