CanViT-B fine-tuned end to end for ImageNet-1k classification at an 8×8 canvas grid, from the ImageNet-21k pretrained flagship. It is the canvas-8 counterpart of canvitb16-add-vpe-finetune-g128px-s512px-in1k-2026-04-06-nnx, which was fine-tuned at canvas 32.

CanViT reads a scene through a sequence of 128 px glimpses, accumulating them into a persistent scene-wide latent canvas. Classification reads the recurrent CLS token. See the base checkpoint's card for the architecture.

canvas_grid is an inference-time choice for CanViT: one learnable initial canvas patch is broadcast at rollout start, so the same weights serve any resolution. That holds for the pretrained model. It was not established for LP-FT, where every parameter moves and the fine-tuning geometry could in principle be baked in. This checkpoint is the arm that measures it.

The canvas grid is deliberately not recorded in config.json, because the weights are not tied to it. What this model is, is a set of weights whose fine-tuning ran at 8.

Full ImageNet-1k val, 50,000 images. Policy C2F, T = 21, scene 512 px, glimpse 128 px, canvas grid 8, --mode finetuned. Top-1 %, mean over n = 8 independent policy seeds with 95% percentile bootstrap intervals over runs (10,000 resamples). Evaluated with CanViT-eval at 232ab11 on an RTX 4090.

Accuracy rises 3.58 points over the first five glimpses and then moves 0.12 points across the remaining sixteen. C2F's first viewpoint is the deterministic full scene, so t = 0 is identical in every repeat and has no interval.

The interval covers viewpoint-order variation under a stochastic policy on this one checkpoint. It does not cover training-seed variance.

Initialized fused_probe from canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02, then all parameters trained — backbone, Canvas Attention and head — which is the LP-FT protocol, not a frozen probe.

The only deviation from the archived flagship LP-FT recipe is --recipe.canvas-grid 8. Every other value above is that recipe's default.

The export restores the orbax checkpoint as host numpy and asserts every model leaf is covered, so a missing or renamed parameter fails the export rather than silently producing a partly-initialized model.

config.json carries n_classes: 1000 and the architecture fields (rw_stride, n_canvas_registers, enable_vpe, …). Consumed by CanViT-PyTorch; pass the canvas grid you want at inference.

Fine-tuned and evaluated on static scenes. Numbers above are ImageNet-1k only — this checkpoint has no segmentation or detection result, and a canvas-8 model is not the configuration the CanViT paper reports.