ViTok-v2: Scaling Native Resolution Autoencoders to 5 Billion Parameters

Vision Transformer (ViT) tokenizers offer a scal- able alternative to convolutional auto-encoders, yet current architectures have two key limitations: their performance degrades when images vary in aspect ratio or resolution, and their reliance on adversarial losses makes them harder to train at scale. To address this, we introduce ViTok-v2, a ViT tokenizer building on ViTok. We add native resolution support via NaFlex with 2D RoPE and stabilize training by replacing the standard LPIPS-plus-discriminator objective with our novel DINO perceptual loss. We scale our model to 5B parameters, training the largest ViT-based image compression auto- encoder to date and demonstrate continued im- provements with scale. In downstream generation experiments with flow matching models, we find that smaller generators perform best with aggres- sive channel compression while larger generators effectively leverage higher channel counts. ViTok- v2 matches state-of-the-art reconstruction at 256p and outperforms across benchmarks at 512p and higher resolutons, while remaining compatible with any pipeline requiring flexible aspect ratios.

Many AI systems that create or understand images first compress each image into a smaller internal representation. Existing image compression methods based on Vision Transformers can work well, but they often struggle when images come in different shapes or resolutions, and some require unstable training procedures that become harder to use at large scale. We introduce ViTok-v2, an image tokenizer designed to handle images of many sizes and aspect ratios while being easier to train. Instead of relying on adversarial training, we use a new perceptual training signal based on DINO, a strong image-understanding model, to help the tokenizer preserve the visual content that matters. We scale this approach up to a 5-billion-parameter model and find that larger tokenizers continue to improve image reconstruction. We also study how these compressed image representations affect downstream image generation, showing that small generators benefit from stronger compression while larger generators can use richer representations. Overall, ViTok-v2 improves high-resolution image reconstruction and makes image tokenization more flexible for modern image generation pipelines.