visionservex 1.0.0rc1

A permissive-license-aware framework for serving modern computer vision models locally and over Cloudflare Tunnel.

VisionServeX

Secure, beginner-friendly Python API serving for permissive computer vision models
Local inference · Cloudflare Tunnel · Stable JSON API · LLM-agent-friendly

Note on the CI badge: it turns green after the workflow runs successfully on GitHub for the first time. Until then the badge shows "no status" — this is expected for a new repository.

---

VisionServeX is a permissive-license-aware Python framework for running modern computer vision models locally, exposing them through a clean HTTP API, and optionally sharing them securely over Cloudflare Tunnel.

• No CUDA expertise required. visionservex doctor tells you what your machine can run. GPU is preferred automatically when available and healthy; broken CUDA runtimes fall back to CPU with a clear warning.
• One download command. visionservex pull rfdetr-nano — weights are cached and verified.
• Stable contracts. Every prediction returns the same JSON envelope, whether from CLI, Python, or curl.
• Honest. Registry entries say wired, partial, or stub. Stubs never silently fake results.
• Secure defaults. Binds to 127.0.0.1, requires auth for public mode, SSRF and bomb guards on.

---

Quickstart (works on CPU, ~5 minutes)

pip install 'visionservex[server,hf,rfdetr]'

visionservex getting-started          # personalized guide for your machine

# RF-DETR detection — real, fast
visionservex pull rfdetr-nano
visionservex predict rfdetr-nano examples/images/street.jpg --save outputs/out.jpg

# Grounding DINO — text-prompted detection
visionservex pull grounding-dino-tiny
visionservex predict grounding-dino-tiny examples/images/street.jpg \
    --prompt "car,person" --save outputs/gd.jpg

# D-FINE — detection via HF Transformers
visionservex pull dfine-s
visionservex predict dfine-s examples/images/street.jpg --save outputs/dfine.jpg

# Start the API
visionservex serve
curl -F "image=@examples/images/street.jpg" \
     -F "model_id=rfdetr-nano" \
     http://127.0.0.1:8080/detect | jq

Recommendation engine:

visionservex recommend --task detect --simple

---

What works today

Family	Model IDs	Task	Status	Install
Mock (built-in)	mock-*	All tasks	stable	base
RF-DETR	rfdetr-nano/small/base/medium/large	detect	beta	[rfdetr]
RF-DETR-Seg	rfdetr-seg-nano/small/medium	segment	beta	[rfdetr]
D-FINE	dfine-n/s/m/l/x	detect	beta	[hf]
Grounding DINO	grounding-dino-tiny/swin-t/swin-b	open-vocab detect	beta	[hf]
SwinV2	swinv2-tiny/small/base/large	classify	beta	[hf]
SAM v1	sam-vit-base/large/huge	foundation segment	beta	[hf]
SAM 2	sam2-hiera-tiny/small/base-plus/large	foundation segment	beta	[hf]
Grounded SAM	grounded-sam	grounded segment	beta	[hf]
OneFormer	oneformer-swin-large/dinat-large/convnext-large	segment (semantic/instance/panoptic)	beta	[hf]

Not yet wired

Family	Why	Alternative
RTMPose	Requires OpenMMLab toolchain	mock-pose for schema
RTMDet-R/R2 (OBB)	Requires OpenMMLab + mmrotate	mock-obb for schema
Co-DINO-Inst	Requires heavy OpenMMLab	rfdetr-seg-* for instance seg
InternImage	Custom CUDA ops, build required	swinv2-* for classification
SEEM	Expert manual install	oneformer-swin-large
Grounded-SAM-2	Needs upstream sam2 package	grounded-sam (works today)
ONNX export	CLI exists; engine-quality varies	Use HF model repos for ONNX
TensorRT	Future roadmap	—

We make no benchmark claims. Pick by task, license, and hardware. See docs/model_zoo.md.

---

Which model to start with?

I want	Start with	CPU?
Fast detection	rfdetr-nano	yes
More accurate detection	dfine-s	yes (slower)
Text-prompted detection	grounding-dino-tiny	yes (slower)
Instance segmentation	rfdetr-seg-nano	yes
SAM-style masking	sam-vit-base or sam2-hiera-tiny	yes (slow)
Text + mask together	grounded-sam	yes (slow)
Image classification	swinv2-tiny	yes
Semantic scene parsing	oneformer-swin-large	yes (slow)
Just testing/CI	mock-detect	yes (instant)
I have no GPU	Any *-nano or *-tiny model	yes
I have NVIDIA GPU	Run visionservex doctor first — GPU is used automatically when available	—

---

Installation

pip install visionservex                        # base: CLI, registry, mock
pip install 'visionservex[server]'              # + FastAPI HTTP server
pip install 'visionservex[hf]'                  # + D-FINE, GD, SwinV2, SAM, SAM2, OneFormer
pip install 'visionservex[rfdetr]'              # + RF-DETR and RF-DETR-Seg
pip install 'visionservex[server,hf,rfdetr]'    # full recommended install

For OpenMMLab models (RTMPose, RTMDet-R, Co-DINO):

pip install openmim
mim install mmengine mmcv mmpose mmdet mmrotate

See docs/installation.md for platform-specific notes.

