ricky42613/reranchor-qwen2.5-3b

Before using the model, clone the RerAnchor library:

git clone https://github.com/cycraft-corp/RerAnchor-for-Visual-Document-Grounding.git

Example Usage

import torch
from PIL import Image
from colpali_engine.models import ColQwen2_5, ColQwen2_5_Processor
from reranchor_lib import Qwen2_5_RerAnchor, ReranchorProcessor, denoise_screenshot

# ----------------------
# Load models
# ----------------------
model = ColQwen2_5.from_pretrained(
    "vidore/colqwen2.5-v0.2",
    device_map="cuda:0"
).eval()

processor = ColQwen2_5_Processor.from_pretrained(
    "vidore/colqwen2.5-v0.2"
)

rerank_model = Qwen2_5_RerAnchor.from_pretrained(
    "ricky42613/reranchor-qwen2.5-3b",
    device_map="cuda:0",
    torch_dtype=torch.bfloat16
).eval()

rerank_processor = ReranchorProcessor.from_pretrained(
    "Qwen/Qwen2.5-VL-3B-Instruct"
)

# ----------------------
# Input
# ----------------------
query = "What is the main finding?"
image = Image.open("page.png")

# ----------------------
# Step 1: encode query
# ----------------------
query_inputs = processor.process_queries([query]).to(model.device)
with torch.no_grad():
    query_embed = model(**query_inputs)

# ----------------------
# Step 2: denoise image (RerAnchor)
# ----------------------
denoised_image = denoise_screenshot(
    rerank_processor,
    rerank_model,
    query,
    image,
    k_tokens=200
)

# ----------------------
# Step 3: encode image
# ----------------------
image_inputs = processor.process_images([denoised_image]).to(model.device)
with torch.no_grad():
    image_embed = model(**image_inputs)

# ----------------------
# Step 4: similarity score
# ----------------------
score = processor.score_multi_vector(query_embed, image_embed)

print("Similarity score:", score.item())