akhaliq HF Staff commited on
Commit
2d11fa1
·
verified ·
1 Parent(s): de38ae9

Card: 5-col showcase grid (H3-Image column, s060 identity-preserving rows)

Browse files
Files changed (1) hide show
  1. README.md +11 -10
README.md CHANGED
@@ -40,16 +40,17 @@ control branch (8 union conditions + inpainting), plus verified, reproducible sc
40
 
41
  ## Showcase: extreme poses → man in a suit
42
 
43
- ![pose showcase grid](assets/showcase2/pose_showcase_grid.png)
44
-
45
- Each row, left to right: a text-to-video **source pose** (woman in athletic wear), its **canny
46
- map** driving the ControlNet-Union-2.0 video stack (22-frame repeated control, 40 steps,
47
- `guidance_scale=1.0`, seed 904231), the **no-control baseline** at the same seed, and the
48
- **controlled output** — the same extreme pose as a man in a tailored business suit in a different
49
- setting. Structure comes from the control; identity, clothing and setting from the prompt.
50
- Control strength matters: at 1.0 the canny's body silhouette can carry the source identity with
51
- it; at 0.6 (backbend row) the prompted identity wins **and** the pose survives. Variants at
52
- 0.6/0.8/1.0 + an animated-control test: [`assets/showcase2/variants/`](assets/showcase2/variants).
 
53
  Full-res frames: [`assets/showcase2/`](assets/showcase2).
54
 
55
  ## Official card demos, reproduced
 
40
 
41
  ## Showcase: extreme poses → man in a suit
42
 
43
+ ![pose showcase grid](assets/showcase2/pose_showcase_grid5.png)
44
+
45
+ Each row is one extreme pose. Columns, left to right: the text-to-video **source pose** (woman in
46
+ athletic wear); its **canny map**; the **H3 video model without control** (same seed, control
47
+ gated off); the **ControlNet-Union-2.0 video stack** (22-frame repeated canny, 40 steps,
48
+ `guidance_scale=1.0`, seed 904231, control scale 1.0 for row 1 and 0.6 for rows 2–4); and the
49
+ **H3-Image model text-only** on the same suit prompt — the fast image model gets the clothing
50
+ right but misses the extreme pose geometry (rows 2 and 4), while the controlled video stack
51
+ nails both pose and identity. Structure comes from the control; identity, clothing and setting
52
+ from the prompt — at scale 0.6 the prompted identity wins **and** the pose survives. Variants
53
+ (0.6/0.8/1.0, animated control): [`assets/showcase2/variants/`](assets/showcase2/variants).
54
  Full-res frames: [`assets/showcase2/`](assets/showcase2).
55
 
56
  ## Official card demos, reproduced