HoLo-FuSe — frozen 0-parameter HSL substrate as a diffusion conditioning door
Honest framing first: this is a minimum-scale baseline training run whose only purpose is to prove that HSL (Holistic Signal Language — a frozen, deterministic 27-D feature frame with 0 learned parameters, a 4.6 KB LUT) can serve as the conditioning substrate of a verified diffusion carrier. Not SOTA, not a product, not "HSL beats embeddings". The carrier is a standard class-conditional DDPM; HSL is the thing under test. FuSe = Frozen Substrate, fused into a verified baseline.
Code & full record: Woojiggun/HoLo-FuSe · live demo: ggunio/HoLo-FuSe-demo · the zero door: hsl-embedding-zero (PyPI) · siblings: HoLo_ZeRo (byte-LM), HoLo-ToLk-STT (audio). DOI (software): 10.5281/zenodo.21322659 Author: Jinhyun Woo (ggunio5782@gmail.com) — independent research, developed in collaboration with AI assistants (Claude Code, Codex); the HSL work and experimental direction are the author's.
What was verified (seed-matched, same budget, step 14000)
| arm | conditioning | result |
|---|---|---|
none | unconditional | readable cat+dog faces, mixed |
hsl | frozen HSL 27-D (0 learned params) → small readout | "Cat"→cats, "Dog"→dogs |
learned | same-budget nn.Embedding control | "Cat"→cats, "Dog"→dogs |
- Flipping the label on the same initial noise morphs the sample between species → conditioning works.
- hsl ≈ learned: the frozen substrate steers class as well as the learned control. Claim is comparable, not better — single seed set, qualitative.
- Specificity update (community-review follow-up, multi-seed quantitative): HSL was not distinguishable from sampled same-shape frozen-random codebook controls under this two-class protocol (0.828±0.016 vs 0.818±0.018 consistency @12k, 3 seeds) — no HSL-specific contribution identified — while the learned control scored modestly higher under this evaluator and protocol (0.891±0.027). Precise wording: a zero-parameter frozen substrate with a learned conditioning readout. Full protocol, exact scripts, evaluator confusion, and caveats: SPECIFICITY.md.
- Known artifact: a background color tint in all arms (under-training of a ~35M model at 14k steps; a sampling sweep showed CFG / dynamic-thresholding does not remove it). Doesn't affect the comparison.
Files
| file | content |
|---|---|
holofuse_hsl_128.pt | HSL-conditioned arm (the demo one) |
holofuse_learned_128.pt | learned-embedding control arm |
holofuse_none_128.pt | unconditional baseline arm |
Each ≈274 MB: {model, cond, ema, step, arch} — EMA included (sample from EMA), optimizer stripped.
Arch: U-Net base128, ch_mults 1,2,2,2, attn@16, ~35M params; DDPM cosine T=250; CFG cond-drop 0.15.
Use — everything ships in this repo (code + weights)
pip install torch hsl-embedding-zero huggingface_hub pillow numpy
import sys, pathlib
from huggingface_hub import hf_hub_download
code = hf_hub_download("ggunio/HoLo-FuSe", "model.py") # inference code lives here too
sys.path.insert(0, str(pathlib.Path(code).parent))
from model import generate
img = generate("Cat", steps=16, cfg=1.6, seed=0)[0] # downloads the hsl checkpoint (274 MB)
img.save("cat.png") # 128px PIL image
generate(label, arm, steps, cfg, seed, n) — label "Cat"/"Dog", arm "hsl"/"learned"/"none",
respaced DDIM (16 steps ≈ 1–3 min on CPU, seconds on any GPU) with CFG + dynamic thresholding,
sampling from the EMA weights. Lower-level pieces (load_holofuse, ddim_sample, UNet,
HSLLabelCond) are in the same model.py. A CLI is included as well:
python generate.py --label Dog --steps 24 --cfg 1.6 --seed 7 --out dog.png
Full-quality ancestral sampling (T=250) and the training harness: Woojiggun/HoLo-FuSe.
Data & license
Trained on AFHQ (StarGAN v2, Choi et al. 2020) animal faces
at 128px (Cat 5153 / Dog 4739, via zzsi/afhq512_16k). AFHQ is CC BY-NC 4.0, therefore these
weights and their outputs are CC BY-NC 4.0 — non-commercial, research/demo only.
Training: 16k steps/arm on a single free Colab T4, crash-resumable harness.