Atlas and LemonSlice.
Forty faces.
The same words.
Atlas scored 77.60 / 80, compared with 66.93 / 80 for LemonSlice, in a model-masked appearance review of 40 matched portraits.
The same portrait and speech input. Two fresh-context LLM reviewers. Every comparison available below.
Preserving the person.
Identity, the scene, visual consistency, and plausible mouth shapes all matter. Greater movement—or greater stillness—does not automatically earn a higher score.
LLM appearance
Score / 80Atlas scored higher on all 40 portraits. The average paired difference was +10.68 points, with a 95% portrait-bootstrap interval of +9.83 to +11.68 points.
Look closer.
Four categories separate the qualities that make an animated portrait convincing.
Identity & composition
Score / 25Scene fidelity
Score / 20Visual consistency
Score / 20Mouth appearance
Score / 15What happens before the first frame?
Both systems stream video. The measurement boundary matters: an API request and an internal rendering batch begin at different points in the delivery path.
First-frame response
| System | Boundary | Result |
|---|---|---|
| Atlas | Rendering batch → first speaking frame | 426 ms median |
| LemonSlice | Audio submission → first received frame | 1.33 s median |
Service price and usage
| System | Price basis | Usage |
|---|---|---|
| Atlas | $7 / hour | $0.0097 per 5 seconds |
| LemonSlice | Starter: $8/month, including 1,000 credits | 6.14 credits per successful session, median |
Video and audio review.
All 280 evaluation responses are preserved. A content audit found transcript evidence that does not match the supplied speech, so motion and audio scores are withheld while that discrepancy is investigated.
The recorded comparisons remain available below. The separate sampled-appearance results above do not establish lip synchronization or full-speed motion quality.
The whole set, in view.
Browse all 40 portraits. Play each full recording with its own audio. Native lead-ins and durations are preserved; simultaneous playback does not align words.
Portrait 1
algebra tutor“Let's solve this equation together, step by step.”
Atlas
Review notes
Source face, curls, wardrobe and background closely maintained. Lip boundaries remain intact but small exposed tooth detail is soft; no added objects or text.
LemonSlice
Review notes
Recognizable person and office, with softened facial detail and minor cheek/eye-shape variation; first crop is noticeably softer. Exposed teeth are slightly merged in several smiles; no added overlays.
Portrait 2
language tutor“Say the sentence slowly, then notice how the ending changes.”
Atlas
Review notes
Source face and room closely retained with consistent hair and neckline. No added content. Slightly soft tooth/lip detail in open-mouth samples, without major deformation.
LemonSlice
Review notes
Recognizable face, sweater and room; first crop is softer than later crops. No new objects or text. Lip contours stay plausible; teeth are mildly smeared at 1.02 and 2.126 s.
Portrait 3
science tutor“First, predict what happens when we change only this variable.”
Atlas
Review notes
Source face, beard, hair and classroom are preserved across samples; no added elements. Small open-mouth shapes have slightly soft lip/interior detail.
LemonSlice
Review notes
Recognizable source face and classroom, but lower face and eye detail soften in some samples, especially first crop. Teeth have slightly merged detail; no extra overlay or props.
Portrait 4
writing tutor“Your opening works; now give the reader one clear example.”
Atlas
Review notes
Source facial detail, hair, sweater and shelves retained. No added content or visible sample-to-sample structural defects. Small openings preserve coherent lip boundaries.
LemonSlice
Review notes
Identity, gray hair and knitted top retained; first crop softer and jaw narrower than later samples. No added items. Some tooth merging and soft lip edges around the smiling/open samples.
Portrait 5
history tutor“Connect this event to its causes using the timeline.”
Atlas
Review notes
Large eyes, face shape, hair, hoops, jacket and study remain close to source. Small mouth openings have coherent outlines with no visible added content.
LemonSlice
Review notes
Face becomes less stylized with narrower eyes and altered cheek proportions; first crop is softened. Some tooth rows and lip corners are uneven, while clothing and study stay intact.
Beyond a short reply.
All 40 portraits also have longer recordings, using roughly thirty seconds of speech. In the separate long-form sampled-appearance review, Atlas scored 77.68/80 and LemonSlice 72.08/80. Watch every complete recording and explore the long-form results below.
Watch the longer recordings →Method.
Same inputs. A fixed rubric. Labels masked.
The cohort uses 40 synthetic human portraits spanning realistic and 3D styles, professional and casual scenes, and varied speaking roles. Each case supplies the same portrait and exact conditioning audio to Atlas and LemonSlice. The selected LemonSlice configuration is its Default flagship model.
Two fresh-context LLM reviewers each evaluated a disjoint set of 20 portraits. For each portrait, five systems were presented in randomized, model-masked positions with ten full frames and ten face crops per clip. Both reviewers used the same scoring anchors; ratings and documented rubric clarifications were finalized before the model key was revealed. Every clip received one review. Historical Human40 scores were not reused.
Identity and composition contribute 25 points, scene fidelity 20, sampled visual consistency 20, and mouth appearance 15. Each category uses integer anchors from one to five. Paired portrait bootstrap intervals describe variation across these 40 cases.
These scores assess sampled appearance, not audible lip synchronization or full-speed motion. The 95% intervals use 2,000 paired portrait resamples and describe variation within this synthetic cohort. Longer conversations are evaluated separately.
LemonSlice playback files reconstruct received video and audio using arrival timing. The capture API does not expose audio RTP timestamps, so these files do not establish precise audiovisual synchronization. Playing recordings together does not align individual words. The capture workflow completed all 40 portraits: 38 on the first attempt and two after retries. One retry followed a receiver failure before a confirmed provider session; the other followed a capture-stage failure. The first successful capture was used, with no best-of selection.
All fresh scores, category results, and per-portrait observations