Positioning the work
Novelty: “A frozen vision foundation model's dense feature field can be represented by a small set of 2D Gaussians whose features are solved in closed form and whose placement is predicted in one pass by an unrolled clustering-and-optimisation network; one model serves 16 to 256 Gaussians and comes within about 5% of per-image optimisation at 64. We found no published method that combines these.”
Limits: every ingredient exists elsewhere (Feature 3DGS, DINOSAUR, variable projection, unrolling, Matryoshka); “not found” is a search; GViT v2 and GPSToken not read in full; 5% is 64 Gaussians, 32 steps, 100 images, one seed (13% at 16 steps, 32% at 256); semantics inherited; teacher at inference.
Contributions: (1) closed-form features, −22%, three seeds; (2) amortised placement, 0.000104 → 0.0000804 → 0.0000587 → 0.0000531 against about 0.0000465; (3) one model for 16 to 256, within about 4% of k-means + 16 steps; (4) target resolution measured, input512 chosen; (5) negative results on hierarchy; (6) Triton kernels, as engineering. Per-count trained models (one seed): 1 to 2% better at 16 and 32, worse at 128 and 256.
Five hardest questions: just k-means / upsampler / GViT? Teacher at inference? Where is the semantics? One seed and reused holdout? Hierarchy failed, what is left?
Venues: CVPR 2027 closed; ICML 2027 full paper 22 January 2027; ICCV and NeurIPS 2027 dates unverified; TMLR. Minimum for ICML: P1 downstream at matched count with baselines, P6 rigour.
Talk: question → representation → placement bottleneck (fig1) → closing the gap (fig2) → target too coarse (fig3) → budget curve (fig4) → what did not work (fig5) → position, limits, next experiment.