HoLo-FuSe uses 0-param HSL substrate for class-conditional image generation

HoLo-FuSe demonstrates that a frozen, zero-parameter HoLo Semantic Layer (HSL) byte substrate can effectively condition a diffusion model for image generation. The project tests whether this deterministic 27-D feature frame can serve as the conditioning mechanism for a class-conditional DDPM, replacing traditional learned embeddings.

• Trained on a single Colab T4 with ~35M U-Net parameters over 16k steps at 128px resolution.
• Uses AFHQ animal faces dataset (Cat/Dog) with classifier-free guidance and cosine schedule.
• Seed-matched qualitative results show HSL conditioning performs comparably to a budget-matched learned embedding control.
• Background color tint observed in all arms is attributed to under-training rather than the substrate.

The authors consider this a proof of operation, establishing that the frozen substrate can steer class generation as well as learned controls, with future work focused on quality improvements through longer training.