chibi-image-nexus-v0.1
Experimental low-rank SVD compression baseline for Agate Preview 003.
What v0.1 is
v0.1 applies a fixed-ratio rank rule to selected large torch.nn.Linear
layers in the Agate Generator.
This repository is intentionally preserved as the baseline that motivated the energy-aware rank selection used in later versions.
Method
For each selected weight matrix:
W ≈ U_r diag(S_r) V_r^T
and the dense layer is replaced by two smaller linear layers.
v0.1 uses a fixed rank fraction rather than choosing rank from the singular value spectrum.
Result
- Base model:
Logolabs/agate-preview-003 - Generator-only compression
- Text encoder / Ettin: not compressed
- VAE / decoder: not compressed
- Saved compressed Generator reload test: passed
- Parameter reduction: see manifest.json
Important limitation
This is not a quality-preserving final model.
The compressed checkpoint can reload and run, but aggressive fixed-ratio compression may significantly alter or degrade generated images.
The purpose of this release is to provide a reproducible baseline for comparison with later automatic rank-selection experiments.
Why v0.2 exists
v0.1 assigns similar rank pressure across layers even though different layers have very different singular-value spectra.
v0.2 instead selects the smallest rank that preserves a target fraction of squared singular-value energy, such as 99%, 99.5%, or 99.9%, and skips layers when low-rank factorization would not actually save parameters.
Files
generator.safetensors— compressed Agate Generatormanifest.json— compression metadata and layer ranksnexus_loader.py— loader for reconstructing the compressed Generator- baseline / compressed test images if produced by the notebook
Example
from nexus_loader import load_nexus
pipe = load_nexus("mikun0202/chibi-image-nexus-v0.1", device="cuda")
image = pipe(
"a minimalist logo of a fox head, orange, flat design, white background",
seed=0,
steps=20,
watermark=False,
)[0]
image.save("output.png")
Status
Experimental / baseline.
Use v0.1 for reproducibility and comparison, not as the recommended quality-preserving compression configuration.