Qwen3-VL-4B-SAMTok Composite

This is a self-contained Hugging Face remote-code wrapper for Qwen3-VL-4B-SAMTok. It loads both the Qwen3-VL language-vision model and the SAMTok VQ-SAM2 mask decoder from this single model repository.

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "godx7/SAMTok-self-contained",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)

result = model.generate_with_masks(
    image="example.jpg",
    question=(
        "Could you please give me a detailed description of the image? "
        "Please respond with interleaved segmentation masks for the corresponding parts of the answer."
    ),
)

print(result["text"])
print(result["tags"])
print(result["masks"].shape if result["masks"] is not None else None)

For lower-level usage, the model forwards normal Qwen3-VL calls:

processor = model.processor
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=512)

You can also decode SAMTok text tokens directly:

masks, tags = model.decode_masks(output_text, image="example.jpg")

Files

• config.json: composite remote-code config.
• qwen_config.json: original Qwen3-VL model config.
• model-*.safetensors: Qwen3-VL weights.
• mask_tokenizer_256x2.pth: SAMTok VQ decoder weights.
• sam2.1_hiera_large.pt: SAM2 image encoder checkpoint.
• modeling_samtok.py, configuration_samtok.py, samtok/: remote-code implementation.