Learning map
Prerequisite links describe a reading dependency, not historical descent or empirical superiority. Each article separately explains what it inherits and what it compares.
1. Compression and discrete generation
Section titled “1. Compression and discrete generation”- A small language for generative models
- Autoencoders: a useful bottleneck
- VAE: learn to reconstruct and to sample
- VQ-VAE: learn a discrete dictionary
- VQ-VAE-2: separate global structure and local detail
- Attention, masks, and next-token prediction
- Three different meanings of “better images”
- VQGAN: a tokenizer that preserves perceptual detail
- VAR: predict the next scale, not the next patch
2. Transformer reconstruction
Section titled “2. Transformer reconstruction”- ViT: an image becomes a sequence
- RoPE-ViT: position enters the comparison
- TiTok: a short sequence can describe an image
3. Diffusion and flow
Section titled “3. Diffusion and flow”- Diffusion and flow: learn a path from noise
- DiT: a Transformer inside latent diffusion
- DiTo: learn the tokenizer through denoising
- FlowMo: a Transformer reconstructs through flow
4. Shared semantic representations
Section titled “4. Shared semantic representations”- Semantic features: useful for which task?
- RAE: decode a frozen representation
- MAETok: shape the space before decoding it
- UniTok: split a code, share the representation
- VILA-U: one visual tower for reading and writing
- AToken: connect semantic, continuous, and discrete views
5. Continuous autoregression
Section titled “5. Continuous autoregression”- AIM: predict patches to learn features
- MAR: autoregression without a codebook
- Fluid: compare value type and generation order
- CausalFusion: two axes of generation
- NextStep-1: continuous images in a causal language model