Skip to content
Research / Human animation30 October 2026 · Research

Atlas.
People in motion.

40 portraits. Four systems. Atlas ranks first by average score: 78.95 / 80 in this LLM appearance review.

Model labels masked. A shared scoring rubric. Every comparison available below.

Preserving the person.

We compared Atlas, Wan, LongCat, and LTX on 40 synthetic human portraits with varied speech. The review considers likeness, scene fidelity, visual consistency, and mouth appearance. Natural movement matters; a larger mouth opening or more head movement does not automatically mean a better result.

Appearance scores out of 80: Atlas 78.95, Wan 78.40, LTX 71.28, LongCat 65.38.
LLM ratings of sampled frames. Whiskers show 95% intervals across the 40 portraits. Full-size chart.

Atlas recorded the highest average score. It scored higher than LongCat on 40 portraits and higher than LTX on 28, with seven ties. Against Wan, Atlas scored higher on 18, tied on eight, and lower on 14.

The details of a convincing face.

Each category measures a different part of appearance. These scores describe the sampled images; audio synchronization and full-speed motion are not included in these scores. Watch the original videos below to inspect those qualities.

Identity & composition

Identity & composition: four-model comparison, maximum 25 points.
Preserving the person and the scene. Full-size chart.

Scene fidelity

Scene fidelity: four-model comparison, maximum 20 points.
Avoiding invented text, objects, and overlays. Full-size chart.

Visual consistency

Visual consistency: four-model comparison, maximum 20 points.
Consistent appearance across sampled frames. Full-size chart.

Mouth appearance

Mouth appearance: four-model comparison, maximum 15 points.
Plausible lip shapes and facial proportions. Full-size chart.

Made for conversation.

24.46Median displayed FPS
1024 × 1024Video output
40Portraits tested

Atlas recordings were captured through Demo Atlas with audio. Thirty-nine of 40 original recordings delivered at least 22 FPS; the range was 18.70–25.04 FPS. The comparisons below retain each system’s original video and audio timing.

Streaming, without waiting for a finished clip.

Atlas delivered a median 24.46 displayed frames per second in these recordings. Its first rendered speaking frame arrived a median 426 ms after batch submission across 40 session warm-ups. That measures renderer response; input buffering, connection setup, and network playback are outside this latency figure.

Atlas: 426 ms to first speaking frame. Median full generation time: LTX 72 seconds, LongCat 166 seconds, Wan 1276 seconds.
Competitor full-clip generation times from the original runs. Atlas streams before completion; its first-frame response is a different timing measure from these full-clip times.
Atlas service price for five seconds: 0.0097 dollars. Competitor estimated compute per clip: LTX 0.033 dollars, LongCat 0.077 dollars, Wan 0.590 dollars.
Atlas shows the service price for five seconds at $7/hour. Other bars show estimated compute per native clip (4.81–6.92 seconds), not their API prices. Competitor setup, storage, and transfer are excluded; these are different cost bases.
ModelGeneration timeOutput lengthGenerated FPSCost / basis
AtlasStreaming · 426 ms first frame5 s pricing example24.46 displayed$0.0097 service price · $7/hour
LongCat166 s6.92 s1.04$0.077
LTX72 s5.04 s1.68$0.033
Wan1,276 s4.81 s0.06$0.590

Generated FPS is the number of output frames divided by generation time, not the clip’s playback frame rate. Atlas’s displayed FPS above measures delivery through Demo Atlas.

These are observed configurations, not equal-hardware speedups: dimensions, settings, and hardware differ. LongCat uses 39 warm runs; its cold first run took 774 seconds. LTX has 39 exact timings; its first run is bounded at 71.23–73 seconds. Wan includes all 40 runs. Cost uses the recorded October 26–27 compute rate of $1.6635/hour for the shared competitor machine.

Method.

How we evaluated

Forty AI-generated human portraits span professional and casual settings, mostly realistic with some 3D styles. Each system received the same portrait and speech input for each case: 2.64–3.93 seconds of speech, padded with silence to five seconds. Atlas, Wan 2.2 S2V, LongCat Avatar 1.5, and LTX 2.3 are compared using their saved outputs. The separate long-form review uses fresh, approximately thirty-second scripts on the same portraits.

An LLM evaluator scored ten full frames and ten face crops per video with model labels hidden and positions randomized for each portrait. A fixed audio-activity threshold selected the sampling window, with ten equally spaced sample times. The same sampling rule and scoring anchors applied to every system. Scores were finalized before model identities were revealed; all 40 cases for each of the four models are included.

Each criterion receives an integer rating from one to five: identity and composition (25 points), scene fidelity (20), sampled visual consistency (20), and mouth appearance (15). The weighted total is out of 80. Natural blinks and expressions are allowed; larger movements or greater stillness earn no automatic advantage. Ties are retained.

This is an internal, single-evaluator appearance review; the evaluator had prior project context. It measures sampled appearance, not audible lip-sync or full-speed motion. Intervals use 2,000 paired portrait resamples. Atlas’s +0.55-point difference from Wan has an interval of −0.40 to +1.53. Atlas ranks first by the observed average; the interval does not establish a statistically conclusive lead over Wan.

Videos have different durations and lead-ins. The speech-start comparison seeks each original recording to its measured audio onset; this is approximate onset alignment, not matched phoneme timing. Full-recording playback starts at zero. LongCat’s audio packaging was repaired without changing its video; Wan’s quiet ending is shorter. Playback preserves those differences. Atlas playback copies use MP4 with AAC audio for browser compatibility; the video is not re-encoded and the original recordings are retained. Comparison playback copies also place the MP4 index first for faster loading without re-encoding video or audio. Generation settings and hardware differ, so this is not an equal-hardware speed comparison.

Scoring protocol · All scores and per-portrait observations