da3mono_loop — qk256fa5b

Training checkpoints of the qk256fa5b version of da3mono_loop (DA3MONO-LARGE with a looped last layer, 1 camera + 4 register tokens, a cross-view coordinate stream and feat_acc -> GeoRoPE), trained on ScanNet++ (nvs_sem_train) and Hypersim (train).

Code, architecture, data and evaluation protocol: GitHub Yusin2Chen/DVLT-Matching, branch qk256fa5b. Training log: wandb yusin/vitgeo / da3mono_loop_qk256fa5b_1792062426 (step axis = global step).

Each checkpoint <name> (qkfa5b_long_<job>_step<global step>) has three files:

file	content
`<name>.pt`	model state dict (all tensors; `LoopModel.load_state_dict(strict=True)`)
`<name>_opt.pt`	AdamW state (loaded automatically by `train.py` next to `INIT_CKPT` for a true resume)
`<name>.step`	global step, pass as `STEP_OFFSET`

The DA3MONO-LARGE architecture config comes from depth-anything/DA3MONO-LARGE (revision f465978e618db8cc79c83b8bbf24964857db1875); scripts/download_weights.py in the code branch downloads everything.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Depth Estimation

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support