Patch Based Image Tokenization With Learned Spatial Embeddings

via “multimodal observation tokenization with flexible sensor composition”

Generalist robot policy model from Open X-Embodiment.

Unique: Implements a modular tokenizer architecture where image tokenizers (learned codebooks or pretrained vision models) and proprioception tokenizers (linear/MLP projections) are independently trained and composed, allowing flexible sensor configuration without retraining the transformer backbone. Supports variable numbers of cameras through dynamic token concatenation.

vs others: More flexible than end-to-end vision models that require fixed camera configurations, and more efficient than raw pixel processing by reducing observation dimensionality 100-1000x while preserving task-relevant information through learned tokenization.

via “feature extraction and embedding generation for downstream tasks”

image-classification model by undefined. 47,71,224 downloads.

Unique: Provides access to hierarchical transformer hidden states (12 layers × 768 dimensions) enabling multi-scale feature extraction; [CLS] token embeddings capture global image semantics superior to average pooling used in CNN-based models, improving downstream task performance

vs others: ViT embeddings achieve better downstream task performance (e.g., 5-10% higher accuracy on image retrieval) compared to ResNet-50 embeddings due to transformer's global attention capturing long-range visual dependencies; embeddings are more semantically aligned with human perception

via “patch-based image tokenization with learned positional embeddings”

image-classification model by undefined. 6,53,291 downloads.

Unique: Uses learned positional embeddings (768-dimensional vectors per patch position) rather than sinusoidal positional encodings, allowing the model to learn task-specific spatial relationships. Combines a learnable [CLS] token (similar to BERT) with patch embeddings, enabling the model to aggregate global image information through a single token rather than pooling all patches.

vs others: More parameter-efficient than CNN feature pyramids (single 768-dim embedding per patch vs multi-scale feature maps), and provides better long-range spatial reasoning than local convolution kernels because each patch attends to all other patches globally.

via “patch-based image tokenization with positional encoding”

image-to-text model by undefined. 1,67,827 downloads.

Unique: Implements 2D positional encoding that explicitly encodes patch grid coordinates (row, column) rather than using 1D sequential positional embeddings, preserving the 2D spatial structure of images. This allows the transformer to learn spatial relationships between patches more effectively than treating them as a flat sequence.

vs others: More spatially-aware than standard ViT positional encoding because it uses 2D coordinates, but less flexible than adaptive tokenization schemes (e.g., DINOv2) that allocate tokens based on image complexity.

via “vqgan detokenization for pixel-space image reconstruction”

min(DALL·E) is a fast, minimal port of DALL·E Mini to PyTorch

Unique: Uses pre-trained VQGan decoder (not a custom decoder), ensuring compatibility with tokens generated by the DALL·E Bart decoder which was trained on VQGan-tokenized images. Supports progressive detokenization via iterator pattern, enabling real-time image rendering without waiting for full token sequence.

vs others: More efficient than diffusion-based decoding (1-2s vs 30-60s) because it's a single forward pass; maintains higher fidelity than upsampling-based approaches because it uses learned reconstruction rather than interpolation.

via “tokenization and text preprocessing for embeddings”

Portable WASM embedding generation with SIMD and parallel workers - run text embeddings in browsers, Cloudflare Workers, Deno, and Node.js

Unique: Implements streaming tokenization for long documents, processing text in chunks and maintaining state across chunk boundaries to handle word-boundary edge cases. Supports custom tokenization rules via pluggable tokenizer interface, allowing domain-specific vocabulary (e.g., code tokens, medical terminology).

vs others: More efficient than calling external tokenization APIs (e.g., Hugging Face Inference API) since tokenization runs locally with zero network latency, and more flexible than hardcoded tokenization since vocabulary is configurable per model.

via “discrete image tokenization for unified sequence representation”

* ⏫ 07/2023: [Meta-Transformer: A Unified Framework for Multimodal Learning (Meta-Transformer)](https://arxiv.org/abs/2307.10802)

Unique: Uses discrete image tokenization to enable unified autoregressive processing of images and text in a single decoder, treating image generation as sequence prediction rather than pixel-space generation

vs others: Simpler than continuous image representations because it reuses text token infrastructure; enables unified architecture but trades off visual fidelity compared to continuous or diffusion-based approaches

via “patch-based image tokenization with learned spatial embeddings”

* ⭐ 02/2023: [Adding Conditional Control to Text-to-Image Diffusion Models (ControlNet)](https://arxiv.org/abs/2302.05543)

Unique: Uses learned 2D positional embeddings that explicitly encode both row and column position information, enabling the model to reason about spatial relationships. Unlike 1D positional encodings used in NLP, this 2D approach preserves the grid structure of images and allows attention heads to develop position-aware patterns.

vs others: More parameter-efficient than CNN feature extraction for large models (saves 50M+ parameters vs ResNet-50 backbone) and enables pure attention-based processing, but requires 2-3x more training data than CNN-based approaches to match accuracy on ImageNet-scale datasets.

via “patch embedding with overlapping windows for feature extraction”

* ⭐ 04/2022: [Hierarchical Text-Conditional Image Generation with CLIP Latents (DALL-E 2)](https://arxiv.org/abs/2204.06125)

Unique: Uses overlapping patch embeddings with learned projections to preserve spatial continuity and reduce boundary artifacts, contrasting with standard non-overlapping patch tiling used in ViT and providing smoother feature transitions

vs others: Produces higher-quality feature representations than non-overlapping patches with better boundary preservation, though at higher computational cost; enables better performance on dense prediction tasks

via “discrete visual tokenization with learned codebook”

* ⭐ 09/2022: [PaLI: A Jointly-Scaled Multilingual Language-Image Model (PaLI)](https://arxiv.org/abs/2209.06794)

Unique: Uses learned discrete codebooks to tokenize images, creating a bridge between continuous vision features and discrete language tokens. This enables applying BERT-style masked language modeling directly to images without pixel-level reconstruction.

vs others: Provides better semantic alignment with language models than continuous feature representations because discrete tokens create a shared vocabulary between modalities, improving joint vision-language learning compared to approaches using separate continuous representations.

via “patch-based image tokenization for transformer input”

### NLP <a name="2022nlp"></a>

Unique: Applies standard vision transformer patch tokenization to diffusion models, enabling direct reuse of transformer optimization techniques (flash attention, tensor parallelism) developed for NLP; patch size becomes a key hyperparameter controlling the speed-quality tradeoff

vs others: Simpler and more efficient than pixel-level processing or hierarchical patch schemes; enables better hardware utilization compared to CNN-based U-Nets which require custom CUDA kernels for efficient convolution