Compare methods

Compare structure rather than ranking incompatible benchmark scores. APE is absolute position encoding. RoPE rotates queries and keys. Not audited does not mean absent.

Method	Representation	Reconstruction or output	Position scheme
VQGAN	grid VQ	single-pass CNN	learned APE (prior)
VAR	multi-scale VQ	single-pass CNN	learned APE + level
ViT	continuous patches	classification head	learned APE
RoPE-ViT	continuous patches	classification head	axial / mixed RoPE
TiTok	1D discrete slots	single-pass hybrid	learned patch / slot APE
DiT	continuous VAE grid	latent denoiser + VAE	fixed 2D sin-cos
DiTo	continuous grid	iterative U-Net	not audited
FlowMo	LFQ slots	iterative flow Transformer	multi-axis RoPE
RAE	high-dimensional features	single-pass ViT	fixed 2D sin-cos (decoder)
MAETok	continuous 1D slots	single-pass decoder	not audited
UniTok	multi-codebook grid	single-pass decoder	hybrid spatial structure
VILA-U	residual-quantized codes	image decoder	not audited
AToken	semantic / continuous / discrete	Transformer decoder	learned APE + 4D RoPE
AIM	continuous pixels	patch regression head	sinusoidal APE
MAR	continuous latent tokens	per-token diffusion MLP	learned APE
Fluid	continuous / discrete	per-token output head	learned positions discussed
CausalFusion	continuous token groups	Transformer denoising	experiment-dependent APE
NextStep-1	text IDs + continuous images	per-patch flow MLP	1D RoPE (paper)
EOSTok	1D IBQ codes	single-pass decoder	learned APE
AdaTok	variable-length VQ	budget-conditioned decoder	not audited
RAE-T2I	shared semantic features	DiT + RAE decoder	module-dependent