# Video breakdown report: social_account_doctor_human_v1.mp4

> English translation of [report.md](report.md), the Markdown report rea-video generated for this video. Section order and content follow the tool output. Text recognized on screen (the "text track" lines) is kept in the original Chinese, since that is what appears in the frame.

**Verdict**: mixed production: code / template rendering 59% + AI-generated video 41%; overlays: text labels, captions, title (composited in code / an editor); real material used: screen recording (8.2 s), UI screenshots (4.8 s); audio: voice + BGM at 121.6 BPM (library music) + 1 sound effect

- Duration 19.042 s · 1080×1920 · 24.0 fps · 9:16
- 10 shots, 1.9 s average; transitions: cut×9, start×1
- Confidence (duration-weighted): 74%; method: deterministic forensics fused with a vision model (VLM)
- VLM active (gemini-3.1-pro-preview, gemini-3.1-pro-preview-pc, gcp-gemini-3-pro-preview); the agent (claude) visually reviewed 1 of 10 shots, merged into the verdict

![storyboard](storyboard.jpg)

## Production method share

| Underlying picture | Duration | Share |
|---|---|---|
| Code / template rendering | 11.17 s | 59% |
| AI-generated video | 7.88 s | 41% |

## Remake plan

**5 AI shots → 1 generation unit, 1 Seedance 2.5 call; 5 code-rendered / screen-recording shots → 1 code project (5 segments)**

- The default engine, Seedance 2.0 (including Mini / Fast), cannot produce several shots in one call; with it you would generate the 5 shots one by one. This plan assumes Seedance 2.5 (≤30 s per call, multi-shot in one call): 1 call in total
- Grouping rules: AI shots in the same set joined by hard cuts are merged into one generation (S01); ≤30 s per call; consecutive hard cuts inside a unit should stay ≤8 s, at most 10 s (S06); split at dissolves / fades and set changes (S03)
- Audio: the model can generate action sound effects and ambience along with the picture; voice-over is better done as separate TTS and mixed in, so you can change the script and timing

### Generation units

| Unit | Shots | Time | Duration | Method | Continuity |
|---|---|---|---|---|---|
| G1 | #2, #3, #5, #7, #9 | 1.4–3.9 s, 5.8–7.5 s, 10.5–11.5 s, 13.5–16.1 s | 7.9 s | multi-shot in one call | First frame: generate an image with the composition of the original shot 2 first frame (shots/shot_002_first.jpg) |

**G1 · AI-generated video · shots #2, #3, #5, #7, #9 · 7.9 s**

- Reference material: `@image1` character reference (one frontal + one half-body per character) to keep faces consistent across shots; `@video1` the original 1.4–3.9 s, 5.8–7.5 s, 10.5–11.5 s, 13.5–16.1 s segments (cut out the inserted UI shots and join into one clip) as a reference for motion, camera and cut rhythm (reference only, do not copy the picture)
- First-frame image prompt: photorealistic, a young Chinese woman with short hair and light-green retro glasses, yellow hoodie, sitting cross-legged on a living-room sofa at night hugging a cushion, holding a phone in one hand, warm floor lamp and houseplants in the background, soft cinematic light, medium shot, eye level, vertical
- These shots are separated by inserted UI segments in the final cut: generate them as one continuous clip, then put the inserts back in the edit

```text
Vertical 9:16; night / low-light ambient, no dedicated fill, soft light (almost no shadow edge), cool tint (daylight / cool-white LED), background slightly soft. Character per @image1. Follow @video1 for motion, camera and cut rhythm. Hard cuts between shots.
Shot 1 (0.0–1.5 s): static shot, the woman speaks with her mouth slightly open, slightly confused and surprised, eyes forward, fingers lightly holding the phone, small head movements, body stays seated and steady, soft night light.
Shot 2 (1.5–2.5 s): static medium close-up, a young woman in a newsboy cap and glasses holds a coffee mug and speaks naturally, blinks, expression shifts from slight thought to a soft smile, both hands steady on the mug, desk-lamp light steady in the background, cinematic interior light.
Shot 3 (2.5–4.2 s): static medium shot, a short-haired woman with glasses on a sofa at night looks down at a glowing phone, then looks up in surprise, mouth open, and speaks toward the camera, lively expression, stays cross-legged, warm floor lamp softly blurred behind, photorealistic cinematic look.
Shot 4 (4.2–5.2 s): static shot, a woman with glasses hugs a cushion on the sofa and looks at her phone, mouth wide in surprise as she speaks, eyes on the screen, cool blue-white phone light flickers on her face, natural small movements.
Shot 5 (5.2–7.9 s): image as the base, audio-driven avatar, lip sync, slight head movement and blinking, body and background static.
```

