EchoDiff: Echo State Reservoirs for Parameter‑Efficient Diffusion Image Generation
Abstract
Diffusion models achieve state-of-the-art image quality but rely on large denoising networks that require repeated, expensive forward passes. This paper introduces EchoDiff, a Diffusion model that integrates Echo State Networks (ESNs) into the denoising pipeline to reduce trainable parameters and memory cost. EchoDiff replaces conventional U-Net downsampling and upsampling blocks with Patch-based ESN modules that decompose feature maps into local patch tokens, process them sequentially through a frozen random reservoir with leaky integration, and map reservoir states to outputs via lightweight, conditioning-aware linear readouts. Because the reservoir weights are fixed and never updated, gradients flow only through small input and readout layers, eliminating backpropagation through recurrent dynamics. We further reduce memory by sharing reservoirs between symmetric encoder–decoder stages and using depthwise-separable convolutions to compress token dimensionality. On MNIST and Fashion-MNIST, EchoDiff cuts trainable parameters by 49.7% (7.19M → 3.62M) and peak GPU memory by up to 33% compared to a U-Net-based DDPM, while achieving lower class-wise FID, lower CFID, and higher recall. An ablation study shows that nonlinear reservoir dynamics, sequential inter-patch context, and per-token conditioning each contribute to these gains over a linear projection baseline. These results position reservoir computing as a promising foundation for lightweight diffusion-based generative…
Read the paper
Related studies
- Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2)
- Commentary on the clinical and preclinical dosage limits of interstitially administered magnetic fluids for therapeutic hyperthermia based on current practice and efficacy models
- Prosthesis: New Technological Opportunities and Innovative Biomedical Devices
- The history and future of neural modeling for cochlear implants
- Ryanodine receptors: waking up from refractoriness
- Novel Transgenic Rabbit Model Sheds Light on the Puzzling Role of Matrix Metalloproteinase-12 in Atherosclerosis
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.