LocateAnything — Vision-Language Grounding

NVIDIA Eagle family VLM, based on Parallel Box Decoding (PBD) for single-step parallel prediction of complete coordinates. 12.7 BPS (H100) ≈ 10× Qwen3-VL.

Architecture: MoonViT-SO-400M → MLP → Qwen2.5-3B → PBD

Installation

git clone https://github.com/NVlabs/Eagle eagle && cd eagle/Embodied
pip install -e .
# Optional MagiAttention (Hopper/Blackwell only, long sequences 32K+):
# git clone https://github.com/SandAI-org/MagiAttention.git && cd MagiAttention && git checkout v1.0.5
# git submodule update --init --recursive && pip install --no-build-isolation .

Inference API

from locateanything_worker import LocateAnythingWorker
from PIL import Image

worker = LocateAnythingWorker("nvidia/LocateAnything-3B")
img = Image.open("e.jpg").convert("RGB")

worker.detect(img, ["person", "car"])                    # Object detection
worker.ground_single(img, "the red car")                  # Single-instance grounding
worker.ground_multi(img, "people wearing hats")           # Multi-instance grounding
worker.detect_text(img)                                   # OCR
worker.ground_gui(img, "search button")                   # GUI box
worker.ground_gui(img, "search", output_type="point")     # GUI point
worker.point(img, "the traffic light")                    # Point grounding

# Low-level: worker.predict(image, question, generation_mode="hybrid")
# mode: fast(MTP) | slow(NTP) | hybrid(default)

Output Parsing

Box: \x3Cref>label\x3C/ref>\x3Cbox>\x3Cx1>\x3Cy1>\x3Cx2>\x3Cy2>\x3C/box>
Point: \x3Cbox>\x3Cx>\x3Cy>\x3C/box>
Empty: \x3Cbox>none\x3C/box>

Coordinates [0,1000] integers, divide by 1000 for relative coordinates.

boxes = LocateAnythingWorker.parse_boxes(answer, w, h)  # Pixel coordinates
points = LocateAnythingWorker.parse_points(answer, w, h)

Data Preparation

JSONL (ShareGPT Format)

{"conversations":[{"from":"human","value":"Detect all objects in \x3Cimage-1>."},{"from":"gpt","value":"\x3Cref>car\x3C/ref>\x3Cbox>\x3C100>\x3C200>\x3C400>\x3C500>\x3C/box>"}],"image":"train/00001.jpg"}

Recipe JSON

{"my_data":{"annotation":["a.jsonl","b.jsonl"],"root":"/data/images/","repeat_time":1.0,"data_augment":true}}

repeat_time: ≥1 oversample, \x3C1 downsample. Coordinates normalized to [0,1000].

8 Task Prompts

Plain text: omit image field. Multi-image: image_list + \x3Cimage-1> \x3Cimage-2>.

Training

torchrun --nproc_per_node=8 eaglevl/train/locany_finetune_magi_stream.py \
  --model_name_or_path nvidia/LocateAnything-3B \
  --meta_path "./recipe.json" --output_dir work_dirs/sft \
  --max_steps 25000 --lr 2e-5 --bf16 True --block_size 6 \
  --attn_implementation magi --max_seq_length 16384 --max_num_tokens 25600 \
  --deepspeed deepspeed_configs/zero_stage2_config.json

Key Parameters

Non-Hopper GPU: --attn_implementation sdpa --max_seq_length 4096. OOM: --grad_checkpoint True + reduce --max_num_tokens.

Streaming Packing: Best-Fit + Big-Rocks-First algorithm, checkpoint resume bit-identical. DeepSpeed recommended zero_stage2.

Evaluation

# COCO / LVIS
bash evaluation/scripts/eval_coco.sh --model_path ... --test_jsonl ... --coco_json ... --output_dir ...
bash evaluation/scripts/eval_lvis.sh --model_path ... --test_jsonl ... --lvis_json ... --output_dir ...

# General grounding (Dense200/DocLayNet/HumanRef/RefCOCOg/VisDrone etc.)
bash evaluation/scripts/eval_grounding.sh --dataset Dense200 --eval_type box_eval ...

# Point evaluation / ScreenSpot-Pro
bash evaluation/scripts/eval_grounding.sh --dataset COCO --eval_type point_eval ...
bash evaluation/scripts/eval_sspro.sh --model_path ... --test_jsonl ... --output_dir ...

Requires Rex-Omni fastevaluate + data Mountchicken/Rex-Omni-EvalData likaixin/ScreenSpot-Pro.

Key Results

Model Info

• Name: nvidia/LocateAnything-3B | LLM: Qwen2.5-3B | Vision: MoonViT-SO-400M | Length: 25K
• HF: https://huggingface.co/nvidia/LocateAnything-3B
• Demo: https://huggingface.co/spaces/nvidia/LocateAnything
• Paper: arXiv:2605.27365

License

Code Apache 2.0 | Model NVIDIA License (non-commercial research)

References

• GitHub: https://github.com/NVlabs/Eagle
• Project page: https://nvlabs.github.io/Eagle/