Can we model images like language — token by token, end to end?...
机器之心 JIQIZHIXIN@jiqizhixin
69 views
Nov 05, 2025
~1 min read
Advertisement
1
Can we model images like language — token by token, end to end?
FARMER is a new framework that directly models raw pixel likelihoods by combining an invertible autoregressive flow (AF) with a causal autoregressive (AR) model, achieving exact likelihoods and high-quality synthesis in one stage.
Three key innovations:
1. Self-supervised channel split separates latent features into informative (Zᴵ) and redundant (Zᴿ) groups, simplifies modeling of high-dimensional pixel data.
2. Resampling-based classifier-free guidance (CFG) substantially enhances generation quality.
3. One-step AF distillation accelerates generation by replacing slow sequential reversal process with a one-step forward process.
On ImageNet-256x256, FARMER-1.9B achieve FID 3.60/ IS 269.21, better than JetFormer-2.8B (FID 6.64), TARFlow (5.56), Pixel-based STARFlow (4.69), FractalMAR (6.15).
🌾 FARMER: Flow AutoRegressive Transformer over Pixels
Paper: arxiv.org/pdf/2510.23588
📬 #PapersAccepted by Jiqizhixin
FARMER is a new framework that directly models raw pixel likelihoods by combining an invertible autoregressive flow (AF) with a causal autoregressive (AR) model, achieving exact likelihoods and high-quality synthesis in one stage.
Three key innovations:
1. Self-supervised channel split separates latent features into informative (Zᴵ) and redundant (Zᴿ) groups, simplifies modeling of high-dimensional pixel data.
2. Resampling-based classifier-free guidance (CFG) substantially enhances generation quality.
3. One-step AF distillation accelerates generation by replacing slow sequential reversal process with a one-step forward process.
On ImageNet-256x256, FARMER-1.9B achieve FID 3.60/ IS 269.21, better than JetFormer-2.8B (FID 6.64), TARFlow (5.56), Pixel-based STARFlow (4.69), FractalMAR (6.15).
🌾 FARMER: Flow AutoRegressive Transformer over Pixels
Paper: arxiv.org/pdf/2510.23588
📬 #PapersAccepted by Jiqizhixin