Skip to content

Latest commit

 

History

History
95 lines (69 loc) · 3.79 KB

File metadata and controls

95 lines (69 loc) · 3.79 KB

DIVISOR

Scope

    Computer software
    Consumer‑grade hardware target
    Mid-range devices (Laptop/Desktop)
    Open source
    Generative AI
    Local & offline generation
    Diffusion models only
    Flow‑match schedulers
    Multimodal output support


Philosophy

    Diffusion is a Journey: Walk your own path
    Principally Subjective: Output is complete when the operator declares it
    Feedback: Tangible, rapid system feel. Instantaneous creativity demands instant tools
    Keep Flow State: Perpetuate the disruptive and unique as long as possible
    Transformation & Iteration: Exploration over refinement
    The End Game: Tractable chaos equilibrium


Objectives

    Step‑wise diffusion: edit parameters while generating (live latent walk)
    Generation window: ≤1s inference for all use‑cases
    Intermediate Preview = Final Product: always display/save current state
    Input: adjust all parameters within the time for Generation window
    Corruptions: ≥10 significantly transformative parameter adjustments (one per finger)
    Code: abstracted from model frameworks, generalizable to any diffusion modality
    Hackable & Branching: Reversible and revisionable output
    Create Value: Store and share every edit and the process leading to it

Related readings: https://diffusionflow.github.io/ Diffusion Meets Flow Matching Flow matching and diffusion models are two popular frameworks in generative modeling. Despite seeming similar, there is some confusion in the community about their exact connection. In this post, we aim to clear up this confusion and show that diffusion models and Gaussian flow matching are the same, although different model specification...

Shadowbox What it does -

Its an ML system that runs models using the fastest library per system type
No-code system, self-arranges to run more models than I even know about
It creates images and generates text
Its totally local and offline (ie private)
As libraries update, it immediately supports them

Where we are -

We are able to recognize any kind of open source AI model in Transformers, Diffusers, and GGUF, and some MFLUX/MLX, ONNX models too
All models supported run right out of the box with no questions asked
We are able to generate several types of media with no configuration required

Why we are here -

share workflows certain to work on any computer (impossible w/o docker & even often with)
create optimized, auto-arranged workflows
hand this system to a non-experienced person and they can use it
allow people to explore with guidance from the system with no fear of breaking it
concentrates and distributes legacy model knowledge
future proof due to versatility of upstream channels, always-up-to-date
deeply reduce complexity and learning required to get to peak performance with any model (otherwise impossible), and then automate its construction per system

Are we there yet? -

For this to run at all/have any purpose on most computers, it needs :

memory management - component recognition done, pipe needs to sequentially load
model chaining - automatic chaining of models towards specific tasks

Goals -

add more libraries and the ability to install them via sdbx
add a fuck ton of upscale/3d/audio/svg/midi models etc (things that arent already recognized by diffusers/transformers/etc)
modularity - tasks need to be dynamically looked up and added to model chain
deep analysis of models

https://darkshapes.org/docs/divisor/divisor/ https://darkshapes.org/docs/zodiac/zodiac/ https://darkshapes.org/docs/mir/mir/ https://darkshapes.org/docs/nnll/nnll/