feat: DINOv2 vitl14 fallback when DINOv3 gated access is pending

[perasang]

Summary

• Adds an opt-in USE_DINOV2_FALLBACK=1 env var in generate.py that swaps the DINOv3 image conditioning model for DINOv2 vitl14
• Resizes input to 448 × 448 px (32 × 32 patches + cls = 1025 tokens), matching DINOv3 vitl16 @ 512 px exactly — no shape mismatch in the pipeline
• Zero-diff when the env var is unset; existing behaviour is unchanged

Motivation

facebook/dinov3-vitl16-pretrain-lvd1689m is a gated Meta model that requires manual review before access is granted. New users are blocked at first run until approval arrives. This fallback lets them test the full pipeline immediately using the publicly available DINOv2 vitl14 weights (~330 MB vs ~1 GB for DINOv3).

Usage

USE_DINOV2_FALLBACK=1 python generate.py path/to/image.png

Trade-offs

• DINOv2 and DINOv3 feature spaces differ; output quality may be slightly lower than with the intended DINOv3 model
• Intended as a temporary workaround — switch back to DINOv3 once access is approved

Test plan

• python generate.py (no env var) — behaviour unchanged, DINOv3 path used
• USE_DINOV2_FALLBACK=1 python generate.py assets/shoe_input.png — pipeline loads DINOv2, runs to completion, outputs .glb

🤖 Generated with Claude Code