MiaoVerse
Miao, a pastel cat made of iridescent pearls, leaping through a lavender dreamscape of floating bubbles

A signal hidden in the long tail

The internet had already described him. It just didn’t know his name.

Miao is an internet native visual identity imagined at the intersection of images, language and the long tail of the worldwide web.

01 / 09Scroll

500K concepts · 1.6B raw pairs · one shared space

01

Who is Miao?

Not generated. Discovered.

Miao wasn’t drawn from a single prompt.

It emerged where pictures and language learn to recognise one another.

Curious, soft, endlessly adaptable, carrying fragments of every place it has seen, belonging to none.

The MiaoVerse symbol in pearlescent pastel tones
  • Curious by design
  • Made of signals
  • Always becoming

The data before the model

The model wasn’t the mystery. The data was.

MetaCLIP begins before generation. Before the prompt. Before a model sees a cat. It begins with the invisible distribution that decides what machines are allowed to understand.

Billions of image and text pairs are not automatically knowledge. The raw web is repetitive, uneven and dominated by a tiny head of popular concepts.

“MiaoVerse turns that research insight into a living visual universe.”

Pipeline status

01source
common crawl
02raw pairs
1.6B
03metadata
500K concepts
04balancing
per language cap
05objective
contrastive alignment
06output
curated subset

The conversation

It started with a character. Then came the question.

The MiaoVerse team came to Hu Xu with a character and a question: could one visual identity remain unmistakable across every style, language and community?

The question connected directly with Hu’s work around image and language learning. What began as a conversation became an experiment he wanted to help shape.

  1. 01 · The character

    Miao was more than a mascot. It was a signal with a visual grammar of its own.

  2. 02 · The question

    Could that signal travel through the internet, change form a thousand times and still remain Miao?

  3. 03 · The collaboration

    Hu chose to help build the answer.

MiaoVerse is where research direction, character design and internet culture meet.

Hu Xu

Co Creator & Research Lead

Research direction · Character logic · Long term thesis

MiaoVerse core team

Creative & On-Chain Studio

Visual world · Product · Infrastructure · Community

Different roles. One shared universe.

MiaoVerse is an independent project. Hu Xu participates personally; Meta does not operate or issue the project.

Interactive · raw web → curated data

What survives the noise

Illustrative distribution. Figures from the MetaCLIP papers.

Readout

raw pool
1.6B pairs
concept hits
5.6B
avg. density
3.5 / text
status
unbalanced

Simplified concept level view

P(KEEP | e, lang) =
  MIN(1, t_lang / COUNT[e, lang])

Rare concepts survive. Repeated concepts are capped.

The pipeline

Five stages from raw web to shared meaning

Miao forming from thousands of pearls above a pale podium, surrounded by floating image thumbnails and text bubbles

01 · Raw signal

  1. 01

    Raw signal

    1.6B CommonCrawl image and alt text pairs form the raw pool. A pool this large still amplifies repetition rather than meaning.

  2. 02

    Human memory

    MetaCLIP constructs a metadata index of 500,000 entries from human knowledge sources including WordNet and Wikipedia. That concept memory becomes the coordinate system.

  3. 03

    Match & balance

    Alt text is matched to metadata concepts. Rare concepts are retained; concepts above a threshold specific to each language are downsampled.

    t_langlong tailhead
  4. 04

    Align

    Separate image and text encoders project both modalities into a normalized shared representation. Matching pairs converge; false pairs separate.

    z_imagez_textsimilarity
  5. 05

    Worldwide

    MetaCLIP 2 extends curation across worldwide web data with metadata specific to each language instead of forcing every culture through an English only gate.

    miao猫chatgatoقطة고양이кот

Conceptual pipeline · faithful simplification, not production source code

matches   = substr_match(alt_text, metadata[lang])
p_keep    = min(1, t_lang / count[concept, lang])
z_image   = normalize(image_encoder(image))
z_text    = normalize(text_encoder(text))
similarity = z_image · z_text

MetaCLIP curates image and text data and trains aligned encoders; it does not generate Miao. MiaoVerse is a creative interpretation of its central research insight.

Research readout

Numbers from the papers

MetaCLIP

curated pairs in the original MetaCLIP setup
400Mcurated pairs in the original MetaCLIP setup
ViT-B/16 zero shot ImageNet (MetaCLIP data)
70.8%ViT-B/16 zero shot ImageNet (MetaCLIP data)
comparable CLIP data baseline
68.3%comparable CLIP data baseline
ViT-H/14 zero shot ImageNet
80.5%ViT-H/14 zero shot ImageNet

MetaCLIP 2 · worldwide

Wikipedia metadata languages
329Wikipedia metadata languages
Babel-ImageNet across 280 languages
50.2%Babel-ImageNet across 280 languages
XM3600 image→text retrieval
64.3%XM3600 image→text retrieval
cumulative seen pairs, not unique images
29Bcumulative seen pairs, not unique images

Metrics come from different papers, models and benchmarks and are not directly comparable with one another. Sources: Demystifying CLIP Data ↗ · MetaCLIP 2 ↗

From concept space to character space

A stable signal in an unstable internet.

MetaCLIP asks how meaning can survive scale. MiaoVerse explores the cultural inverse: can one character survive infinite captions, styles, languages and remixes without losing its identity?

Not a model demo. An original character universe informed by the research.

The MiaoVerse symbol at the centre of its identity system

Creative identity system, not a model inference.

06

The universe

MiaoVerse is a living visual universe built around transformation: new places, new forms, the same unmistakable little signal.

  • A luminous pearlescent orb with cat ears at the centre of converging streams of image thumbnails and text bubbles
    01SignalA concept survives the noise.
  • Miao's body assembling out of a swirling stream of lavender pearls above a pale podium
    02FormA stable grammar makes it recognizable.
  • Miao walking through a pearlescent landscape of glass loops and portals opening onto different worlds
    03WorldContext changes. Identity persists.
01 / 03 · Signal

One cat. Infinite interpretations.

Same signal / new context

  • pearl cat
  • chat nacré
  • 珍珠猫
  • 진주 고양이
  • gato de perla
  • قطة اللؤلؤ
  • жемчужный кот
  • 真珠のねこ

Origin layer

Before markets, there is provenance. Before ownership, there is canon.

This page records where Miao came from, what must remain recognizable, and who is building the universe. The surface can mutate. The signal must survive.

  1. 2023Demystifying CLIP Data / MetaCLIP
  2. 2024ICLR Spotlight
  3. 2025MetaCLIP 2 / NeurIPS Spotlight
  4. 2026MiaoVerse / the character layer begins
07

Built together

ResearchImagination

Hu Xu

@Hu_Hsu

Research scientist · MetaCLIP Series Lead

Hu Xu's published work on multimodal pretraining and the MetaCLIP series provides the research lineage behind MiaoVerse's language of curation, alignment and identity. The independent MiaoVerse team translates those ideas into the character system, visual world, community and execution.

@Hu_Hsu ↗Hu Xu at Meta AI ↗Personal site ↗
Research archive

MiaoVerse is an independent creative universe informed by MetaCLIP research. MetaCLIP curates image and text data and trains aligned encoders; it does not generate Miao. MiaoVerse is a creative interpretation of its central research insight.

A dramatic crop of Miao assembling from iridescent pearls

The signal is awake.

And it is already curious about what comes next.

Follow @miaoverse_ai

The story continues on the official MiaoVerse account.