<details><summary>English prompt (as generated by the tool)</summary>

```text
Shot 1 (0.0-1.5s): Static shot, a young woman with short hair and glasses talking with slight confusion, holding a phone, sitting on a sofa at night, minimal head movement, cinematic cozy lighting.
Shot 2 (1.5-2.5s): Static medium close-up shot of a young East Asian woman with a newsboy cap and round glasses sitting on an armchair, holding a ceramic mug with both hands, speaking and subtly smiling. Warm cinematic lighting, blurred bookshelf and cozy practical lamp in the background.
Shot 3 (2.5-4.2s): Static medium shot, a short-haired Asian girl wearing turquoise glasses and a yellow hoodie sits cross-legged on a sofa at night holding a glowing phone, looks up in surprise and starts talking, realistic movie cinematography, subtle screen light on face, warm floor lamp in soft-focus background.
Shot 4 (4.2-5.2s): Static shot, a young woman wearing round glasses sits on a couch hugging a cushion, looking at her smartphone with an astonished expression and speaking, phone light illuminating her face.
Shot 5 (5.2-7.9s): Audio-driven avatar from image, lip sync, slight head movement and blinking, static body and background.
```

</details>

### Code compositing

One Remotion / HTML project rebuilt from the text tracks and layout (text and UI 100% controllable; changing copy means changing data). An LLM can write this project directly.

- Segment 1: shot #1 (0.0–1.4 s)
- Segment 2: shot #4 (3.9–5.8 s)
- Segment 3: shot #6 (7.5–10.5 s)
- Segment 4: shot #8 (11.5–13.5 s)
- Segment 5: shot #10 (16.1–19.0 s)

## Framing, lighting and editing

- Framing: close-up 100%; eye-level angle; subject is a face (framing measured by face height)
- Camera: fixed (tripod / stand / on a table) 66%, moving (gimbal / handheld pan) 34%
- Lighting: night / low-light ambient, no dedicated fill. Key light frontal (from the camera direction, both sides of the subject equally bright), from below (screen / table bounce / low source); contrast ratio about 1.3:1 (0.4 stops); medium softness; cool color temperature (daylight / cool-white LED); lower half of the subject brighter: possibly light from below (screen, table bounce); subject and background differ in color temperature: mixed sources (e.g. warm lamp + cool window / screen); background slightly soft
- (The picture is not live footage: the above describes the light in the frame, for writing generation prompts, not a shooting plan)
- Grade: natural (close to source); normal saturation, normal contrast, normal blacks, neutral overall
- Editing: 28.4 cuts per minute, 1.9 s average shot; no live-action shots

| Dimension | Item | Conclusion | Source |
|---|---|---|---|
| Framing | Shot size | Close-up | Measured: close-up; VLM: medium shot (half body) |
| Framing | Angle | Eye level | VLM: eye level |
| Framing | Composition | Subject centered, vertical | VLM: subject centered, vertical |
| Camera | Support | Tripod / stand, fixed | Measured: mixed, fixed 66% / moving 34%; VLM: tripod / stand, fixed |
| Camera | Lens | Camera portrait / telephoto | VLM: camera portrait / telephoto |
| Camera | Setups / scene | Two facing seated indoor setups (a woman in a yellow hoodie hugging a cushion, frontal medium close-up; a woman in a white shirt and green vest holding a mug, frontal medium close-up), intercut with phone diagnostic-report and code-repo screenshots | VLM: same |
| Camera | Likely gear | Mirrorless or cinema camera (natural shallow depth of field) with a fast prime, lavalier mic | VLM: same |
| Lighting | Scene | Indoor; night; living-room sofa | VLM: same |
| Lighting | Key | Front left, slightly high, softbox | VLM: same |
| Lighting | Softness | Medium | Measured: medium; VLM: soft |
| Lighting | Fill | Weak fill from the right | VLM: same |
| Lighting | Rim | None | VLM: none |
| Lighting | Catchlight | Square | VLM: square |
| Lighting | Visible sources | Warm desk lamp, background left | VLM: same |
| Lighting | Color temperature | Cool (daylight / cool-white LED) | Measured: cool; VLM: warm |
| Lighting | Setup | Front-left soft key + warm practical lamp in the background | VLM: same |
| Lighting | Overall | Cozy warm home-at-night lighting, a desk lamp as a warm practical in the background, a large softbox giving the face a clean soft key | VLM: same |
| Grade | Style | Natural (close to source) | Measured: natural |

- Note: contrast ratio, softness, color temperature and shot size are measured on the subject's pixels (the face when there is one, otherwise the sharpest region, which is an estimate); platform transcoding crushes shadows, so softness and fill read conservative; angle, light source type, catchlight and rim come from visual review. Where measurement and visual review disagree, both are listed.