---

Python API

from visionservex import VisionModel

# Object detection
m = VisionModel("rfdetr-nano")
result = m.predict("image.jpg")
for det in result.detections:
    print(det.label, f"{det.score:.2f}", det.box.to_xyxy())
result.save("annotated.jpg")

# D-FINE detection (HF Transformers)
m = VisionModel("dfine-s")
result = m.predict("image.jpg")

# SAM 2 (point prompt)
m = VisionModel("sam2-hiera-tiny")
result = m.predict("image.jpg", points=[[x, y]], point_labels=[1])

# SAM 2 (box prompt)
result = m.predict("image.jpg", boxes=[[x1, y1, x2, y2]])

# OneFormer (choose task)
m = VisionModel("oneformer-swin-large")
result = m.predict("image.jpg", task="semantic")       # or "instance", "panoptic"

# Grounding DINO
m = VisionModel("grounding-dino-tiny")
result = m.predict("image.jpg", prompts=["red car", "person walking"])

# Auto-pull on first use
m = VisionModel("dfine-s", auto_pull=True)
result = m.predict("image.jpg")

Stable result fields: kind, model_id, task, device, precision, backend, latency_ms, model_loaded_from, fallback_reason, warnings.

---

HTTP API

Stable response envelope:

{
  "request_id": "...",
  "status": "completed",
  "model_id": "dfine-s",
  "task": "detect",
  "backend": "huggingface_dfine",
  "device": "cpu",
  "precision": "fp32",
  "latency_ms": 187.4,
  "results": [{"box": {...}, "score": 0.72, "label": "person", "class_id": 0}],
  "warnings": [],
  "metadata": {}
}

Error envelope:

{
  "request_id": "...",
  "error": {
    "code": "MODEL_MISSING",
    "message": "Model weights for 'dfine-s' are not cached.",
    "hint": "Run: visionservex pull dfine-s",
    "details": {}
  }
}

Key endpoints: GET /health, GET /devices, GET /models, POST /detect, POST /segment, POST /classify, POST /open-vocab/detect, POST /grounded-segment, GET /jobs/{id}, GET /metrics. Full reference: docs/api_reference.md.

---

Security defaults

Setting	Default
Server bind	127.0.0.1 only
Public mode	disabled (explicit opt-in)
Authentication	disabled — enable before exposing
Remote URL inputs	disabled (SSRF protection)
CORS	disabled
Upload limit	20 MiB
Image pixel limit	~33 MP (decompression-bomb guard)
Rate limit	120 req/min per IP
Token redaction	enabled in all logs

See docs/security.md and SECURITY.md.

---

Safe Cloudflare Tunnel

export VISIONSERVEX_AUTH__ENABLED=true
export VISIONSERVEX_AUTH__API_KEY=$(python -c "import secrets; print(secrets.token_urlsafe(48))")

visionservex tunnel doctor
visionservex tunnel create visionservex
visionservex tunnel route visionservex api.yourdomain.com
visionservex tunnel config api.yourdomain.com --out tunnel.yaml
visionservex serve &
visionservex tunnel run tunnel.yaml --i-understand-this-is-public

The CLI refuses without auth enabled and the explicit confirmation flag. The generated config always ends with a catch-all http_status:404 rule. See docs/cloudflare_tunnel.md.

---

Documentation

Beginner quickstart	First prediction in 5 min
Device check	GPU/CPU/MPS diagnostics
Model zoo	All models, license table, "which model?"
Model downloads	Download system, auto-pull
Model licenses	Per-model license details
Cloudflare Tunnel	Safe public exposure
Security	Threat model, all protections
HTTP API reference	Endpoints, error codes
Python API	VisionModel, result types
CLI reference	Every command
Troubleshooting	Common errors
LLM agent guide	Stable CLI/JSON for agents
About	Author, citation, acknowledgment

---

License and upstream models

VisionServeX is Apache-2.0 (SPDX-License-Identifier: Apache-2.0). See LICENSE and NOTICE.

Each integrated model retains its own upstream license. Review the model, checkpoint, and training-data licenses before commercial use. VisionServeX does not provide legal advice. See docs/model_licenses.md.

---

Citation

@software{sajjadi2026visionservex,
  author = {Arash Sajjadi},
  title  = {{VisionServeX: A permissive-license-aware framework for
             local computer vision model serving}},
  year   = {2026},
  url    = {https://github.com/arashsajjadi/VisionServeX},
  note   = {Developed under the supervision of Prof. Mark Eramian,
            Department of Computer Science, University of Saskatchewan.}
}

Author: Arash Sajjadi — PhD Candidate, Department of Computer Science, University of Saskatchewan
Supervision: Prof. Mark Eramian, Computer Vision Lab, University of Saskatchewan
(This project is not an official product of the University of Saskatchewan.)