AcoustiTrace: When Plausible Sound Violates Physics
Paper • 2608.02035 • Published
AcoustiTrace RT60 Evaluator
This repository contains the learned visual RT60 component used by AcoustiTrace, an audio-visual benchmark for diagnosing acoustic plausibility in generated videos.
The checkpoint estimates scene-consistent reverberation time at 500 Hz from
visual acoustic cues. It contains a LoRA adapter for
Qwen/Qwen3-VL-8B-Instruct
and a Sabine-guided continuous physics head. It does not include a copy of
the Qwen3-VL base model.
Files
physics_head.pt
vlm_adapter/
adapter_config.json
adapter_model.safetensors
processor/
acoustitrace_rt60_config.json
SHA256SUMS
Use with AcoustiTrace
git clone https://github.com/Leader-sheng/AcoustiTrace.git
cd AcoustiTrace
python scripts/download_weights.py \
--repo-id Missouter/AcoustiTrace-RT60 \
--local-dir checkpoints/acoustitrace-rt60
python scripts/verify_checkpoint.py checkpoints/acoustitrace-rt60
The frozen base model is downloaded separately from
Qwen/Qwen3-VL-8B-Instruct. Full evaluator setup and end-to-end commands are
provided in the AcoustiTrace repository.
Training sources and license
The RT60 supervision combines Matterport3D RGB-D observations, SoundSpaces 2.0 room-acoustic simulation, and absorption coefficients from the PTB Room Acoustics Absorption Coefficient Database. No upstream training data is included in this repository.
The model artifacts are available for non-commercial academic and research use under the terms in LICENSE. In particular, use remains subject to the Matterport3D Academic Use Terms. The Qwen3-VL base model is separately licensed under Apache-2.0.
Links
Scope
This model is released as the visual RT60 evaluator for AcoustiTrace. Its outputs are diagnostic estimates for benchmark evaluation rather than direct room-acoustic measurements.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for Missouter/AcoustiTrace-RT60
Base model
Qwen/Qwen3-VL-8B-Instruct