GEAR: Guided End-to-End AutoRegression for Image Synthesis
Abstract
GEAR trains a vector-quantized tokenizer and autoregressive generator jointly end-to-end using representation alignment, overcoming non-differentiability issues through a dual read-out approach that improves convergence speed and feature quality.
Visual generative models are typically trained in two stages. A tokenizer is first trained for reconstruction and then frozen, after which a generator is trained on its discrete indices or continuous latents. This decoupling leaves the tokenizer unaware of what the generator finds easy to model. We present GEAR (Guided End-to-end AutoRegression), which trains a vector-quantized (VQ) tokenizer and an autoregressive (AR) generator jointly and end-to-end, guided by representation alignment. The key obstacle is that the VQ index fed to the AR model is non-differentiable, so gradients cannot reach the tokenizer, and a straight-through estimator collapses. GEAR resolves this with a dual read-out of the codebook assignment. A hard, one-hot branch trains the AR with next-token prediction, while a differentiable soft branch carries a representation-alignment loss that flows back to guide only the tokenizer. The AR model thereby steers its tokenizer toward an index distribution it can predict more easily. This shifts the alignment burden from the tokenizer to the AR: the tokenizer's own features become less DINOv2-like while the AR's become more so, the opposite of diffusion-side recipes that make the latent itself semantic. GEAR speeds up ImageNet gFID convergence by up to 10x relative to the strong LlamaGen-REPA baseline, learns markedly better patch-level and spatially-coherent features, and generalizes across quantizers (VQVAE, LFQ, IBQ) and to text-to-image generation.
Community
📄 Paper: https://arxiv.org/abs/2606.32039
💻 Code: https://github.com/Tencent-Hunyuan/GEAR
🤗 Models: https://huggingface.co/collections/BinLin203
🏠 Homepage: https://linb203.github.io/gear
[1/6] 🚨 Stop freezing your tokenizers! Visual Autoregressive (AR) models are stuck in the 2-stage era. We present GEAR ⚙️ (Guided End-to-End AutoRegression). By jointly training the VQ tokenizer & AR generator, we achieve up to 10x faster convergence than strong baselines like LlamaGen-REPA! 🧵👇
[2/6] 🧠 The Bottleneck & The Solution Why is end-to-end discrete AR so hard? The non-differentiable argmax! Naive Straight-Through Estimator (STE) causes the codebook to collapse. 💥 GEAR solves this with a brilliant Dual Read-out mechanism: 1️⃣ Hard branch: Trains AR with next-token prediction. 2️⃣ Soft branch: A differentiable path carrying alignment loss back to guide only the tokenizer.
[3/6] 🤯 The Mind-Blowing Finding (Representation Shift) Here is the most surprising part: Unlike diffusion models (where end-to-end training makes latents more semantic), GEAR does the exact OPPOSITE! The tokenizer becomes less DINOv2-like, reorganizing for pure predictability (lower entropy). The semantic alignment burden shifts entirely to the AR model's hidden states!
[4/6] 🚀 The Results: Faster & Better The numbers speak for themselves. On ImageNet 256x256, GEAR consistently beats baselines across B/L/XL scales (gFID drops to 2.52). On Text-to-Image (GPIC), it reaches the baseline's final REPA alignment loss 11.1x faster and NTP loss 2.5x faster! ⚡️
[5/6] 🔧 Generality Across Quantizers GEAR isn't just a one-trick pony for VQ-VAE. The soft-guidance mechanism is highly general! It works seamlessly across different quantizers: VQVAE, LFQ, and IBQ. In every single case, GEAR improves BOTH generation quality and reconstruction fidelity. 📈
[6/6] 🔮 The Future of Visual AR Despite the reconstruction ceiling of discrete tokens, end-to-end VQ-AR is the key to unified, long-context generation. GEAR paves the way for applying LLM-style alignment (RLHF, DPO) directly to visual tokens! Dive into the paper and code below: 👇
📄 Paper: https://arxiv.org/abs/2606.32039
💻 Code: https://github.com/Tencent-Hunyuan/GEAR
🤗 Models: https://huggingface.co/collections/BinLin203
Visual generative models typically suffer from decoupled two-stage training. GEAR solves this by enabling end-to-end joint training of a tokenizer and an AR generator. It overcomes the non-differentiability of discrete tokens via a dual read-out mechanism: a hard branch for AR prediction and a soft branch to pass gradients back to the tokenizer. This shifts the semantic alignment burden to the AR model, guiding the tokenizer to produce easily predictable indices. Consequently, GEAR accelerates convergence by up to 10x, improves spatial feature coherence, and demonstrates strong generalization across different quantizers (VQVAE, LFQ, IBQ) and text-to-image (T2I) generation.
Get this paper in your agent:
hf papers read 2606.32039 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 6
Browse 6 models citing this paperDatasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper