STILL-Qwen3.5-9B
Scene Change Verification VLM
1 min readupdated
The STILL-Qwen3.5-9B model solves the task of scene change detection between two visits, determining whether an object has stayed in place, disappeared, or moved. Built on the Qwen/Qwen3.5-9B base model (around 9.4 billion parameters) and the authors' research paper, the model is trained to match frames and object coordinates from two viewpoints. It is useful for researchers and developers working on computer vision, robotics, and spatial monitoring systems.
Key features
- Full fine-tuning of the
Qwen/Qwen3.5-9Bbase model in bf16 format. - Specialized in detecting changes between visits to the same scene.
- Supports two evaluation formats: Scene view (by frames) and Place view (by crops).
- Returns probabilities for
same,changed, andunseenstatuses.
What you can do
- Integrate the model into video surveillance systems for automated change tracking.
- Apply it in robotics to analyze room conditions over time.
- Run evaluations and adjust thresholds using scripts from the official repository.
Example
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
name = "oran9er/STILL-Qwen3.5-9B"
proc = AutoProcessor.from_pretrained(name)
model = AutoModelForImageTextToText.from_pretrained(name, torch_dtype=torch.bfloat16).cuda().eval()
Read more
DetailsBenchmarks: AP 58.1, F1 59.9 on the SceneDiff SD-V test.
Requirements: VRAM for a 9.4B parameter model in bf16.
License: Apache-2.0.










Comments