## Metadata clues

- Encoder: `Lavf61.7.100`

## Audio breakdown

![audio](audio.png)

- Overall loudness −16.8 dBFS, peak −3.5 dBFS
- **Voice**: present, 79% of duration, fundamental about 247.8 Hz (reads female)
- **Background music**: 39% of duration (0.0–10.5 s), about 121.6 BPM (or 60.8 / 243.2), key F major (confidence 0.795), section changes at 4.5 s, 8.5 s
- **Cuts vs beat**: 0% of cuts land on a beat (random baseline about 12%)

**Audio source: AI-generated or existing material**

- **Conclusion**: video type: AI video with native audio (each segment carries its own sound) → voice: native AI-video dialogue (generated with the picture) (leaning, to be confirmed); music: library music; sound effects: generated by the video model with the picture (native audio)
- **Separated stems**: `stems/original.wav` (original), `stems/vocals.wav` (dry voice; transient effects land here too), `stems/music.wav` (music + ambience)
- Video form: narrative; AI picture 41%, live footage 0%
- **Background music**: library music (existing_library 73%, ai_generated 13%, popular_song 13%)
  - The music is a bed under the picture: for an ordinary video the bed is most likely from a library (prior)
  - The voice is speech / narration rather than singing (sustained-pitch ratio 0.12): this is the ordinary "narration + BGM" form
  - VLM listening judgment: library music
  - How to confirm: on the Douyin / Xiaohongshu page check the music tag: a song title → existing track; "original sound by @author" → uploaded by the author (for music videos this is usually an AI song)
  - How to confirm: run Soda Music / QQ Music song recognition or ACRCloud: a hit means an existing song
  - How to confirm: listen: AI songs often have slurred diction, thin metallic highs and lyrics that track the picture closely; library beds are mostly instrumental, standard structure, cut off or faded at the end
- **Voice**: to be confirmed (ai_video_native 44%, human_recording 44%, tts 12%)
  - VLM judged it a human voice
  - Background sound jumps at every cut and the picture is AI video: the voice was most likely generated by the video model together with the picture (e.g. Seedance / Veo / H3 native audio)
- **Sound effects**: generated by the video model with the picture (native audio) (ai_native 71%, library_editor 24%, live_sync 6%)
  - Background / ambient sound jumps at cuts (24.7 dB at cuts vs 9.4 dB within shots) and the picture is AI video: each segment's effects / ambience are native to the video model
- Per-segment native-audio check: background change 24.7 dB at cuts, 9.4 dB within shots (ratio 2.63) → each segment carries its own sound

**Sound effects** (1; source: vlm = heard by the model, deterministic = signal-detected candidate, low confidence)

| Time | Type | Description | Matching picture | Source |
|---|---|---|---|---|
| 10.76 s | pop |  |  | deterministic |

**Audio remake plan**

- voice (A01-narration): human voice-over / dialogue, two female voices: character A is a young short-haired woman at home (bright, puzzled), character B is a woman in a cap with a composed, mentoring tone; moderate pace, brisk rhythm, natural everyday dialogue → capsule-cinema DoubaoTTSTool / UniversalTTSTool (or a MiniMax cloned voice); use duration-controlled TTS (A02) when you must fit the original timing
- music (A03-music): low-key, gentle, melody-free ambient bed; focused, slightly suspenseful and mild; 121.6 BPM; F major → the original leans library: search a library by BPM / mood / key, or generate a royalty-free track in a similar style with AI
- sfx (A04-sfx): 1 effect → the original effect was generated with the AI video segment: for a remake use a video model with native audio (Seedance 2.x / Veo 3.1 / Kling), or video-to-audio (HunyuanVideo-Foley)
- ambience (A04-sfx): quiet indoor conversation ambience → ambience stock or a video-to-audio model
- Mix: voice crisp and clearly in front, BGM very low as a bed only, no noticeable effects

