Playground experiment
Add noise to a generated image, then step backward through a simplified denoising process. It is a canvas intuition for generative diffusion models, not a trained model.
Noise schedule
The timestep slider controls how much noise has been added. Play the reverse process to watch the noisy image move back toward a structured sample.
Clean target
Noisy timestep
Denoised estimate
Model-backed sampler
After the notebook exports the tiny ONNX denoiser, this panel samples from pure noise in the browser with DDIM-style steps.
Current noise
Predicted clean
Final sample
The forward process gradually mixes a clean sample with random noise. At high timesteps the image keeps less visible structure, which is why real diffusion models learn many small denoising steps.
The reverse panel is a deterministic canvas approximation. It uses the hidden clean sample as a guide so the interaction stays understandable; a real model would learn that guide from data.