Diffusion & Generative Media

The story, event by event

Every point below is traced to a real source — nothing on this page is invented.

1. 10 Jun 2014impact 76

   GANs — the first generative breakthrough

   A two-network adversarial game — a generator against a discriminator — became the first method to reliably synthesize convincing images from noise.

   Goodfellow et al. framed generation as a minimax game: a generator tries to fool a discriminator into accepting synthetic samples as real, and the two networks improve together until the generator's output is indistinguishable from real data.

   arXiv

   GANs went on to dominate image synthesis from 2014 to roughly 2021 — style transfer, face generation, and super-resolution all ran on adversarial training before diffusion overtook it on fidelity and training stability.

   arXiv

   GANs proved neural networks could generate convincing novel images, not just classify them — the opening event of the entire generative-media lineage.

2. 19 Jun 2020impact 78

   Diffusion origin → DDPM makes it work

   Diffusion models learn to generate by reversing a gradual noising process — an idea introduced in 2015 that only became competitive with GANs once Ho et al. made it practical in 2020.

   Diffusion probabilistic models were first proposed by Sohl-Dickstein et al. in 2015: destroy data by adding noise step by step, then train a network to reverse each step, turning noise back into data.

   arXiv

   Denoising Diffusion Probabilistic Models (DDPM) simplified the training objective to plain noise prediction and showed the resulting samples matched or beat GANs on image fidelity — the paper that made diffusion the field's default generative method.

   arXiv

   DDPM is the handover event: from here on, diffusion — not adversarial training — is the default way to build a generative image model.

3. 6 Apr 2022impact 74

   DALL·E 2 — frontier text-to-image

   DALL·E 2 showed a single model could turn an open-ended text prompt into a photorealistic, editable image — the demo that made text-to-image a mainstream category.

   Built on unCLIP, DALL·E 2 generates a CLIP image embedding from the text prompt and then diffuses an image conditioned on that embedding, producing higher-resolution and more caption-faithful images than its GAN-era predecessors.

   arXiv

   Beyond raw generation, DALL·E 2 supported inpainting and variation from a reference image, moving text-to-image from a research curiosity to a controllable creative tool and defining the commercial product category that followed.

   OpenAI

   DALL·E 2 is the frontier demonstration that put text-to-image in front of the public and set the bar every open competitor was measured against.

4. 22 Aug 2022impact 84

   Latent / Stable Diffusion — democratized

   Running the diffusion process in a compressed latent space instead of pixel space cut the compute cost enough to put high-quality image generation on a consumer GPU.

   Rombach et al. showed that training and sampling diffusion models in the latent space of a pretrained autoencoder, rather than directly on pixels, preserved image quality while sharply reducing the compute needed for both training and inference.

   arXiv

   Stability AI, CompVis, and Runway released Stable Diffusion publicly in August 2022 with open weights — the first frontier-quality text-to-image model anyone could run and fine-tune on their own hardware, which triggered the open-source generative-image ecosystem.

   Stability AI

   This is the democratization event: open weights turned high-end image generation from a closed API into infrastructure anyone could build on.

5. 15 Feb 2024impact 70

   Sora — text-to-video (a Diffusion Transformer)

   Sora generated up to a minute of coherent video from a text prompt by running diffusion on a transformer backbone instead of the usual U-Net, moving the generative frontier from images to temporal media.

   Sora is a diffusion transformer: it operates on space-time patches of a video's latent representation, replacing the convolutional U-Net that earlier diffusion models relied on with the same attention-based architecture behind large language models.

   OpenAI

   Maintaining subject and scene consistency across up to a minute of generated video — not just a single frame — was the qualitative jump that made synthetic video look plausible, and it sharply raised the stakes on provenance and detection of AI-generated media.

   arXiv

   Sora shifted the generative frontier from static images to coherent video, and fused the diffusion and transformer lineages into one architecture in the process.

6. 13 Jun 2025impact 60

   Integration of AI in Filmmaking

   film: ANCESTRA; partnership: Google DeepMind and Primordial Soup; technologies: Gemini, Veo, Imagen

7. 26 Feb 2026impact 60

   Introduction of Nano Banana 2 Model

   features: advanced image generation capabilities, high speed, accessibility, precise text rendering, improved visual fidelity; model name: Nano Banana 2; integration: available across various Google products

8. 3 Mar 2026impact 50

   24-hour speedrun for text-to-image diffusion model training

   budget: constrained; techniques used: x-prediction formulation, Muon optimizer; training duration: 24 hours

9. 6 Jul 2026impact 60

   Data Strategy for PRX Model Training

   key argument: The article argues for a data strategy that prioritizes diversity and breadth in pre-training datasets to enhance model learning.; long captions: We found that long, detailed captions are a very strong lever on output quality.

Lineage