Watch on YouTubeWatch on Instagram
About this video
Most AI pictures start as pure static. Here's how they become photos.
How diffusion works, why you never get the same picture twice, and why ChatGPT's images are made differently.
Episode 6 of Learn AI in Under 3 Minutes, the last of level 1.
Made and measured for this video on FLUX.2 [klein] 4B (Black Forest Labs, open weights, Apache 2.0, January 2026), a 4-billion-parameter rectified flow transformer distilled to 4 steps, run on a 2020 MacBook Air (M1, 16 GB). Prompt: "A red hot-air balloon floating over the misty valleys of Cappadocia at sunrise, photograph", seed 7, 1024 x 576. Every picture on screen is the model's own output: the noisy state at each step and the model's guess of the finished picture (x_t minus noise level times the predicted velocity), decoded. Step 1's guess against the final picture: layout 92% (correlation of the pictures shrunk 32 times), fine detail 32% (correlation of what a blur removes). With 90% noise left (step 3), layout 99.5%. Same seed with "red" changed to "blue": layout 91% the same; the same words from another seed: 56%. The model works on a latent of 294,912 numbers against 1,769,472 for the pixels (6 times fewer); the "compressed sketch" shows 3 of its 32 channels as colours. The four video frames are an illustration made from the same steps, not a video model's output. One picture and one model are an example, not a benchmark.
History: Ho, Jain and Abbeel, "Denoising Diffusion Probabilistic Models" (2020): 1,000 steps. Google DeepMind, Veo 3 model card (January 2026): "Veo 3 utilizes latent diffusion", applied to the audio and the video latents. OpenAI, GPT-4o System Card addendum on native image generation (March 25, 2025): "Unlike DALL·E, which operates as a diffusion model, 4o image generation is an autoregressive model". Inception, Mercury 2.5 (September 2026), a diffusion language model.
0:00 Static to picture
0:18 How it learns
1:07 Prompt and seed
1:36 Faster, and video
1:57 Not all diffusion
2:20 My take
Music: Voice Over Under by Kevin MacLeod, CC BY 3.0, via Wikimedia Commons (https://commons.wikimedia.org/wiki/File:Voice_Over_Under_(ISRC_USUAN1600001).mp3)
Follow @PranavInnovates for the tech story that matters, every day.
What the video says
Most AI pictures you've seen started like this: pure static. Four steps later, this. No camera, no artist. An open model on my laptop cleaned the noise away. That's diffusion. Here's how, why you never get the same picture twice, and why ChatGPT doesn't do it this way. Look closely: there's a balloon in there. Like a song on a crackly radio, you hear it because you know songs. This model knows pictures. It learned backwards. Take a real photo, add noise, more noise, until it's static. Its job: guess the clean picture. Now run it forwards. Start from fresh static. The model looks, and guesses the whole picture at once. The first guess is a blur. It mixes a little of that guess into the static, looks again, and guesses sharper. Quick one. After step one of four, what's already settled: the layout, or the details? The layout. I measured it. The first blur matches the final layout 92 percent. The fine detail? Only 32. It plans the picture before it paints it. Your prompt steers every guess. Same static, one word changed, red to blue: same scene, repainted. Keep the words, change the static, and it's a new picture. Four different cats on episode three's mat. That starting static is the seed. Builder tip: like an image? Keep its seed, edit the words. Same layout, your change. And it never touches pixels. It cleans a compressed sketch, six times smaller, then a decoder paints the pixels. It got fast, too. The 2020 paper that made diffusion work took a thousand steps. This model, from Germany's Black Forest Labs: one step gives a blur, two get close, four, done. Video is the same trick, bigger. Google's model card says Veo uses latent diffusion: the frames and the sound, all cleaned up from noise. So, was ChatGPT's image maker built this way too? No. When OpenAI put images into ChatGPT, it said this one is autoregressive: it builds the picture piece by piece, like the text in episode one. And text can go the other way: Mercury 2.5 writes text by diffusion. My take: steps are the price. Four-step models made pictures cheap enough for a laptop. So don't polish one perfect prompt. Make twenty, pick the best. Static, guess, mix, repeat. That's diffusion, and that's level one done. Next, level two: from prompting to context engineering.