## Structure and copy (VLM)

- Summary: a two-person scripted dialogue between a home creator and a professional mentor shows an open-source Xiaohongshu account-diagnosis tool diagnosing a post's cover images and title and suggesting fixes.
- Form: narrative; hook: opens straight on the pain point with a screenshot and a provocative claim: a post published 21 minutes ago is judged 90% likely to flop, which grabs creators' attention instantly.
- Narration: human, two female voices (character A young, short hair, bright and puzzled; character B in a cap, composed mentor); music: low-key, gentle, melody-free bed; captions: white sans-serif with black outline, centered low; a persistent bold title at the top reads "笔记没人看，先别删" ("Nobody's reading your post? Don't delete it yet")
- Overall judgment (VLM): live footage combined with screen recording / screenshots. Two characters (or one actor playing both) shot on a dressed set with careful lighting and wardrobe, intercut with a Markdown diagnostic report and the GitHub repo page to pitch the product. (Note: the per-shot forensics and the agent's visual review overruled this; see shots #2–#9.)
  - 0.0–2.83 s · pain point · "Published 21 minutes ago and already judged 90% likely to flop?"
  - 2.83–5.83 s · diagnosis · "The cover order is reversed; the one with the hook is buried as image 3."
  - 5.83–10.54 s · how it works · "How does it know which one has a hook? It compared the top 9 in the same niche; every first image has a hook."
  - 10.54–15.7 s · solution and caveat · "It even gives you fixes? Three new titles, ready to go. The catch: you have to bring your own vision-capable model."
  - 15.7–19.3 s · call to action · "Fine. Next time a post gets no views, don't delete it, run it through this first."

## Real material in the frame

5 shots, 11.17 s (59%), contain real material (screen recordings, screenshots, photos, embedded video). A remake needs the original files; they cannot be replaced by generated output:

- Shot #1 (00:00.000): screen recording, region x0% y31% w100% h37%
- Shot #4 (00:03.875): screen recording, region x0% y29% w100% h41%
- Shot #6 (00:07.542): screen recording, region x0% y36% w100% h27%
- Shot #8 (00:11.542): screen recording, region x0% y35% w100% h28%
- Shot #8 (00:11.542): UI screenshot, partial frame, the white central area with an illustrated checklist and title-rewrite suggestions
- Shot #10 (00:16.125): UI screenshot, region x0% y22% w100% h69%
- Shot #10 (00:16.125): UI screenshot, partial frame, GitHub mobile repo page (JuneYaooo/social-account-doctor)

## Shot list

| # | Time | Duration | Underlying picture | Confidence | Material | Overlays | Camera | Remake route |
|---|---|---|---|---|---|---|---|---|
| 1 | 00:00.000 | 1.38 s | Code / template rendering | 92% | Mixed (real material): screen recording | — | Smooth move | R01-code-graphics → code_render_spec |
| 2 | 00:01.375 | 1.46 s | AI-generated video | 72% | AI-generated | Text, title | Static (subject moving) | R07-product-hero → capsule_cinema |
| 3 | 00:02.833 | 1.04 s | AI-generated video | 58% | AI-generated | Text, title | Static (subject moving) | R07-product-hero → capsule_cinema |
| 4 | 00:03.875 | 1.96 s | Code / template rendering | 91% | Mixed (real material): screen recording | — | Smooth move | R01-code-graphics → code_render_spec |
| 5 | 00:05.833 | 1.71 s | AI-generated video | 72% | AI-generated | Text, title | Static | R07-product-hero → capsule_cinema |
| 6 | 00:07.542 | 3.00 s | Code / template rendering | 99% | Mixed (real material): screen recording | — | Smooth move | R01-code-graphics → code_render_spec |
| 7 | 00:10.542 | 1.00 s | AI-generated video | 92% | AI-generated | Text, title | Static (subject moving) | R07-product-hero → capsule_cinema |
| 8 | 00:11.542 | 1.92 s | Code / template rendering | 59% | Mixed (real material): screen recording, UI screenshot | — | Smooth move | R01-code-graphics → code_render_spec |
| 9 | 00:13.458 | 2.67 s | AI-generated video | 66% | AI-generated | Captions, title | Smooth move | R07-product-hero → capsule_cinema |
| 10 | 00:16.125 | 2.92 s | Code / template rendering | 50% | Mixed (real material): UI screenshot | — | Smooth move | R01-code-graphics → code_render_spec |

## Shot details

### #1 00:00.000–00:01.375 · Code / template rendering; introduces real material (screen recording)

![shot_001](shots/shot_001.jpg)

- Probabilities: code / template rendering 91%, screen recording / screenshot 7%, 3D rendering 1% (fused)
- Material: mixed (real material): screen recording
  - Screen recording: 9 text items inside a light UI panel (9 of them small type)
- Likely tool: CapCut (JianYing)
- Rendered texture: flat color blocks 74%, 95% of pixels use only 23 colors, horizontal / vertical straight edges 31%
- Dense fine high-frequency detail (small text / UI) on a flat background, like a screenshot or document
- [VLM] Dark-gray solid background with large title typography at top and bottom
- [VLM] The centered card is a flat text box; text is sharp and undistorted
- [VLM] The central card has a horizontal slide motion
- Picture: a vertical feed-style layout video, dark background with title captions top and bottom, a white card in the center shows the post-diagnosis text and slides slightly left.
- Text track (in-frame text / bottom): 这条才发二十一分钟
- Text track (in-frame text / middle): 0.TL;DR一这条会爆吗？ / 不会。9成扑街概率。预计24h数据：5-15赞/1-3评/3-8收藏 / …
- Text track (in-frame text / middle): 核心诊断（三个一票否决项）： / 1.封面致命错位一第1张是无字插画（D实物展示型，0钩子），唯一带钩子的「设计师 / 了J大字图被埋到第3张。小红书CTR只看第1张→直接判死刑。 / …
- Text track (in-frame text / middle): 2.正文承诺没兑现一标题/desc都说「附提示词J，但3张图全是效果图+1句吐槽 / … / 3.冷启信号已经走弱一21分钟3赞1评2收藏，作者历史10条平均3.6赞 / …
- Remake: code rendering: rebuild from the text tracks and layout in Remotion / HTML, 100% controllable, no generative model needed; the user must supply the real material: screen recording (region x0% y31% w100% h37%)

