docs: head-to-head vs Wan2.2-Animate, with clips

.gitattributes

@@ -39,3 +39,6 @@ examples/media/reference.png filter=lfs diff=lfs merge=lfs -text
 examples/media/compare-monkey.mp4 filter=lfs diff=lfs merge=lfs -text
 examples/media/compare-parka.mp4 filter=lfs diff=lfs merge=lfs -text
 examples/media/compare-horns.mp4 filter=lfs diff=lfs merge=lfs -text
+examples/media/compare-costume.mp4 filter=lfs diff=lfs merge=lfs -text
+examples/media/compare-prop.mp4 filter=lfs diff=lfs merge=lfs -text
+examples/media/compare-style.mp4 filter=lfs diff=lfs merge=lfs -text

MODIFICATIONS.md

@@ -67,13 +67,20 @@ crop of it (`crop=512:768:389:0`, no scaling):
 The driving clip itself is **not** bundled. `examples/demo.sh` rebuilds it, with the
 documented crop, from your own copy of the base model.
 
-**The comparison clips** — `compare-monkey.mp4`, `compare-parka.mp4`, `compare-horns.mp4` — are a
+**The comparison clips** — `compare-prop.mp4`, `compare-costume.mp4`, `compare-style.mp4` — are a
 different matter. Their driving videos are
 real filmed footage that we hold the rights to, and they are among the clips this model was
-evaluated against. Each file is a four-panel stack: painted reference, driving video,
-Wan2.2-Animate-14B, this model. The Wan2.2-Animate panels were rendered by us from the official
+evaluated against. Each file is a four-panel stack: painted reference, driving video, this model,
+Wan2.2-Animate-14B. The Wan2.2-Animate panels were rendered by us from the official
 [`Wan-AI/Wan2.2-Animate-14B`](https://huggingface.co/Wan-AI/Wan2.2-Animate-14B) weights and code,
-unmodified, at its documented replacement-mode settings — that panel is Wan's output, not ours.
+unmodified, at the replacement-mode settings its own README documents (20 steps, `sample_shift 5.0`,
+`--refert_num 1 --replace_flag --use_relighting_lora`, preprocessing at `--w_len 1 --h_len 1`) —
+that panel is Wan's output, not ours.
+
+Wan's `generate.py` hardcodes a 30 fps output timebase regardless of the source, so its raw files
+claim 4.10 s for motion that is 24 fps. The panels are retimed (`setpts`), not resampled, so no
+frames are dropped and both models play at the same speed. Every panel is letterboxed into the
+driving clip's own geometry; nothing is stretched.
 
 ## Not redistributed here
 

README.md

@@ -56,9 +56,9 @@ The repainted frame comes from an image editor, not from this model — the demo
 
 ## Compared to Wan2.2-Animate-14B
 
-One B200, 480×832, 124 frames at 24 fps, bf16, no compile, no offload, idle card. Wan ran its
-official replacement recipe — 20 steps, `sample_shift 5.0`, `--refert_num 1 --replace_flag
---use_relighting_lora` — after its own preprocessing pass.
+One B200, 480×832, 124 frames at 24 fps, bf16, no compile, no offload. Wan ran its documented
+replacement recipe — 20 steps, `sample_shift 5.0`, `--refert_num 1 --replace_flag
+--use_relighting_lora`, `--w_len 1 --h_len 1` — after its own preprocessing pass.
 
 | | Viggle-Animate | Wan2.2-Animate-14B |
 |---|---|---|
@@ -66,34 +66,40 @@ official replacement recipe — 20 steps, `sample_shift 5.0`, `--refert_num 1 --
 | Render, after weights load | **26 s** | 160 s |
 | — of which sampling | **13.6 s** | 140 s |
 | Forward passes | **3** | 40 (20 steps × 2 chunks) |
-| Parameters | 33 B | 14 B |
+| Parameters | 33.1 B | 17.3 B |
 
-**6× faster per render, 10× on sampling alone** — while being the 2.4× larger model. Wan's
-preprocessing pass is not counted in its 160 s, so the real gap is wider. The win is step
-distillation, not a faster backbone: per forward pass we are 4.5 s to Wan's 3.5 s.
+**6.1× faster per render, 10.3× on sampling alone**, on a model with 1.9× the parameters. Wan's
+preprocessing pass is not counted in its 160 s. The win is step distillation, not a faster
+backbone: per forward pass we are 4.5 s to Wan's 3.5 s.
 
-Four panels each: **painted reference · driving video · Wan2.2-Animate-14B · this model.**
+Four panels each: **painted reference · driving video · this model · Wan2.2-Animate-14B.**
 
-<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-monkey.mp4"></video>
+<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-prop.mp4"></video>
 
-The reference has a monkey's head. Wan replaces it with a bald human one.
+A light-up toy held in both hands. Wan renders it as a featureless red blob, then as a blown-out
+white rectangle for the rest of the clip.
 
-<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-parka.mp4"></video>
+<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-costume.mp4"></video>
 
-A padded parka with a fur-trimmed hood, driven by a man in a fitted dress shirt. Wan returns a slim jacket.
+A white bow headwrap, floral-print leggings and navy slippers, driven by a bearded man. Wan flattens
+the leggings to plain white, drops the bow and the slippers, and lets the driver's beard through.
 
-<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-horns.mp4"></video>
+<video controls muted loop playsinline width="100%" src="https://huggingface.co/Viggle/Viggle-Animate/resolve/main/examples/media/compare-style.mp4"></video>
 
-Horns and a heavy fur coat, driven by a man in a suit. Wan drops both.
+A flat 2D anime character. Wan drifts toward a semi-realistic face and recolours the outfit.
 
-This has a mechanical cause rather than a mysterious one. In replacement mode Wan composites its
-result back inside a segmentation mask traced from the **driving** performer, so anything that
-silhouette does not already contain has nowhere to go. We condition on the painted frame directly,
-with no mask and no skeleton, so the silhouette is free to change.
+The two pipelines carry appearance very differently. In replacement mode Wan regenerates everything
+inside a per-frame mask over the driving performer, and what it has to rebuild that region from is
+one CLIP embedding plus one VAE latent of the character image, against a pose-skeleton video, 512×512
+face crops through a motion encoder, and a background plate with the region painted out — four of the
+five conditioning streams derived from the driving clip. At the documented `--w_len 1 --h_len 1` the
+painted-out region is the performer's whole bounding box, so a held prop sits inside it and is not
+represented in any of those streams. We condition on the painted frame and the driving video, with no
+pose extraction, no segmentation, no background plate and no prompt.
 
 These three are selected. Across the 24 clips we ran head to head, Wan came out ahead on one — a
-tight talking-head shot, where its explicit facial-landmark channel is more expressive than our
-face. That is the same weakness listed under Limitations below.
+tight talking-head shot, where its explicit face channel tracks expression that we render flat. That
+is the same weakness listed under Limitations below. Roughly six were a wash.
 
 ## What this is
 

examples/media/compare-costume.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a45ac3e161df80bff9e4b220e3282855821d450b1d14974b4e886609e001d8e
+size 447259

examples/media/compare-prop.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4128c0e35bf144b5d0e72c5837cc1958421a749337af395977de04fa07b92fea
+size 678070

examples/media/compare-style.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:692ff38363a0162b965d0c6944e9adb98455c8a4e92419b3f9bb764ce9bd6778
+size 536776