Block-wise training
Train multiple network blocks simultaneously via diffusion interpretation, instead of one monolithic end-to-end backward pass.
← Store · // product
unifur
unifur (DiffusionBlocks++) is a Rust crate on the Burn framework for block-wise neural-network training: parallel depth denoising, consistency regularization, flow-matching objectives, mixture-of-experts routing and block distillation. Support and integration from $499.
from $499support & integration
⚠ Research-grade softwareunifur is experimental ML research code that reproduces and extends a recent paper. Expect to validate results on your own models and hardware. Provided as-is, without warranty.
// what it is
Train the blocks in parallel, not end to end
unifur is DiffusionBlocks++: block-wise neural-network training that uses parallel depth denoising trajectories, consistency regularization, flow-matching objectives, mixture-of-experts routing and block distillation — aiming to match end-to-end backpropagation while training multiple blocks at the same time.
It is built as a Rust crate on the Burn framework, and is based on the ICLR 2026 paper DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation (Shing, Koyama, Akiba) and Sakana AI’s original implementation.
The repository documents a consolidated, research-backed pipeline — shared vector space, scene graph, generation, training-efficiency stages, optional reinforcement learning and output translation — with references to the underlying work, so what it does and why it does it stay together.
// capabilities
What it does
Parallel depth denoising
Depth-wise denoising trajectories with consistency regularization keep independently trained blocks coherent.
Flow-matching objectives
Flow-matching training objectives for stable, efficient block optimisation.
Mixture-of-experts routing
MoE routing at the block level so capacity goes where the data needs it.
Block distillation
Distil trained blocks for smaller, faster deployment targets.
Rust on Burn
A typed, memory-safe Rust crate on the Burn framework — the same language family as our Peregrine inference engine.
// under the hood
Specifications
- Language
- Rust
- Framework
- Burn
- Based on
- DiffusionBlocks (ICLR 2026) — Sakana AI
- Type
- Crate + documented pipeline
- Pairs with
- Peregrine (inference) and subnext (context runtime)
- Intended use
- ML research, training-efficiency work and experimentation
// what you pay for
Support, not a paywall
unifur is open research code. What you pay for is making it work for your models and your hardware.
- Integration & training setup — stood up on your hardware, wired to your data pipeline and model targets.
- Custom blocks & objectives — new block types, routing or distillation strategies built for your architecture.
- Reproduction & benchmarking — paper results reproduced and measured against your current training approach.
- A direct line to the engineer who built it — no support tier in between.
// the fine print
Terms & liability
As-is, without warranty. unifur is research-grade machine-learning software provided for evaluation and integration. To the maximum extent permitted by law, Cyber Sec and Stephan Botes accept no liability for any damage, data loss or consequence arising from its use.
Credit where due. DiffusionBlocks++ builds on the DiffusionBlocks paper (Shing, Koyama, Akiba, ICLR 2026) and Sakana AI’s original code; unifur is an independent Rust implementation and extension.
What the fee buys. The price shown is a starting point and covers the engineer’s time — integration, custom blocks, reproduction and benchmarking. You pay for the work, not for access to the source.
Questions about scope, licensing, or authorization? Get in touch first.