mhamilton723/FeatUp

Health v2 · maintenance only

Adoption not part of the score

1654 stars · 96 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

FeatUp is a model-agnostic framework that upsamples the spatial resolution of deep neural network features by 16-32x without changing their semantics. It provides pretrained upsamplers and implicit upsampler fitting for vision models, released as official code for an ICLR 2024 paper.

Use cases

• upsample features from vision transformers to high resolution
• improve spatial resolution of backbone features for segmentation
• get high-resolution feature maps from DINO or other models
• boost dense prediction tasks with sharper features
• fit an implicit upsampler to a single image
• use pretrained feature upsamplers via torch hub

When to choose

• you need high-resolution feature maps from a frozen vision backbone
• you're doing dense prediction like segmentation or depth estimation
• you want a drop-in feature upsampler for multiple model architectures

When to avoid

• you need general image super-resolution rather than feature upsampling
• you can't run GPU inference
• you need a production-supported library with long-term maintenance

Facets

library · maturity active

machine-learning deep-learning image-processing computer-vision machine-learning computer-vision deep-learning python cross-platform feature-upsampling vision-transformers iclr-2024 pretrained-models torch-hub dense-prediction research gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="mhamilton723/FeatUp")