REVIEW 7 cited by
Scaling the Codebook Size of VQGAN to 100,000 with a Utilization Rate of 99%
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
In the realm of image quantization exemplified by VQGAN, the process encodes images into discrete tokens drawn from a codebook with a predefined size. Recent advancements, particularly with LLAMA 3, reveal that enlarging the codebook significantly enhances model performance. However, VQGAN and its derivatives, such as VQGAN-FC (Factorized Codes) and VQGAN-EMA, continue to grapple with challenges related to expanding the codebook size and enhancing codebook utilization. For instance, VQGAN-FC is restricted to learning a codebook with a maximum size of 16,384, maintaining a typically low utilization rate of less than 12% on ImageNet. In this work, we propose a novel image quantization model named VQGAN-LC (Large Codebook), which extends the codebook size to 100,000, achieving an utilization rate exceeding 99%. Unlike previous methods that optimize each codebook entry, our approach begins with a codebook initialized with 100,000 features extracted by a pre-trained vision encoder. Optimization then focuses on training a projector that aligns the entire codebook with the feature distributions of the encoder in VQGAN-LC. We demonstrate the superior performance of our model over its counterparts across a variety of tasks, including image reconstruction, image classification, auto-regressive image generation using GPT, and image creation with diffusion- and flow-based generative models. Code and models are available at https://github.com/zh460045050/VQGAN-LC.
Forward citations
Cited by 7 Pith papers
-
Hita: Holistic Tokenizer for Autoregressive Image Generation
Hita's holistic-to-local tokenization lets vanilla autoregressive image models generate global tokens first, improving FID, convergence, and enabling zero-shot style transfer and inpainting.
-
Scalable Training for Vector-Quantized Networks with 100% Codebook Utilization
A training-only ViT-based projector, VQBridge, combined with learning annealing, achieves full codebook utilization in vector-quantized networks at large codebook sizes, improving reconstruction and autoregressive ima...
-
Quantize-then-Rectify: Efficient VQ-VAE Training
A pretrained VAE can be converted into a high-compression VQ-VAE by freezing it and training only a multi-group quantizer plus a post rectifier, cutting training cost by over two orders of magnitude while keeping rFID...
-
Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis
DisCon treats discrete image tokens as conditioning signals rather than targets, letting a continuous autoregressive model refine details and reach gFID 1.38 on ImageNet-256.
-
UniCode$^2$: Cascaded Large-scale Codebooks for Unified Multimodal Understanding and Generation
UniCode² builds a 500K-entry codebook from clustered SigLIP embeddings and uses a cascaded frozen-plus-trainable codebook to unify multimodal understanding and generation with stable training and high token utilization.
-
MagiCodec: Simple Masked Gaussian-Injected Codec for High-Fidelity Reconstruction and Generation
A single-layer streaming Transformer codec with masked Gaussian noise injection during training reports state-of-the-art reconstruction and better downstream generation and understanding in 16 kHz English speech.
-
UniTTS: An end-to-end TTS system without decoupling of acoustic and semantic information
The authors propose DistilCodec, a 32,768-code single-codebook audio codec, and UniTTS, a Qwen2.5-7B TTS model trained with audio, text, and cross-modal autoregressive tasks on interleaved prompts.
Discussion (0). Sign in to comment.