How Diffusion Turns Random Numbers Into an Image | Diffusion Visually Explained

Video Model Series: https://www.youtube.com/playlist?list=PLWow_bYuhBL8

You press Generate. The process begins with a grid of random numbers—not a hidden picture.
How can one model examine that grid many times and turn it into a clear image that follows your words?
This video builds diffusion from that visible puzzle. We create one noisy training example with known numbers, watch many local predictions form a field, and then follow a seeded starting grid as a sampler moves it through that learned field. Broad shapes become recoverable before fine texture. Stable Diffusion performs these repeated changes in a smaller latent grid, while the text request changes every noise prediction.
The central picture is simple:
real image + known noise → a target the model can learn
random grid → model prediction → sampler change → repeat → final image
The image is not stored in the starting noise. It develops through a sequence of choices guided by patterns learned from real images and by information from the text request.
Along the way, we’ll answer:
• How does adding noise create a known training answer?
• What should the model predict when several clean images fit the same noisy input?
• How do local predictions become a field that guides generation?
• What is the difference between the model’s prediction and the sampler’s change?
• Why do broad shapes usually become stable before fine details?
• Why does Stable Diffusion work in a smaller latent grid?
• How does text change every prediction?
• What do the seed, step count, sampler, and guidance value each control?

Chapters:
0:00 When Does the Image Exist?
0:28 Training Creates a Known Target
1:56 Denoising Estimates Define a Score Field
3:19 Sampling Follows the Learned Field
4:41 Large-Scale Structure Becomes Recoverable First
5:50 Latent Diffusion and Text Guidance
8:34 The Complete Mechanism

Sources and further reading:
• Denoising Diffusion Probabilistic Models — Ho, Jain, and Abbeel (2020): https://arxiv.org/abs/2006.11239
• The Annotated Diffusion Model — Rogge and Rasul (2022): https://huggingface.co/blog/annotated-diffusion
• Generative Modeling by Estimating Gradients of the Data Distribution — Yang Song (2021): https://yang-song.net/blog/2021/score/
• Perspectives on Diffusion — Sander Dieleman (2023): https://sander.ai/2023/07/20/perspectives.html
• Hugging Face Diffusers Schedulers: https://huggingface.co/docs/diffusers/using-diffusers/schedulers
• Diffusion Is Spectral Autoregression — Sander Dieleman (2024): https://sander.ai/2024/09/02/spectral-autoregression.html
• The Illustrated Stable Diffusion — Jay Alammar (2022): https://jalammar.github.io/illustrated-stable-diffusion/
• High-Resolution Image Synthesis with Latent Diffusion Models — Rombach et al. (2022): https://arxiv.org/abs/2112.10752
• Classifier-Free Diffusion Guidance — Ho and Salimans (2022): https://arxiv.org/abs/2207.12598
• Hugging Face Diffusers Reproducibility: https://huggingface.co/docs/diffusers/main/using-diffusers/reusing_seeds

#diffusionmodels #stablediffusion #aiimagegeneration #generativeai #denoising #classifierfreeguidance #latentspace #machinelearning #artificialintelligence #mlvisualized

User Queries:
diffusion models explained visually
how diffusion turns noise into images
how stable diffusion works
diffusion model training explained
score field diffusion intuition
diffusion sampler explained
why diffusion images form coarse to fine
latent diffusion explained
classifier free guidance explained
seed steps sampler guidance explained