### #2 00:01.375–00:02.833 · AI-generated video + code / editor overlays (text labels, title)

![shot_002](shots/shot_002.jpg)

- Probabilities: AI-generated video 72%, talking avatar 21%, 3D rendering 2% (fused)
- Material: AI-generated
- Likely tool: Hailuo
- Low-light frame: 36% of the area is near-black but grainy (not a solid fill), flat blocks in the highlights only 0%, and there is a face: the flat blocks / few colors come from crushed shadows, so they are not counted as rendering features
- Photo / video texture: noise σ=0.42, 123 colors, flat blocks 9%
- Per-frame sharpness oscillates periodically (period 2 frames, correlation 0.55): the picture went through frame interpolation / rate conversion; the original frame rate was lower
- Duplicate frames follow an interpolated pattern (0%, period None): source was 24/25 fps then transcoded, typical of AI video, 3D / animation renders or film footage
- 100% of sampled frames have a frontal face with a fixed camera, and the audio track has voice
- [VLM] The hand holding the phone looks slightly stiff and plasticky
- [VLM] Where the glasses frame meets the skin, the light and shading blend with a slight AI-render unnaturalness
- [VLM] The face and skin are extremely smooth with the typical AI-render sheen
- Picture: night interior, a short-haired woman with glasses sits cross-legged on a sofa hugging a cushion, holding a phone, talking to herself with a slightly surprised expression. Fixed medium shot.
- Framing / light: close-up; eye level; tripod / stand, fixed; night / low-light ambient, no dedicated fill; key frontal (from the camera direction, both sides of the subject equally bright); contrast about 1.2:1 (0.2 stops); soft light (almost no shadow edge)
- Text track (text / bottom): 就被判九成扑街？
- Text track (title / top): 笔记没人看，先别删
- Remake: the original is AI-generated: take the original shot's first and last frames for first/last-frame generation, which reproduces composition and motion closely; to swap the subject, edit the first frame with an image model first
  - First-frame image prompt: photorealistic, a young Chinese woman with short hair and light-green retro glasses, yellow hoodie, sitting cross-legged on a living-room sofa at night hugging a cushion, holding a phone in one hand, warm floor lamp and houseplants in the background, soft cinematic light, medium shot, eye level, vertical
  - Video prompt: static shot, the woman speaks with her mouth slightly open, slightly confused and surprised, eyes forward, fingers lightly holding the phone, small head movements, body stays seated and steady, soft night light

### #3 00:02.833–00:03.875 · AI-generated video + code / editor overlays (text labels, title)

![shot_003](shots/shot_003.jpg)

- Probabilities: AI-generated video 54%, live footage 38%, talking avatar 4% (fused)
- Material: AI-generated
- Low-light frame: 47% of the area is near-black but grainy (not a solid fill), flat blocks in the highlights only 0%, and there is a face: flat blocks / few colors come from crushed shadows, not counted as rendering features
- Photo / video texture: noise σ=0.38, 98 colors, flat blocks 8%
- 100% of sampled frames have a frontal face with a fixed camera, and the audio track has voice
- [VLM] Micro-expressions around the eyes, blinking and lip movement are natural and fluid, no AI facial warping or flicker
- [VLM] Finger joints, nail edges and the contact with the mug are physically and perspectivally correct
- [VLM] The bookshelf, lamp bokeh and depth-of-field falloff match a real optical lens
- [claude visual review] Same set and same character as #2, #5, #7, which were judged AI video
- [claude visual review] Skin is too smooth and evenly lit, bookshelf detail in the background smears together, none of the noise live footage would have
- Picture: a short-haired woman in a black newsboy cap and round glasses sits in an armchair in a study, holding a mug with both hands and speaking; warm desk lamp and bookshelf behind; fixed medium close-up.
- Framing / light: close-up; eye level; tripod / stand, fixed; night / low-light ambient, no dedicated fill; key front left; contrast about 2.0:1 (1.0 stop); medium softness
- Text track (text / bottom): 封面顺序反了
- Text track (title / top): 笔记没人看，先别删 / 笔记没人看，先别册
- Remake: the original is AI-generated: first/last-frame generation from the original shot; to swap the subject, edit the first frame with an image model first
  - First-frame image prompt: photorealistic, a young East Asian woman with short hair, black newsboy cap and thin round glasses, white shirt and dark-green vest, holding a light ceramic mug with both hands, seated in an off-white armchair, blurred night bookshelf and warm yellow desk lamp behind, cinematic, soft warm light, vertical composition
  - Video prompt: static medium close-up, a young woman in a newsboy cap and glasses holds a coffee mug and speaks naturally, blinks, expression shifts from slight thought to a soft smile, both hands steady on the mug, desk-lamp light steady behind, cinematic interior light

