This study adapts the Super-Resolution via Repeated Refinement (SR3) conditional diffusion model for ECG noise removal. We aimed to develop a single model capable of eliminating electrode motion (EM), muscle artifacts (MA), and composite noise, which are particularly challenging to remove at low signal-to-noise ratios (SNRs) without data-driven methods. We employed Gram Angular Fields for lossless embedding of 1D signals into the image domain. The model’s performance was compared against various filters, including an LMS adaptive filter, FIR low pass filter (LPF), moving average, and a hybrid filter (LPF with LMS), depending on the noise type. Initially, the model was trained on MIT-BIH Arrhythmia Database (ARDB) data with added noise from the MIT-BIH Noise Stress Test Database. We further assessed the model’s performance on the MIT-BIH Atrial Fibrillation Database (AF) to evaluate its generalization ability. Our model effectively tackled EM, MA, and composite noise across SNRs ranging from 0 to 15 decibels (dB). It outperformed baseline models in handling single noise types (EM, MA) at low SNRs (0 and 5 dB). While excelling in removing composite noise on ARDB across all SNRs, its effectiveness diminished for single noise types at high SNRs. Retraining our model on atrial fibrillation data significantly improved its ability to denoise these types of signals. Our methodology presents an innovative yet computationally demanding approach to ECG noise removal, excelling in low SNR and multi-noise settings.