---
title: GroundAnything-VLM
emoji: 🎯
colorFrom: purple
colorTo: green
sdk: gradio
sdk_version: 5.49.1
python_version: 3.12.12
app_file: space_app.py
pinned: false
license: apache-2.0
models:
  - GroundingPI/GroundAnything-VLM
---

# GroundAnything-VLM

Public English visual grounding demo owned by FanLR, with native inference on ZeroGPU. The purple and lime UI offers nine image tasks, 22 input examples, upload, zoom, target inspection, downloads, and structured GAM output. Video is disabled.

The model is [GroundingPI/GroundAnything-VLM](https://huggingface.co/GroundingPI/GroundAnything-VLM), pinned to `da7d6cdd1b2c58389c551d56c707faa78f1d5b11`. Its own custom processor, chat template, tokenizer and autoregressive generation are used. Model weights load at startup; GPU generation runs inside `spaces.GPU`. Truncated or malformed outputs are reported as errors. An explicit GAM `None` target is a valid absence result.

The separate [GroundingPI frontend](https://huggingface.co/spaces/GroundingPI/groundanything-vlm) calls this Space's `/choose_upload` and `/run_console` endpoints. Each client has its own Gradio session and image state. The Hub's official caller-header bridge forwards the visitor's ZeroGPU identity; no shared owner token is embedded in either demo. GPU availability and quota are checked at inference time.

Only the image and the current task prompt enter inference. Examples contain no stored model answers. Chinese labels use the bundled Noto Sans CJK font. Fonts, research images and model weights retain their respective source licenses; the model repository includes its upstream Kimi-K3 license terms.

For a CPU UI preview, set `GROUNDING_BACKEND=http` and run `python app.py`. This requires an external OpenAI-compatible inference service for predictions. Production uses the pinned native ZeroGPU backend and never retries through that preview service.