01
Passion
Scratch your own itch.
Every project here started because somebody hit a wall and refused to walk around it. There is no roadmap handed down from above. The roadmap is whatever is currently annoying us enough to fix.
About
A volunteer-run, non-profit community building open infrastructure for anime and 2D image research. Founded 2026. No funding round, no revenue target, and no roadmap handed down from above.
This started the way most of these things do: one person, one stubborn bug, and a training script that only worked on one machine. Making that script reproducible turned into a trainer, the trainer turned into a stack, and at some point it stopped being reasonable to call it a side project.
Nobody here is paid and nothing here is sold. It is a community of people who care about the same awkward corner of machine learning: anime, illustration, 2D art, and the models that keep falling over on it. General-purpose vision was trained on photographs and quietly gives up on a drawing, so somebody has to build the taggers, the corpora and the evaluation sets. That is the whole reason this exists.
What holds it together is one commitment. Everything is published: the weights, the corpora, the configs, and the failure cases we would rather nobody saw. Closed research is unreviewable research, and a result you cannot reproduce is a rumour with a chart attached. Being unfunded is not an excuse for being unrigorous.
Three rules, and not many more.
A lab this size does not need a process document. It needs three things everyone agrees on and the discipline to keep them.
02
Joy
Happiness before deliverables.
Research is a long game and burnout ends it early. We keep the pace humane, celebrate the small wins, and let people wander off into strange ideas without asking for a justification memo.
03
Evolution
Stay embarrassingly curious.
The stack we love today is the legacy code of next year. We adopt fast, discard faster, and treat rewriting our own work as a feature of the process rather than an admission of failure.
Six directions, one thread.
Everything here comes back to the same question: how do you make a capable model that somebody else can actually reproduce?
Generative Vision
Diffusion and flow-matching models, controllable generation, and the evaluation harnesses needed to tell whether any of it actually improved.
Anime Image Understanding
Taggers, detectors and aesthetic scorers trained on illustration data, where general-purpose vision models quietly fall apart.
Parameter-Efficient Training
Adapter families, low-rank decompositions and quantised training paths that let a single consumer GPU do work that used to need a cluster.
Model Infrastructure
Registries, artefact stores and scheduling layers. The unglamorous plumbing that decides whether a research idea ever reaches a second machine.
Data Pipelines
Crawling, dedup, captioning and quality filtering at a scale where a bad heuristic costs a week of compute rather than an afternoon.
Open Evaluation
Reproducible benchmarks published with their failure cases attached, because a leaderboard without error analysis is marketing.
How it got here.
Nine months, four turning points and a lot of runs that are not on this list.
- Jan 2026
One stubborn bug
One shared frustration with fine-tuning tooling, and a repository nobody expected anyone else to clone.
- Apr 2026
First release in the wild
MochiCore went public. The issue tracker turned into an accidental research agenda within a fortnight.
- Jul 2026
Generative vision starts
PrismStream split off as its own stack once the conditioning experiments outgrew a single notebook.
- Now
Building the plumbing
Shifting focus from single models to the systems around them. Registries, evaluation harnesses and reproducible pipelines.
The ones we get asked.
Mostly about licensing, joining and whether this is a company. It is not.
No. It is an independent research lab. There is no funding round, no board and no revenue target. Just a public repository and a lot of open pull requests.