Truncate Kronecker

Kronecker-factorized adapters - LoKr in LyCORIS terms - became the community's default for character LoRAs on the newer architectures, and for good reason: they pack a lot of expressiveness into very few parameters. Truncate Kronecker lets you take an existing delta and re-factorize it into that same Kronecker form, instead of training one from scratch. Same family as Truncate Rank, but a different compression philosophy.

What it actually does

Where Truncate Rank keeps the top rank SVD directions, Truncate Kronecker does a rank-1 decomposition with a twist: it reshapes the tensor and factors it as a single Kronecker product w1 ⊗ w2, then rebuilds. The kronecker_ratio (default 0.5) controls the shape of the split - it picks the factor dimensions that land closest to that balance, from the tensor's integer divisors, so the decomposition stays square-ish instead of lopsided. It's aggressive compression: one rank-1 pair per tensor, no rank dial to tune. That's why it's compact, and why it's very much a "know what you're doing" node.

It shares its controls with Truncate Rank: use_approximate_basis (default on) swaps the exact SVD for torch.svd_lowrank with approximate_basis_iters and approximate_basis_seed, and there's a cache input for a Cache Unit so re-running at a different kronecker_ratio reuses the previous factorization. Output is a recipe in the delta merge space.

Wiring it up

Feed it a delta - a Mecha Lora Recipe load, or a subtract of two models - and the Kronecker-truncated delta comes out. From there it merges back into a base via a weighted-sum recipe into Mecha Merger (UNet + text encoder out). The practical use: take a regular LoRA you already have, fold it into Kronecker form, and see whether the compressed version keeps the likeness - the KB's read on LoKr is that it converges better on post-Flux architectures, so converting can genuinely win.

Installing it

This is one node in the Mecha Merge Node Pack by Louis-Jacob Lebel (ljleb), author of the sd-mecha library underneath. Install via ComfyUI Manager (search "mecha" → "Mecha Merge Node Pack") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Only dependency is sd-mecha==1.1.7. No model downloads - it works on the LoRAs and checkpoints already in your ComfyUI models folder.

Gotchas

This is the fiddliest of the three. The input must be a delta; the node validates the merge space, but you still need the mental model that a full checkpoint won't behave here. kronecker_ratio is the dial everyone mis-sets first - it's not a strength, it's a shape parameter, and the effect is subtle until the factorization goes lopsided and quality drops. And single-factor Kronecker is a lossy hammer: great for a compact character LoRA, wrong tool if you need fine control over what's preserved. The merge_checkpointing toggle (fp16 CPU caching of the branch) is there like on the other recipe nodes, but for a one-shot conversion it rarely earns its memory. Use this node when you specifically want Kronecker form - otherwise Truncate Rank is the more predictable first stop.