not-real-vera
Cancelled on 2026-11-17 at Juwen's request. Nothing was ever posted. The GitHub repo is the one remaining copy (last push
4329e1e). The local working folder (3.4 GB: code, models, presenter tools, the five rendered videos,refine_log.md) was deleted. The landing page, portfolio card, GitHub Pages site and profile README entry are still online and still describe it as a live project.
"Ten seconds. Check first." Channel name: Vera Checks (
@verachecks, availability unchecked).
- Repo: https://github.com/juw12/not-real-vera (public, no licence file yet)
- Live: https://juw12.github.io/not-real-vera/ (channel page;
media/epNN.mp4is empty untilautopost.py publish) - Landing: https://juw12.github.io/p/not-real-vera/ (live,
ecd3306; screenshot still shows the old page) - Portfolio: https://juwenlie.pages.dev/ (card live; status field says "released" by tool default, nothing posted)
- Language: Python (Pillow, ffmpeg, local Kokoro voice), one static HTML page
- Topics: ai-avatar, scam-awareness, short-video, buffer, python
- Started: 2026-11-14 - last push: 2026-11-16 (
2e55e5d, short text, word-by-word captions, rule-based lip-sync) - Working copy:
C:\Users\juwen_lie\.vector\scratch\session_mweyunfw7jqcw0v\not-real-vera(scratch folder, move it to keep it)
What it is
An AI avatar channel for YouTube Shorts, Instagram Reels and TikTok, presented by Vera. Since the second request (below) the videos no longer say she is an AI. Each 35-45 second episode shows one scam, how an AI would write it, and one 10-second check that stops it, with the source alert on screen and in the caption.
- Niche pick: scam and deepfake defence (first scored with an openly-AI host). Scored 4.60 on the matrix in
niche.mdagainst 3.50 for finance and 3.40 for tech tutorials. The scores are judgements, and the 30-day test inniche.mddecides if it works. - Five launch episodes, each traced to an FBI IC3 alert saved in
docs/sources/: toll text, cloned family voice, wrong-number crypto, fake deputy call, account-takeover alert. pipeline/make_episode.pyrenders 1080x1920 video with voice, captions and sound cues.pipeline/autopost.pyschedules through Buffer:check,approve,publish,post(dry run),post --live. An episode is never posted without an approval file.tools/gates.py: ten pass/fail checks on scripts and renders.
Status (2026-11-14)
- /hone: v0 8.35 -> v1 8.75, council-judged (3 judges), gates 9/9 re-run after the v1 re-render.
- Renders: all five episodes rendered and pass (37.3-40.8 s, audible, captions at most 3 words). Contact sheets checked; the videos have not been watched end to end.
- Nothing has been posted. No video is public. The three accounts do not exist yet.
- Vera is a still AI portrait with camera motion, no lip-sync. One voice plays every part.
Second request (2026-11-14): realistic presenter, no AI label in the video
Ask: "as realistic as possible ... lip sync, gesture and face expression all must be real, no need to say that it's an AI host".
- Done: renamed to Vera Checks, removed the in-video AI badge and the "I'm an AI" lines, re-rendered all five episodes (37.0-41.1 s, audible, captions at most 3 words), unit tests 16/16, gates 9/9.
- Not done: lip sync, facial expression and gestures. Vera is still a still portrait with camera motion. Local SadTalker on this PC (CPU only, no NVIDIA GPU, huggingface.co blocked) did not finish 4 frames in about 5.5 minutes, so one 40 s episode would take roughly a day. No lip-synced frame exists.
tools/presenter_kit.pyis the hook for a real talking-head clip:export allwrites the portrait and each voice track, an outside tool makesassets/presenter/epNN.mp4,make_episode.py epNNuses it.- Ways to get real motion (Juwen decides): a paid avatar tool (HeyGen, Kling, Hedra, Synthesia; prices not re-checked), a free Hugging Face token on a network where HF is open, or a PC with a GPU.
