ego_gen_vroid_single — VRoid synthetic SINGLE-EGO ablation

Ablation of the dual-ego ego_gen_vroid model where each ego view is generated independently (no cross-view information). Same clips/two ego views as ego_gen_vroid, three conditions changed:

1. No reference frames — reference appearance/pose/camera conditioning removed.
2. Own-view-only warp — the main warp selects max-coverage over the SHARED same-seq pool (ego_a + ego_b); ego_b targets end up ~90% sourced from ego_a, so holing is not viable. Here warped_cond is re-warped from the view's OWN ego render pool only (persons masked out) → typical 80–98% coverage, holes where own context cannot reach. src_frame indexes the own pool.
3. Self-only pose — pose keeps only the wearer's identity skeleton (ego_a = p1 = EGO palette, ego_b = p2 = PARTNER palette); the partner body is removed.
4. Per-view Plücker cameras — w2c canonicalised to THIS view's frame-0 (w2c[0] = identity); fed as camera_w2c, the model's shared canon inv(view0-frame0) = I becomes a no-op → per-view coordinate.

Layout

groups/<gid>/train/clips/<clip_id>__ego_a|b/single.npz   # self-contained training unit (below)
manifest/manifest_all.csv, manifest/train_*.csv          # pairing + splits (use train_avatar_disjoint.csv)
captions/                                                 # universal caption + detail originals

single.npz keys (self-contained — no dependency on the main clip.npz)

key	shape	meaning
target_rgb	(77,504,504,3) u8	GT video to generate
warped_cond	(77,504,504,3) u8	own-view-only warp (hole where uncovered)
visibility_packed, vis_shape	packed (T,504,504)	warp visibility
pose	(77,504,504,3) u8	self-only pose (wearer identity)
src_frame	(77,) i32	own-pool source frame index
w2c	(77,4,4) f32	per-view canonical world→cam (frame0 = identity)
K	(3,3) f32	intrinsics; src_res = 504
warp_mode, pose_mode, cam_frame	—	provenance markers

Loader contract (needs a single_ego mode; not wired in the public loader)

Read ONLY single.npz: video←target_rgb, control_input_pose←pose, control_input_warped←warped_cond, control_input_visibility←unpack(visibility_packed,vis_shape), camera_w2c←w2c (per-view; feed directly), camera_K←K, camera_src_res=504. Do NOT load refs / reference_frames / reference_pose, and do NOT enable enable_reference_plucker. Caption = universal A first-person view of C interacting with a partner.