AGREE

Official model weights for AGREE, introduced in the paper "Few-shot Acoustic Synthesis with Multimodal Flow Matching" by Amandine Brunetto (CVPR 2026).

AGREE is a joint embedding model for acoustics and room geometry. It learns a shared representation between room impulse responses (RIRs) and panoramic depth maps captured at the receiver position.

The model can be used for:

• Evaluating geometry consistency of generated RIRs (via retrieval metrics and Fréchet distance, as done in the paper)
• Downstream multimodal learning tasks involving acoustics and geometry
• Audio-visual representation learning

This repository contains the pretrained weights. To run AGREE, please use the official codebase.

---

Available checkpoints

file	description
AGREE_AR.ckpt	Model trained on the Acoustic Rooms (AR) training set. Intended for downstream tasks.
AGREE_fullAR.ckpt	Model trained on the full AR dataset. Used in the paper for evaluation of RIR generation.
AGREE_fullHAA.ckpt	Model fine-tuned on the full HAA dataset, used for evaluation of RIR generation.

---

Download

Weights can be downloaded with:

huggingface-cli download AmandineBtto/AGREE --local-dir weights/AGREE