Architecture timeline · image generation

How we got to FLUX.2.

Twelve years of ideas that taught machines to paint, from the variational autoencoder to today’s frontier image model. Every stop links to a plain-English explanation or a live interactive graph.

11 stops 6 papers 6 model graphs 2013 to 2025
Concept Model Paper and model in one release
Learning to generate2013 to 2020

Two ideas make images something a network can produce, not just recognize.

ConceptDec 2013VAE (latent autoencoder)Compress images into a smooth latent space and decode them back. Every diffusion model below paints inside a VAE’s latent, not in pixels.latent Auto-Encoding Variational BayesarXiv 1312.6114 · Dec 2013Read the glossary entry →
ConceptJun 2020Denoising diffusion (DDPM)Generation becomes gradual denoising: start from pure noise and remove a little at each step. A timestep embedding tells one network which stage it is at.28 steps Denoising Diffusion Probabilistic ModelsarXiv 2006.11239 · Jun 2020Timestep embeddingRead the glossary entry →
Words steer pictures2020 to 2022

Transformers learn to see, text and images share a space, and the recipe goes open source.

ConceptOct 2020Patch embedding (ViT)An image becomes a sequence: slice it into patches and project each one into a token. The trick that lets transformers see. An Image is Worth 16x16 Words: Transformers for Image Recognition at ScalearXiv 2010.11929 · Oct 2020Vision encoderRead the glossary entry →
Paper + modelJan 2021openai/clip-vit-base-patch32The paper and the open weights land together. Images and text share one embedding space, giving generators a way to know how well a picture matches a prompt.openai/clip-vit-base-patch32 architecture graphzero-shot-image-classification · ↓ 23.9M · ♡ 975Learning Transferable Visual Models From Natural Language SupervisionarXiv 2103.00020 · Mar 2021Open in visualizer →
ConceptDec 2021Cross-attention steers the pictureA 2014 translation idea finds its calling: latent diffusion wires the prompt’s tokens into every denoising step through cross-attention.other sequence Neural Machine Translation by Jointly Learning to Align and TranslatearXiv 1409.0473 · Sep 2014Read the glossary entry →
ModelAug 2022stable-diffusion-v1-5/stable-diffusion-v1-5Latent diffusion with CLIP conditioning, weights open to everyone. Image generation stops being a demo and becomes an ecosystem.stable-diffusion-v1-5/stable-diffusion-v1-5 architecture graphtext-to-image · ↓ 1.7M · ♡ 1kDenoising loopOpen in visualizer →
The transformer takes over2022 to 2023

The U-Net gives way to patches and blocks, and scaling laws arrive for images.

ConceptDec 2022DiT blockReplace the U-Net with a transformer over latent patches, modulated by the timestep. Scaling laws arrive for image generation.attentionMLPt Scalable Diffusion Models with TransformersarXiv 2212.09748 · Dec 2022Read the glossary entry →
ModelJul 2023stabilityai/stable-diffusion-xl-base-1.0The open recipe grows up: a much bigger backbone and two text encoders. Still the workhorse behind a thousand fine-tunes.stabilityai/stable-diffusion-xl-base-1.0 architecture graphtext-to-image · ↓ 1.4M · ♡ 8kCross-attentionOpen in visualizer →
Flow to the frontier2024 to 2025

Rectified flow plus a big DiT becomes the frontier recipe for pictures.

ModelAug 2024black-forest-labs/FLUX.1-devA 12B rectified-flow DiT from the original Stable Diffusion authors. One of the most liked models in Hugging Face history.black-forest-labs/FLUX.1-dev architecture graphtext-to-image · ↓ 625k · ♡ 14kDiT blockOpen in visualizer →
ModelAug 2025Qwen/Qwen-ImageThe DiT recipe spreads: a 20B image model that lays out crisp text inside pictures, in English and Chinese.Qwen/Qwen-Image architecture graphtext-to-image · ↓ 190k · ♡ 3kOpen in visualizer →
Nov 2025 · where it all lands

black-forest-labs/FLUX.2-dev

Every idea above, alive in one graph: a VAE latent, a denoising loop, DiT blocks with the LLM stack’s own RMSNorm inside. Open it and trace a prompt’s path to pixels.

black-forest-labs/FLUX.2-dev architecture graph
VAEDenoising loopDiT blockTransformer blockRMSNorm
Open FLUX.2 in the visualizer →

Want the definitions behind each stop? The glossary explains all 39 concepts, and the catalog has interactive graphs for 2,300+ models.

AlsoHow we got to DeepSeek-V4-Pro · the LLM timeline →AlsoHow we got to Qwen3-VL · the vision-language timeline →