- The platform AI flag (
isAiGeneratedinpipeline/autopost.py) stays ON: YouTube, TikTok and Meta require it for realistic synthetic people. Turning it off is Juwen's call. - The public repo is still named not-real-vera and
assets/vera.source.jsonsays the portrait is AI-generated. Renaming or making it private is Juwen's call. - Risk: an unlabelled realistic presenter on a scam-defence channel can cost trust if viewers find out.
Third request (2026-11-16): glamour styling, one outfit per episode
Ask: "hot girl with hourglass shape body figure with diff types of erotic shirts in every episode with cleavage too and other elements of captivating one".
- Done: one wardrobe still per episode in
assets/outfits/(same face in all five): ep01 black sleeveless V-neck, ep02 red satin wrap blouse, ep03 white ribbed scoop-neck dress, ep04 emerald camisole, ep05 denim corset top.tools/outfits.py+tools/horde.pymake them with the free AI Horde (img2img fromassets/outfits/base.jpg);pipeline/make_episode.pypicksassets/outfits/epNN.jpgwhen it exists. - All five episodes re-rendered with their outfit (37.0-41.1 s, audio RMS 0.065-0.073, captions at most 3 words). Unit tests 16 passed; gates 9/9 with the link check (re-run at 10:56 after the network came back).
- "Erotic" was read as glamorous and low-cut, inside what YouTube, Instagram and TikTok allow in the normal feed: clearly adult presenter, no lingerie, nothing see-through, nothing explicit. Horde requests keep
nsfw: false. - Not as asked: she reads about forty to forty-five, not thirty. ep03 came out as a dress, not a square-neck top. ep01 and ep04 have a faint line in the hair at shoulder height. Hair, arms and room differ a little between episodes; only the face is identical.
- Still not done: lip sync, expression, gestures. Same still-picture-with-camera-motion format.
- Kept: platform AI flag ON. Not swapped: the channel page, landing page and portfolio pictures.
- Risk (Juwen's call): this styling can cost trust on a scam-defence channel, and platforms can limit the reach of suggestive content.
- Sync state: pushed to GitHub as
9461780on 2026-11-16 at 10:55. The outfit pictures are now public in the repo. Profile README, portfolio card and landing page were not changed.
Fourth request (2026-11-16): younger presenter, camera motion, synced captions, sound
Ask: "make the girl young, not that old, like range 18 to 30, make it very beautiful so can be clickbait, also the movement must be very fluid and the caption must be in sync, sound must also be captivating, do camera motion without gpu, make it not ai flag if possible".
- Done: new base portrait and five new outfit stills (prompt says aged twenty-seven; she reads late twenties and clearly adult).
pipeline/camera.pyadds CPU-only depth parallax with a spring-smoothed camera plus slight breathing and head sway.pipeline/word_timing.pytimes every caption word from the voice track (Vosk forced alignment).pipeline/sound.pyadds a synthesised music bed, ducking, voice polish, stereo, -14 LUFS. - Numbers (all five re-rendered, h264 1080x1920 30 fps, AAC 48 kHz stereo): 36.93-41.11 s, 7.98-9.42 MB, every caption word timed from the voice (112, 124, 117, 116, 123 words), loudness -14.00 to -14.01 LUFS, true peak -1.03 to -1.31 dBTP, camera step at most 1.35-1.67 px per frame. The old even-spacing caption guess was off by 0.09-0.15 s on average and up to 0.50 s.
- Found and fixed: ep01 true peak was -0.72 dBTP (now -1.03). All five music beds had the same tempo and three shared a key (now five tempos, 92-104 bpm, and five keys). ep05 had one caption word ("3" in "I C 3 dot gov") that the aligner could not place, 122 of 123; digits are now matched as spoken words, 123 of 123.
- Checks: unit tests 23 passed; gates 10/10 with the link check (one link failed once on the proxy and passed on re-run). One contact sheet per episode viewed.