### #4 00:03.875–00:05.833 · Code / template rendering; introduces real material (screen recording)

![shot_004](shots/shot_004.jpg)

- Probabilities: code / template rendering 90%, screen recording / screenshot 8%, 3D rendering 1% (fused)
- Material: mixed (real material): screen recording
  - Screen recording: 10 text items inside a light UI panel (10 of them small type)
- Likely tool: CapCut (JianYing)
- Rendered texture: flat color blocks 60%, 95% of pixels use only 41 colors, straight edges 34%
- Dense fine high-frequency detail (small text / UI) on a flat background, like a screenshot or document
- [VLM] Vertical feed-style layout, dark-gray background with a centered card
- [VLM] The card shows an analysis of a Xiaohongshu image carousel with pan / zoom motion (pans left to reveal images 2 and 3)
- [VLM] Fixed typeset caption and title at top and bottom
- Picture: a vertical explainer animation; on a dark background a centered card analyses the carousel, panning and zooming from image 1 to images 2 and 3 and their hook copy.
- Text track (in-frame text / bottom): 带钩子那张埋在第三张
- Text track (in-frame text / middle): image2王炸 / 效果，设计师 / 又又又要失业了
- Text track (in-frame text / middle): BRASS / SIRMINGHAM / 第3张一唯一带钩子的「设计师又又又 / 要失业了」，被理
- Text track (in-frame text / middle): BRASS BIRMINGHAM,0字钩子 / 封面）一蒸汽朋克插画 / 第2张一双曝光精灵女王，美但仍无字 / …
- Remake: code rendering in Remotion / HTML from the text tracks and layout; the user must supply the real material: screen recording (region x0% y29% w100% h41%)

### #5 00:05.833–00:07.542 · AI-generated video + code / editor overlays (text labels, title)

![shot_005](shots/shot_005.jpg)

- Probabilities: AI-generated video 72%, talking avatar 21%, 3D rendering 2% (fused)
- Material: AI-generated
- Likely tool: Kling
- Low-light frame: 37% near-black but grainy, flat highlight blocks 0%, face present: not counted as rendering features
- Photo / video texture: noise σ=0.41, 119 colors, flat blocks 9%
- Per-frame sharpness oscillates periodically (period 4 frames, correlation 0.46): frame interpolation / rate conversion; the original frame rate was lower
- Duplicate frames follow an interpolated pattern (0%, period None): 24/25 fps source then transcoded, typical of AI video, 3D / animation renders or film footage
- 100% of sampled frames have a frontal face with a fixed camera, and the audio track has voice
- [VLM] Face and skin have the typical Midjourney / AI photoreal look, too smooth with a faint glow
- [VLM] Lip movement and eye motion during speech show the slight warping and floating of AI image-to-video
- [VLM] Small structures where the hand holding the phone meets the cushion look slightly melted and unstable
- Picture: a short-haired woman in blue-framed glasses sits cross-legged on a sofa in a dim living room, hugging a cushion and looking at her phone, then looks up in surprise and speaks; fixed shot, warm room light against cool phone light.
- Framing / light: close-up; eye level; tripod / stand, fixed; night / low-light ambient, no dedicated fill; key frontal; contrast about 1.3:1 (0.4 stops); soft light
- Text track (text / bottom): 它怎么知道哪张有钩子？
- Text track (title / top): 笔记没人看，先别删
- Remake: first/last-frame generation from the original shot; edit the first frame with an image model to swap the subject
  - First-frame image prompt: photorealistic, dim living room at night, a young Asian woman with short hair and light-blue thick-framed glasses, yellow hoodie, cross-legged on a sofa hugging an off-white cushion, holding a glowing smartphone, looking up in surprise with her mouth open, a warm floor lamp blurred behind, cinematic, warm/cool light contrast, 8k detail
  - Video prompt: static medium shot, a short-haired woman with glasses on a sofa at night looks down at a glowing phone, then looks up in surprise, mouth open, and speaks toward the camera, lively expression, stays cross-legged, warm floor lamp softly blurred behind, photorealistic cinematic look

