Same faces.
Thirty seconds of speech.
Forty portraits, each with its own longer script. 40 of 40 paired recordings are available.
Appearance through longer speech.
Twenty full frames and twenty face crops per recording sample the longer speaking interval. The same four scoring categories apply to both systems.
Long-form LLM appearance
Score / 80Across 40 scored pairs, Atlas scored higher in 37, tied in 3, and scored lower in 0. The average Atlas-minus-Lemon Slice difference was 5.60 points; the 95% portrait-bootstrap interval was 4.70 to 6.55.
Identity & composition
Score / 25Scene fidelity
Score / 20Visual consistency
Score / 20Mouth appearance
Score / 15These are sampled appearance scores. They do not measure full-speed motion, exact lip synchronization, or every frame of the recording.
Longer recordings: delivery and usage.
40 of 40 Lemon Slice captures completed on their first attempt, with verified playback files. Source speech recordings ranged from 28.35 to 31.44 seconds.
Lemon Slice delivery
| Measurement | Median · 40 captures |
|---|---|
| API session request | 6.86 s |
| Audio submission → first received video | 1.22 s |
| Full-span received SDK cadence | 24.01 frames/s |
Credits and service-price example
| Scope | Usage or advertised price |
|---|---|
| Lemon Slice long session | 16.49 credits, median |
| All 40 long captures | 670.11 credits |
| Short + long runs, all attempts | 928.91 of 1,000 included credits |
| Atlas advertised service price | $7/hour · $0.0583 per 30 seconds |
Received SDK cadence uses the frame count minus one divided by the first-to-last native timestamp span, including idle tails and internal gaps. It measures delivery cadence, not provider compute speed.
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.
Watch the complete recordings.
Each pair uses the same portrait and source audio. These are separate long generations, not repeated short clips. Native starts and quiet endings are retained; simultaneous playback does not align individual words.
Lemon Slice playback uses approximate receiver-arrival audio/video alignment, so it cannot establish exact lip synchronization.
Portrait 1
algebra tutorAtlas
Review notes
identity_composition: Identity, source framing, hairstyle and clothing remain closely preserved at 2.200, 16.249 and 31.848 seconds.; no_added_content: No visible invented props, text, scene changes or malformed anatomy.; sampled_stability: Stable facial proportions and background in early, middle and late samples; blink at 25.607 seconds is normal.; mouth_appearance: Lip opening looks mildly softened at 3.760, 16.249 and 24.047 seconds; no major mouth deformation in the sampled frames.
Lemon Slice
Review notes
identity_composition: Identity, hairstyle, blazer, blouse and classroom remain recognizable at 0.767, 14.809 and 30.431 seconds.; no_added_content: No invented text or props; small lower-edge hand presence is plausible natural gesture.; sampled_stability: Face and scene remain coherent across early, middle and late samples; normal blinks and head shifts are not defects.; mouth_appearance: Open smiles at 0.767, 13.273 and 28.850 seconds show slightly uniform bright teeth and softened mouth detail; lip outline otherwise plausible.
Spoken script
Let's solve this equation together, step by step. Imagine that three identical boxes and two loose pencils make fourteen pencils altogether. First remove the two loose pencils from both sides. Now divide what remains equally among the three boxes. Substitute your answer into the original equation to check it. There is no need to rush; understanding each part makes the whole process easier. Take a moment to think about that example. Then explain the next step in your own words.
Portrait 2
language tutorAtlas
Review notes
20 full frames and 20 face crops inspected at original sheet resolution, including early/middle/late samples.
Lemon Slice
Review notes
20 full frames and 20 face crops inspected at original sheet resolution, including early/middle/late samples.
Spoken script
Say the sentence slowly, then notice how the ending changes. Try describing what you do today, then describe what you did yesterday. Listen for the change in the verb rather than rushing through the sentence. Repeat your example with a different subject, and notice which words stay exactly the same. There is no need to rush; understanding each part makes the whole process easier. Take a moment to think about that example. Then explain the next step in your own words.
