Add DreamSim (human-aligned) encoder: LoRA-tuned DINO ViT-B/16 patch tokens + CLS attention

Add MAE and CLIP encoders alongside DINOv3 (per-model prefix/grid/normalization; MAE ids_restore un-shuffle)

Color final/leaf patch numbers green in reveal animation

add DINOv3 model selector (ViT-L / ViT-H+ / ViT-7B); single-slot cache, dynamic prefix, bf16 for 7B

consistent palette (blue=focus, red=past only); tooltip shows % + raw mean attn value

tab title -> Visual Tree

tooltip: left of cursor + add mean attention; rename to Visual Tree

use latest gradio (6.19) compatible with py3.13 env; sdk_version match

re-add gradio==4.44.1 (was uninstalled); align sdk_version

deps: drop gradio from requirements (let SDK manage), CPU torch, transformers>=4.53; add startup prints

fix Space startup: lazy model load + module-level launch (no __main__ guard)

add .gitignore, drop __pycache__

DINO Hierarchies Gradio app: upload image -> condensed tree + reveal