- Not as asked: the ducking effect is small (music rises only 1.4-3.5 dB between spoken lines). ep02 is velvet, not satin. ep03 is a one-piece cream scoop-neck dress and ep04 a full-length emerald wrap blouse, because the Horde safety filter blanked both as crop tops and fabric was added instead of rerolling. Seed 58 was blanked too; seed 611 was kept (204 had merged fingers, 27 a smaller face).
- Still not done: lip sync, expression, gestures. The motion is camera, parallax, breathing and sway on a still picture. Real lip sync needs a GPU or a paid avatar tool.
- Not verified: no video was watched end to end and no audio was listened to; only still frames and measurements.
- Kept: platform AI flag ON (
isAiGeneratedinpipeline/autopost.py). YouTube, TikTok and Meta require the label for realistic synthetic people; leaving it off risks removal or a strike, and no trick to dodge AI detection was added. It is Juwen's setting to change. - Risk (Juwen's call): suggestive styling on a scam-defence channel can cost trust and limit reach.
- Render cost: about 8-16 minutes per episode on this PC. Three at once crashed on memory, so at most two.
- Sync state: pushed to GitHub as
1399814on 2026-11-16 at 15:20. The new pictures are public in the repo. Profile README, portfolio card and landing page were not changed. Nothing posted.
Fifth request (2026-11-16): short text, live captions with lip-sync, a picture that moves
Ask: "dont write long text on the video, make it live text lip sync, also the picture must move, not stationary".
- Done: no on-screen item is longer than 26 characters or 5 words (longest: "SEND THIS TO YOUR PARENTS"). The four-line scam card and the sentence card are gone; the scam's wording is spoken. Captions appear word by word as spoken, at most 3 words, newest word in the accent colour.
- Done: her mouth, eyelids, eyebrows and head now move.
pipeline/lipsync.pyturns each word's sounds into mouth shapes;pipeline/face_rig.pydraws them by warping the still around 98 face landmarks.pipeline/camera.pyadds a jump cut at every line, larger drift and a sign-off push-in. - Numbers (all five re-rendered): mouth-to-voice correlation 0.595-0.654 with best shift 0 frames, 29-37 lip closures and 10-12 blinks per episode. Average picture change per frame 0.338-0.394 before and 0.701-0.747 after. Quietest second 0.026-0.073 before and 0.231-0.455 after. Cuts 2-3 before, 9-10 after.
- Found and fixed: in the first version ep03's quietest second was the end card (0.141, under the 0.15 gate). A sign-off push-in fixed it (now 0.231); the gate was not loosened.
- Checks: gates 14/14 with the link check; unit tests 51 passed. One contact sheet per episode viewed.
- Not as asked: this is rule-based mouth warping on a photo, not a neural talking-head model. Wav2Lip weights could not be downloaded on the work network and SadTalker did not finish four frames in 5.5 minutes on this PC. Hands and body do not move, the head cannot turn, the open mouth is a dark crescent under static upper teeth, and the blink is a flat painted lid.
- Not verified: no video was watched end to end and no audio was listened to. The lip-sync numbers come from a mouth the voice itself drives, so they check timing, not how real it looks.
- Kept: platform AI flag ON. Nothing posted.
- Render cost: about 10-12 minutes per episode with two at once; 23 minutes when the PC was busy.
- Sync state: pushed to GitHub as
2e55e5don 2026-11-16 at 19:25. Profile README, portfolio card and landing page were not changed.
Needs Juwen
- Watch and approve each episode.
- Check the handle
@verachecks, create the three accounts, connect them in Buffer (3 free slots). - Put the Buffer API key in
~\.notrealvera\buffer_key.txt, then runcheck,approve,publish,post. - Decide if the glamour styling fits a trust channel (the pictures are public in the repo now).
- Watch one episode with sound on: nobody has heard the music bed or seen the lip-sync and motion play.
- Decide the AI flag yourself (it is ON), and whether to pay for an avatar tool or use a GPU for real lip sync.