Portrait 3
science tutorAtlas
Review notes
identity_composition: Identity and source composition closely preserved at 2.201, 16.266 and 31.869 seconds.; no_added_content: No visible new object, text, clothing or malformed anatomy.; sampled_stability: Hair, face proportions, shirt and classroom remain stable across early, middle and late samples; blink at 19.386 seconds is ordinary.; mouth_appearance: Mild softness around teeth/lip opening at 11.571, 16.266 and 25.625 seconds; no severe contour deformation.
Lemon Slice
Review notes
identity_composition: Source identity, dark curly hair, blue shirt and classroom preserved at 1.426, 15.467 and 31.008 seconds.; no_added_content: Hands at 1.426 seconds are plausible gestures; no invented object or text is apparent.; sampled_stability: First face sample at 1.426 seconds is noticeably softer than later samples; face contours remain coherent at 15.467 and 31.008 seconds without progressive drift.; mouth_appearance: Teeth appear slightly fused/bright at 6.112 and 29.432 seconds; rounded lips at 10.765 and 26.341 seconds remain anatomically plausible.
Spoken script
First, predict what happens when we change only this variable. Imagine testing how sunlight affects two seedlings. Give both plants the same soil, water, and container size. Place one near the window and the other farther away. Record what you observe each day before deciding whether your prediction was supported. There is no need to rush; understanding each part makes the whole process easier. Take a moment to think about that example. Then explain the next step in your own words.
Portrait 4
writing tutorAtlas
Review notes
20 full frames and 20 face crops inspected at original sheet resolution, including early/middle/late samples.
Lemon Slice
Review notes
20 full frames and 20 face crops inspected at original sheet resolution, including early/middle/late samples.
Spoken script
Your opening works; now give the reader one clear example. Choose a familiar moment, such as waiting for a bus in the rain. Replace a general description with a sound, a small action, and one concrete detail. Read the paragraph aloud and remove any sentence that repeats what readers already understand. There is no need to rush; understanding each part makes the whole process easier. Take a moment to think about that example. Then explain the next step in your own words.
Portrait 5
history tutorAtlas
Review notes
identity_composition: Stylized face, large eyes, earrings, curled hair and navy jacket retain source identity at 2.281, 15.492 and 30.215 seconds.; no_added_content: Clothing, belt, map, books and globe remain coherent without visible invented props/text.; sampled_stability: Stable stylized proportions and background across the complete early/middle/late sample groups.; mouth_appearance: Mildly blurred small teeth and lip interior at 5.201, 12.562 and 24.337 seconds; outline remains plausible.
Lemon Slice
Review notes
identity_composition: Source stylized identity, hair, earrings and outfit remain recognizable at 0.697, 14.184 and 29.118 seconds. Expression changes alone are not identity defects.; no_added_content: No new text or objects and no obvious malformed anatomy in the observed frames.; sampled_stability: First sample at 0.697 seconds has noticeably softened/blurred face detail; middle and late samples recover coherent contours without sustained drift.; mouth_appearance: Teeth appear somewhat uniform at 0.697, 11.178 and 27.622 seconds, with slight softness at 17.156 seconds. Lip shapes remain plausible.
Spoken script
Connect this event to its causes using the timeline. A useful timeline shows more than dates. Put the major event in the center, then place earlier decisions on one side and later consequences on the other. Distinguish a condition that made change possible from the immediate action that triggered it. There is no need to rush; understanding each part makes the whole process easier. Take a moment to think about that example. Then explain the next step in your own words.
Method.
The same 40 portraits each use a longer script, with about thirty seconds of source speech. Two fresh-context LLM reviewers each assess a disjoint set of twenty portraits, with Atlas and Lemon Slice randomly labeled A or B for each pair. Each clip receives one review. The rubric is fixed before review and both rating sets are finalized before model labels are revealed.
Twenty full frames and twenty expanded face crops are sampled across each recording’s audio-active interval. The rubric assigns 25 points to identity and composition, 20 to scene fidelity, 20 to sampled visual consistency, and 15 to mouth appearance. Greater movement or greater stillness alone does not earn a higher score. Intervals use 2,000 paired portrait bootstrap resamples.
Native recording gaps remain visible. In portrait 009, Atlas has a 3.2-second video pause during speech; two sampling targets therefore select the same subsequent frame. This capture remains in the gallery and review. The sampled scores do not replace full-recording playback or quantify exact audio/video synchronization.