research · geometry

fusion perception

a frozen dino backbone projected into the space, making geometry and place language-searchable.

family fusion embedding 2·base dinov2 / dinov3·output 2048-d·license research preview

github ↗research · not yet publicly released

fusion perception is the "perceive" path to the base's "search". it projects a frozen dino backbone into the fe2 space so geometric and place structure (the kind of signal behind segmentation, depth, and place recognition) becomes language-searchable.

it is a research effort and is not yet publicly released, so there is no hugging face repo for it yet.

how it works

a frozen dino backbone (dinov2 / dinov3) is projected into the fe2 space. that projection is what turns geometry-heavy features into something a text query can reach, letting the base's semantic "search" path run alongside a structural "perceive" path in the same coordinate system.

fusion perception architecture: an image passes through a frozen dino backbone and a trained projector into the shared 2048-d space. — architecture: frozen components in grey, the trained / new path in terracotta, all reading out into one shared 2048-d space.

training data

the dino backbone is frozen; only the projection into the fe2 space is trained. it is a research effort; the weights are not yet publicly released.

results

place recognition reaches about 0.804 and projected r-oxford reaches about 0.764: language-searchable geometry, evaluated through the shared readout. it complements the semantic search path with structure; it is not a general geometry solver.

geometry / place retrieval

benchmark	score
place recognition	≈0.804
r-oxford (projected)	≈0.764

honest limitations

—research preview, not yet publicly released, and there is no hugging face repo for it yet.
—it complements the semantic search path with structure; it is not a general geometry solver.

usage

import torch
from transformers import AutoModel

# research preview — weights are not yet publicly released.
# once released, the readout matches the rest of the family:
model = AutoModel.from_pretrained(
    "EximiusLabs/fusion-embedding-2-perception",  # placeholder id
    trust_remote_code=True,
).eval()

image = model.embed_image("street.jpg")
query = model.embed_text("a cobblestone plaza")

similarity = (image * query).sum()

citation

@software{eximius_fusion_perception_2026,
  author = {Eximius Labs},
  title  = {Fusion Perception: language-searchable geometry via a
            frozen DINO backbone projected into Fusion Embedding 2},
  year   = {2026},
  note   = {Research preview},
}

how it fits the family

fusion perception is the geometry-facing member of the family: same shared space, a frozen backbone, a learned projection.

for a physical agent it adds the structural, place-aware channel that pairs with the semantic search the base already provides.

← all models