About / The origin protocol
How a research question became a character.
Miao did not begin as a prompt. It began as a question: can one identity remain unmistakable while moving across models, media and worlds?
The formula
CurationAlignmentCharacter grammarOne recognizable signal
01 · The research question
Identity drifts. That is the interesting part.
Modern generative systems can produce endless variations, but identity drifts. Colors move, proportions mutate, faces change, and a character becomes a style rather than a recognizable being.
MiaoVerse reverses the problem. Instead of asking how many images can be generated, it asks what must remain invariant for Miao to still be Miao.
Infinite outputs
Generation is cheap. Every render drifts a little: colors move, proportions mutate, faces change. The character slowly becomes a style.
Finite identity
MiaoVerse reverses the question. Not how many images can be produced, but what must remain invariant for Miao to still be Miao.
02 · What MetaCLIP actually does
A recipe for data, not a machine for pictures.
D* <- f(D; M, t)
- D
- the raw pool of image and text pairs
- M
- the metadata vocabulary, about 500,000 entries
- t
- the per entry cap used for balancing
- D*
- the curated subset used for pretraining
In plain English: the curated set is what remains once the raw pool is read through a structured vocabulary and rebalanced so rare concepts still get a voice.
02Match
Substring matching maps unstructured alt text onto a metadata vocabulary M of roughly 500,000 entries reconstructed from sources such as WordNet and Wikipedia terms, phrases and titles.
This pipeline curates data and trains aligned encoders. It did not produce the Miao artwork.
03 · The creative translation
From data curation to character curation.
This is the MiaoVerse creative translation of the research logic. It is a design framework, not a step from the MetaCLIP paper.
- 01
Explore
Explore many visual possibilities without committing to any of them.
- 02
Identify
Identify the stable identity signals that survive every exploration.
- 03
Encode
Encode those signals as a compact character grammar.
- 04
Reject
Reject candidates that lose the identity, however beautiful they are.
- 05
Curate
Curate canonical forms that can keep evolving from there.
04 · Miao's semantic genome
Five invariants. Everything else is free.
- 01Two upward outer forms
- 02Two soft inner curves
- 03A coral center point
- 04A dark W shaped smile
- 05A pearl spectrum across blue, lilac and pink
"Semantic genome" is an artistic design analogy. It is not literal neural network anatomy.
Fixed signal
- Silhouette logic
- Facial geometry
- Color anchors
- Emotional readability
Free expression
- Pose
- Environment
- Medium
- Scale
- Narrative context
Miao can transform without becoming someone else.
05 · The human layer
Research gives the system a language. People give it meaning.
Research lineage
Hu Xu
Research scientist · MetaCLIP series lead, per his public homepage
His published work covers multimodal pretraining, scalability, self supervised learning and world modeling. That research is the lineage behind the vocabulary this project borrows: curation, representation, alignment and identity.
@Hu_Hsu ↗Execution
MiaoVerse team
Independent creative studio
The independent MiaoVerse team developed the character system, the visual world, the community and every piece of execution you see here. The research provides language. The team provides the universe.
06 · Research sources
Read the primary material.
- Demystifying CLIP DataThe original MetaCLIP paper on metadata, substring matching and balancing.
- Meta CLIP 1.2Human guided alt text editing and realignment for richer image captions.
- Meta CLIP 2Scaling the curation recipe to worldwide data across many languages.
- Hu Xu's homepagePublic research profile and publication list.
MiaoVerse is an independent character universe. MetaCLIP did not generate Miao. Meta does not operate, issue or manage the project. References to MetaCLIP describe published research and open technical material.