### #6 00:07.542–00:10.542 · Code / template rendering; introduces real material (screen recording)

![shot_006](shots/shot_006.jpg)

- Probabilities: code / template rendering 98%, 3D rendering 0%, screen recording / screenshot 0% (fused)
- Material: mixed (real material): screen recording
  - Screen recording: 7 text items inside a light UI panel (7 small type)
- Likely tool: CapCut (JianYing) / After Effects
- Rendered texture: flat color blocks 79%, 95% of pixels use only 11 colors, straight edges 33%
- Dense fine high-frequency detail on a flat background, like a screenshot or document
- [VLM] Solid flat background
- [VLM] Sharp-edged geometric color block (white card)
- [VLM] Clean, lossless typeset text
- [VLM] Standard animation: the card scrolls up in a straight line at constant speed
- Picture: on a dark-gray background a white card with a list scrolls upward at constant speed, with fixed captions above and below.
- Text track (in-frame text / bottom): 它对比了同赛道前九名 / 第一张全有钩子
- Text track (in-frame text / middle): 6.共性总结一用户该抄什么 / top9都在做的事（你这条1个都没做）： / …
- Text track (in-frame text / middle): 1.封面有大字/强视觉钩子一9/9全做了，你0/3 / 2.明确身份锚定一设计师/老师/家长/品牌方/二次元一你的封面没有人群 / 3.反差/反认知一7/9（Altman直播/AI写试卷/Al仿官方海报...）一 / …
- Text track (in-frame text / middle): 4.具体可见的「原型对比」一5/9用截图/海报/试卷做「Alvs真品」对照，你纯 / 5.垂直号沉淀一7/9是AI/设计/工具垂类号，标签明确，你是「AI资讯日报」杂 / …
- Remake: code rendering in Remotion / HTML; the user must supply the real material: screen recording (region x0% y36% w100% h27%)

### #7 00:10.542–00:11.542 · AI-generated video + code / editor overlays (text labels, title)

![shot_007](shots/shot_007.jpg)

- Probabilities: AI-generated video 77%, talking avatar 15%, live footage 7% (fused)
- Material: AI-generated
- Likely tool: Kling
- Low-light frame: 38% near-black but grainy, flat highlight blocks 0%, face present: not counted as rendering features
- Photo / video texture: noise σ=0.39, 119 colors, flat blocks 10%
- 100% of sampled frames have a frontal face with a fixed camera, and the audio track has voice
- [VLM] The edge of the hand holding the phone has a slight AI smoothness
- [VLM] Face and lighting are realistic but carry the typical AI image look (MJ / Flux facial traits)
- [VLM] Lip movement has the micro-motion signature of AI image-to-video
- Picture: a young woman with glasses sits cross-legged on a sofa at night hugging a cushion and looking at her phone, opens her mouth in surprise and speaks. Fixed medium shot, warm room light mixed with cool phone light.
- Framing / light: close-up; eye level; tripod / stand, fixed; night / low-light ambient, no dedicated fill; key frontal; contrast about 1.1:1 (0.1 stops); soft light
- Text track (text / bottom): 连改法都给了？
- Text track (title / top): 笔记没人看，先别删
- Remake: first/last-frame generation from the original shot; edit the first frame with an image model to swap the subject
  - First-frame image prompt: photorealistic, cozy living room at night, a young East Asian woman with short hair and teal round glasses, yellow hoodie, cross-legged on a sofa hugging a beige cushion, holding a glowing smartphone, mouth slightly open in surprise looking at the screen, cool phone light on her face, warm floor lamp and a window behind, shallow depth of field
  - Video prompt: static shot, a woman with glasses hugs a cushion on the sofa and looks at her phone, mouth wide in surprise as she speaks, eyes on the screen, cool blue-white phone light flickers on her face, natural small movements

### #8 00:11.542–00:13.458 · Code / template rendering; introduces real material (screen recording, UI screenshot)

![shot_008](shots/shot_008.jpg)

