Natural-language object masks
Extract a region by describing its identity, attributes or position instead of supplying manual points and boxes.
Independent tool overview
EVF-SAM-2 at a glance
EVF-SAM2 is an open-source text-prompted segmentation model that combines EVF-SAM's early vision-language fusion with Meta's SAM 2 backbone for image masks and zero-shot video propagation.
Visit the official EVF-SAM-2 site ↗
- Primary task
- Text-prompted image and video segmentation
- Backbone
- SAM 2 with BEiT-3 language-vision encoding
- Model size
- Approximately 0.9B parameters
- Access
- Open weights, code and Hugging Face demo
- License
- Apache 2.0 project and model card
- Hosted API
- No official inference-provider deployment
Overview
What EVF-SAM-2 is
EVF-SAM2 lets a user describe an object—such as “the zebra at the top left”—and returns a pixel-level mask for that target. It is intended for referring-expression segmentation, where the model must connect natural-language meaning with a specific region rather than segmenting from a click or bounding box.
The project combines a BEiT-3 vision-language encoder with SAM 2. The published checkpoint has roughly 0.9 billion parameters, supports images and videos, and can be run through the project's Python code or explored in a public Hugging Face demo.
The name can be confusing: EVF-SAM2 is the SAM-2-backed release of EVF-SAM, not the separately previewed “EVF-SAM v2” research roadmap. It is a developer-oriented model rather than a polished commercial editor or hosted production API.
Use cases
Who EVF-SAM-2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Research prototypes
Evaluate referring segmentation and early vision-language fusion with released checkpoints and benchmark tooling.
Video object segmentation experiments
Use a text-selected object in the initial image context and propagate masks across frames with the SAM 2-based implementation.
Specialized semantic masks
Multitask checkpoints add support for body parts, background regions and broader semantic concepts through documented prompt formatting.
Capabilities
Core EVF-SAM-2 features
Text-prompted segmentation
Accepts a natural-language referring expression and produces a binary mask for the described object or region.
Early vision-language fusion
BEiT-3 jointly processes image and text signals before sending the fused representation to the segmentation components.
SAM 2 image and video support
The SAM 2-backed checkpoint has fewer parameters than the original SAM-H version and adds experimental zero-shot video capability.
Multitask checkpoints
Additional weights support parts, background objects and semantic-level masks beyond standard referring expressions.
Local inference controls
The repository documents FP16 inference plus optional 8-bit and 4-bit loading for developers managing their own hardware.
Evaluation code
Scripts are included for evaluation on referring-segmentation datasets such as RefCOCO, RefCOCO+ and RefCOCOg.
Process
How the EVF-SAM-2 workflow works
Step 1
Test the task in the demo
Upload an image, describe one visible target precisely and inspect the returned mask before committing to local setup.
Step 2
Choose a checkpoint
Use the standard EVF-SAM2 weights for referring segmentation or a multitask checkpoint for parts and semantic masks.
Step 3
Install the project
Clone the official repository, install a compatible PyTorch and CUDA stack, and download the checkpoint from Hugging Face.
Step 4
Run image or video inference
Call the provided scripts with the model type, image or extracted video frames, precision and text prompt.
Step 5
Validate on representative data
Measure mask quality, failure cases, latency and memory on the exact scenes and language your application will use.
Cost
EVF-SAM-2 pricing and free plan
EVF-SAM2 is an open-source research project rather than a paid SaaS product. The model and code can be downloaded without a license fee, while self-hosted compute and any production infrastructure are the user's responsibility.
Hugging Face demo
Free shared demo
Browser-based testing for individual images and sample workflows.
- No local installation for basic evaluation
- Shared availability and resource limits
- Not a production service-level commitment
Self-hosted
No model fee
Run the released code and checkpoint on compatible local or cloud hardware.
- Apache 2.0 repository and model card
- Infrastructure and GPU costs are separate
- Integration, monitoring and scaling are self-managed
- Review all dependency and dataset licenses for the intended use
Pricing checked . Check current pricing at the source ↗
Assessment
EVF-SAM-2 strengths and limitations
Where it stands out
- Selects objects with natural language instead of only clicks, points or boxes
- Released code, weights, demo and evaluation scripts support reproducible testing
- SAM 2 integration extends the approach from static images to video experiments
- Multitask weights cover parts and semantic-level regions in addition to object references
- Quantized-loading options can reduce local memory requirements
What to consider
- Setup assumes Python, PyTorch, model-weight and GPU experience
- The checkpoint is not served by an official Hugging Face inference provider
- Text descriptions can be ambiguous when scenes contain similar objects or unclear relationships
- Video support is described as zero-shot capability from image-based training and needs application-specific evaluation
- The model card does not support a simple AutoModel.from_pretrained workflow
- A public research demo is not a substitute for a secured, monitored production deployment
Compare
EVF-SAM-2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Design
Magic Studio Background Remover
Choose a simple background remover when the only goal is a no-code foreground cutout rather than developer-controlled semantic masks.
Explore Magic Studio Background Remover →Miscellaneous
Hunyuan Vision 1.5 Thinking
Consider a broader vision-language model when image understanding and reasoning matter more than returning a pixel-level mask.
Explore Hunyuan Vision 1.5 Thinking →Miscellaneous
Sam 3D
Consider Meta's SAM 3D tools when the desired output is a reconstructed 3D object or human rather than 2D image segmentation.
Explore Sam 3D →Questions
EVF-SAM-2 FAQs
What is EVF-SAM2?
EVF-SAM2 is a text-prompted segmentation model that combines EVF-SAM's early vision-language fusion with a SAM 2 backbone. It converts a natural-language object description into a pixel mask.
Is EVF-SAM2 free?
The project code and Hugging Face model card use the Apache 2.0 license, and there is no model subscription fee. Running it locally or at scale still creates hardware and infrastructure costs.
Can EVF-SAM2 segment video?
Yes, the repository includes a SAM 2-based video inference path and describes zero-shot text-prompted video segmentation. Teams should test temporal consistency on their own footage.
How is EVF-SAM2 different from SAM 2?
SAM 2 is prompted primarily with visual inputs such as points, boxes or masks. EVF-SAM2 adds a vision-language encoder so a natural-language referring expression can identify the target.
Is EVF-SAM2 the same as EVF-SAM v2?
No. EVF-SAM2 refers to the EVF-SAM implementation built on the SAM 2 backbone. The project separately previewed a future EVF-SAM v2 with additional salient-object and matting capabilities.
Does EVF-SAM2 have an API?
The official project provides local Python inference code and a Hugging Face demo, but the model card is not deployed by an official inference provider. A production API would need to be self-hosted or built through another service.
What hardware does EVF-SAM2 need?
The published checkpoint is roughly 0.9 billion parameters and is designed for GPU inference. The repository documents FP16 plus optional 8-bit and 4-bit loading, but actual memory and latency depend on resolution and workload.
Bottom line
Our EVF-SAM-2 verdict
EVF-SAM2 is a useful open-source bridge between language grounding and SAM 2 masks. It is most compelling for technical teams that need text-selected regions in images or experimental video pipelines and are prepared to own deployment. For a one-click background cutout or a supported commercial API, a simpler hosted product will be easier.
Visit EVF-SAM-2 website ↗Similar Design Tools
View all Design AI tools →
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.