Falcon Perception and Falcon OCR Release
Falcon Perception and Falcon OCR Release
TL;DR
Hugging Face and the Technology Innovation Institute (TII) have released Falcon Perception, a 0.6B-parameter early-fusion Transformer designed for open-vocabulary grounding and segmentation, and Falcon OCR, a 0.3B-parameter model for high-throughput document understanding. By replacing modular pipelines with a single-stack Transformer architecture, these models achieve state-of-the-art performance in compositional grounding and OCR throughput while remaining computationally efficient.
Falcon Perception: Architecture and Design
Falcon Perception utilizes a single early-fusion Transformer backbone that processes image patches and text tokens in a shared parameter space from the first layer, eliminating the need for separate vision backbones and late-fusion decoders.
Hybrid Attention and Sequence Processing
To handle the differing structures of visual and textual data, the model employs a hybrid attention mask:
- Image tokens use bidirectional attention to build global visual context.
- Text and task tokens use causal attention, attending to the visual prefix and all preceding text.
Chain-of-Perception Interface
To manage variable-length dense predictions (from zero to hundreds of instances), the model uses a structured autoregressive interface called Chain-of-Perception. Each instance is decomposed into three sequential steps:
<coord>: Predicts the center of the instance.<size>: Predicts the spatial extent.<seg>: Produces a single embedding that is dot-producted with upsampled image features to generate a full-resolution binary mask.
Specialized Output Heads
- Coordinate & Size Heads: Use Fourier feature encoding (random Gaussian projection into high-dimensional sinusoidal space) to overcome spectral bias and improve localization precision. Decoded coordinates are re-injected into the sequence to condition subsequent tokens.
- Segmentation Head: Performs a dot product between the
<seg>token's hidden state and content-aware upsampled image features, removing the need for Hungarian matching or separate mask-query machinery.
PBench: A New Diagnostic Benchmark
TII introduced PBench, a benchmark designed to isolate specific perception failures by categorizing samples by the dominant capability required:
| Level | Capability | Example Prompt |
|---|---|---|
| L0 | Simple objects | "car" |
| L1 | Attributes & subtypes | "red car", "broken fence" |
| L2 | OCR-guided identification | "Diet Coke bottle", "Nike shoes" |
| L3 | Spatial understanding | "car on the left", "third window from left" |
| L4 | Relations & interactions | "person holding umbrella", "tallest building" |
| Dense | Crowdedness stress test | Hundreds of instances per image |
Training Methodology
Multi-Teacher Distillation
Falcon Perception is initialized via distillation from two vision teachers to provide a strong foundation:
- DINOv3 (ViT-H): Provides local features for segmentation.
- SigLIP2: Provides language-aligned features for open-vocabulary understanding.
Data Pipeline and Scale
The model was trained on 54M images, 195M positive expressions, and 488M hard negatives. The pipeline included hierarchical clustering via DINOv3, VLM-driven listing, and an ensemble consensus (SAM 3, Qwen3-VL-30B, and Moondream3) requiring an IoU > 0.8 for automatic acceptance.
Three-Stage Training Recipe
- In-Context Listing (450 GT): Learns to list scene inventories autoregressively, capturing object co-occurrence.
- Task Alignment (225 GT): Modifies the attention mask to simulate independent queries at inference, focusing gradients on presence classification and localization.
- Long-Context Finetuning (10 GT): Adapts the model for extreme crowd density by raising the mask limit to 600 per expression.
Performance Results
SA-Co Benchmark
Falcon Perception (0.6B) achieved a 68.0 Macro-F1 on the SA-Co open-vocabulary segmentation benchmark, surpassing SAM 3's 62.3. It showed significant gains in attribute-heavy (+8.2), food & drink (+12.2), and sports equipment (+4.0) splits, though it lags SAM 3 in presence calibration (MCC 0.64 vs 0.82).
PBench Results
Falcon Perception demonstrates a significant advantage over SAM 3 as prompt complexity increases:
| Capability | SAM 3 | Falcon Perception | Gap |
|---|---|---|---|
| L0: Simple objects | 64.3 | 65.1 | +0.8 |
| L1: Attributes | 54.4 | 63.6 | +9.2 |
| L2: OCR-guided | 24.6 | 38.0 | +13.4 |
| L3: Spatial | 31.6 | 53.5 | +21.9 |
| L4: Relations | 33.3 | 49.1 | +15.8 |
| Dense | 58.4 | 72.6 | +14.2 |
Falcon OCR
Falcon OCR is a 0.3B-parameter variant of the early-fusion architecture trained from scratch to optimize for fine-grained glyph recognition and stroke-level discrimination.
Capabilities and Benchmarks
Falcon OCR handles multi-column layouts, mathematical formulas, tables, and handwriting. It achieved 80.3% on olmOCR (within 1.7 points of the top system) and 88.64 on OmniDocBench, leading all models on Multi-Column (87.1%) and Tables (90.3%) tasks.
Throughput and Efficiency
Due to its small size, Falcon OCR offers high serving throughput. On a single A100-80GB using vLLM, it achieves 5,825 tok/s and 2.9 img/s for a full Layout + OCR pipeline.
Inference Stack
The release includes an inference stack based on PyTorch’s FlexAttention, featuring:
- Paged KV cache and continuous batching to eliminate padding waste.
- CUDA graph capture for the decode loop.
- HR feature cache: An LRU cache for upsampled image features to skip expensive upsampling on subsequent queries for the same image.
Sources
- OriginalFalcon Perception