- Probabilities: code / template rendering 59%, screen recording / screenshot 18%, 3D rendering 10% (fused)
- Material: mixed (real material): screen recording, UI screenshot
  - Screen recording: 10 text items inside a light UI panel (10 small type)
- Likely tool: CapCut (JianYing)
- Rendered texture: flat color blocks 78%, 95% of pixels use only 9 colors, straight edges 31%
- Dense fine high-frequency detail on a flat background, like a screenshot or document
- [VLM] The subject is a typeset document screenshot / card that scrolls or pans down slowly and smoothly
- [VLM] Text edges are clean and undistorted, not AI video
- [VLM] Fixed decorative title and subtitle at top and bottom; the central card has a slow displacement motion
- Picture: a card with a guide to rewriting the post on a dark background, title captions top and bottom, the content in the middle scrolls vertically.
- Text track (in-frame text / bottom): 三个新标题都写好了
- Text track (in-frame text / middle): P0一现在就改（5分钟内，决定生死） / 1.把第3张设成封面一编辑笔记→拖图重排序→第3张拉到第1位。预期CTR从估算2 / 8%，可能进二级池。 / …
- Text track (in-frame text / middle): 2.正文补提示词一至少给一段ChatGPTimage-2的prompt例子，兑现 / of British Industrial Revolution illustr / sepia tone..." / 3.改标题一套公式1（数字+人群+效果）重写：
- Text track (in-frame text / middle): o候选1：「3个提示词，设计师又要失业了|GPTimage-2实测J / o候选2:「我用GPTimage-2复刻了爆款桌游封面，附提示词] / o候选3:「GPTimage-2双曝光神技，设计师抄作业] / …
- Remake: code rendering in Remotion / HTML; the user must supply the real material: screen recording (region x0% y35% w100% h28%), UI screenshot (partial frame)

### #9 00:13.458–00:16.125 · AI-generated video + code / editor overlays (captions, title)

![shot_009](shots/shot_009.jpg)

- Probabilities: AI-generated video 66%, live footage 16%, talking avatar 9% (fused)
- Material: AI-generated
- Likely tool: HeyGen / D-ID / CapCut avatar
- Low-light frame: 46% near-black but grainy, flat highlight blocks 0%, face present: not counted as rendering features
- Photo / video texture: noise σ=0.40, 96 colors, flat blocks 8%
- [VLM] The body and the hand holding the mug are completely still
- [VLM] Only the head and mouth move, consistent with an audio-driven avatar
- [VLM] The background is completely static
- Picture: medium shot, a woman in a cap and glasses holds a mug and speaks, cozy study behind.
- Framing / light: close-up; eye level; tripod / stand, fixed; night / low-light ambient, no dedicated fill; key front left; contrast about 1.8:1 (0.9 stops); medium softness
- Text track (captions / bottom): 前提是你得自己配 / 一个会看图的大模型
- Text track (title / top): 笔记没人看，先别删 / 笔记没人看，先别册
- Remake: first/last-frame generation from the original shot; edit the first frame with an image model to swap the subject
  - First-frame image prompt: medium shot, young Asian woman, short hair, round glasses, black newsboy cap, white shirt, dark-green vest, seated on a sofa holding a white mug. Cozy study behind, warm desk lamp, blurred. Soft light from front left, photorealistic
  - Video prompt: image as the base, audio-driven avatar, lip sync, slight head movement and blinking, body and background static

### #10 00:16.125–00:19.042 · Code / template rendering; introduces real material (UI screenshot)

![shot_010](shots/shot_010.jpg)

- Probabilities: code / template rendering 50%, 3D rendering 25%, screen recording / screenshot 21% (fused)
- Material: mixed (real material): UI screenshot
  - UI screenshot: 30 text items inside a light UI panel (29 small type)
- Likely tool: CapCut (JianYing)
- Rendered texture: flat color blocks 73%, 95% of pixels use only 13 colors, straight edges 51%
- Dense fine high-frequency detail on a flat background, like a screenshot or document
- [VLM] The subject is a screenshot of a GitHub project's mobile web page (JuneYaooo/social-account-doctor)
- [VLM] The whole frame zooms in slowly and evenly over time, no internal interaction or cursor: a Ken Burns push on a static screenshot
- [VLM] Solid-color bands with a title and captions added in post at top and bottom
- Picture: a screenshot of a GitHub repo home page on mobile showing the file list; the camera pushes in slowly.
- Text track (in-frame text / bottom): .env.example / 2weeks ago / .gitignore / 2weeksago
- Text track (in-frame text / bottom): 口 / LICENSE / 5 monthsago / README.md
- Text track (in-frame text / bottom): 行，下回发了没人看 / 先别删，拿它查查
- Text track (in-frame text / middle): <>Code / Issues / Pullrequests / Actions
- Remake: code rendering in Remotion / HTML; the user must supply the real material: UI screenshot (region x0% y22% w100% h69%), UI screenshot (partial frame)

## Notes

- Deterministic forensics are reliable for "is it moving, is it a whole-image push, is the texture rendered, was there a 24→30 frame-rate conversion"; telling live footage, AI and 3D apart relies on the VLM and human review.
- Platform transcoding usually strips metadata; missing metadata does not mean there was none.