Skip to content
Research / Character animationUpdated 1 November 2026 · Research

Zankai.
Characters in motion.

40 portraits. Four systems. Zankai scores 76.8 / 80 in this AI appearance review.

Shared inputs. Model labels masked. One scoring rubric. Every portrait included.

Preserving the character.

We compared the same 40 portraits and speech track across four systems. Zankai scored highest in this model-masked appearance review. The score covers identity, composition, visible artifacts, sampled stability, and mouth appearance; it is not an overall ranking of talking-head quality.

AI appearance scores out of 80: Zankai 76.8, Wan 67.9, LongCat 65.8, LTX 49.5. Horizontal bars with 95 percent case-bootstrap intervals.
Sampled-frame appearance, rated by one AI evaluator with model labels masked. Whiskers show 95% intervals across the 40 portraits. Full-size chart

This developer-led AI review measures sampled appearance on a curated cohort. Read the method below and inspect all 160 clips.

Motion and preservation.

These measurements compare mouth motion with audio and measure preservation of the source background. Each captures a different aspect of the output.

Pixel / audio correlation. Zankai 0.659; LongCat 0.634; Wan 0.329; LTX -0.031.
Mouth-region motion compared with the reference audio. A diagnostic proxy, not a perceptual lip-sync score. Full-size chart
Best-lag proxy. LongCat 12.5 ms; Zankai 72.9 ms; Wan 210.4 ms; LTX 272.9 ms.
Offset with the strongest pixel/audio correlation. Lower is better; this does not measure response latency. Full-size chart
Speech motion coverage. Zankai 0.869; Wan 0.866; LongCat 0.840; LTX 0.683.
The share of speech frames with detected mouth-region motion. Movement alone does not establish correct mouth shapes. Full-size chart
Background similarity. Wan 0.985; LongCat 0.841; LTX 0.830; Zankai 0.812.
Similarity to the source before motion registration. Intended movement and reframing can lower this score. Full-size chart

Motion diagnostics are mixed: in the paired Zankai development comparison, fixed-region correlation increased from 0.559 to 0.659, while tracked-mouth correlation decreased from 0.784 to 0.756. Improved appearance does not establish improved speech synchronization.

Generation speed.

Zankai supports realtime interactive animation. In this comparison, Wan, LongCat, and LTX outputs were generated offline.

Recorded generation time Median bar · observed range · lower is faster

Recorded separately from this appearance evaluation: Zankai warm capture and historical competitor worker records. Different timer boundaries and output lengths. Not a matched speed ranking.

Zankai12.5 sec11.0 sec–15.7 sec
Wan21 min 14 sec2 min 3 sec–21 min 24 sec
LongCat5 min 15 sec5 min 3 sec–24 min 19 sec
LTX1 min 14 sec1 min 14 sec–3 min 3 sec

Method.

How we evaluated

Inputs and comparison

Each system uses the same 40 illustrated portraits and five-second speech input. The gallery includes every scored portrait, with no cases removed. Zankai was captured for this evaluation; Wan, LongCat, and LTX use archived outputs. The portraits were curated in Zankai’s favor, so these results describe this showcase rather than a representative sample of all faces.

Videos use a common 1024 × 1024, 24 FPS, five-second window with an aspect-preserving center crop. Wan’s final frame is held for 0.1875 seconds; longer LongCat clips are trimmed. Returned audio is not replaced or retimed. Source artwork is shared, but input encoding and resizing are not byte-identical.

Appearance review

One AI evaluator reviewed ten full frames and twenty mouth crops per clip at shared sample times. Model labels were hidden, presentation order was randomized, and ratings were finalized before labels were restored. One portrait also received a wider crop at the same four times for every system because the initial crop clipped the mouth.

The same rubric applies throughout: identity and composition (25 points), absence of added content (20), sampled stability (20), and mouth appearance (15). Each criterion is rated from 1 to 5, weighted proportionally, and combined into a score from 16 to 80. The evaluator had prior familiarity with the work; masking reduces identification cues but does not make this an independent or unbiased study.

Scores cover sampled appearance, not audible sync, phoneme accuracy, full-speed motion, or glitches between samples. There are no human ratings. Intervals use 10,000 paired portrait resamples and describe variation within these 40 cases, not reviewer uncertainty. All ratings and measurements.

AI ratings can vary between reviews. In earlier reviews of unchanged Zankai clips, Zankai scored 76.6 then 77.0, while Wan scored 76.8 then 66.2. Those results remain available in the earlier chart and review record. The current gallery contains newly evaluated Zankai outputs.

Diagnostic limits

Pixel metrics use reference audio and fixed source regions, which can miss a moving or reframed mouth. Returned audio differs between systems, so the lag proxy is not a listening assessment. Only three cases passed tracking checks across every arm of the development comparison; tracked results do not establish a cross-system ranking. Background similarity also penalizes intended motion. This benchmark does not validate fresh uploads, current competitor releases, or production readiness.