From script to an illustrated reel, at no cost

🎨 Illustrated, no lip sync

A large illustrated avatar of Nei at the bottom (≈ 43 % of the height), in 11 poses tied to the emotion of the speech (startled, unsure, pointing, explaining, thumbs-up…), one hero illustration per scene, and a robot mascot. No one needs to record anything.

🎙️ Nei’s voice, local

Voice generated with local TTS (inemavox chatterbox), at 160–190 words per minute. Local Whisper marks the timing of each word for captions and scene triggers.

✅ Automatic QA

Measures duration, speech pace, volume (−14 LUFS), visual changes per minute, proof within 20 s, options planted before the prompt, CTA without a delivery promise, final card, and silence before the maxim. It produces a qa.md and a contact.jpg.

1. A hook in frame 0, of one of these types: contrary (tells people to stop doing something everyone does), loss (what if X disappeared?), fear statistic with a source, or a third-party story with a number. Not allowed: lists ("5 things…"), "did you know…", "I did…".
2. Frame 0 headline of up to ~6 words, in uppercase, with one word in yellow, more forceful than the speech. Draw an object with visible damage (scratched in red, broken); the avatar reacts in the first second, and something changes every ~1 s in the first three seconds. Frame 0 serves as the cover.
3. Proof within 20 s: a real screenshot of the source page, the number in the headline, and the fonte: line below.
4. The viewer’s problem, with reassurance ("it’s not your fault; no one gave you the map"). The final question is planted here: the options appear in yellow.
5. One topic per video, in 2–5 points with a visible index and a pill bar at the top; one object per sentence, changing every 2–3 s.

6. Every number has a written source on screen (fonte: …).
7. An honest caveat, spoken aloud, on an alert card (red = problem, green = correct).
8. A maxim in antithesis, after 0.8 s of real silence (voice and soundtrack stop).
9. An opinion CTA: Nei states his side and asks a strong question that can be answered with one word ("comment PAGO or LOCAL"). The card shows the question and options; the comment box types one option, posts it, and confirms it. It stays for ~10 s, with no sign-off. Never promise to deliver anything ("I’ll send it to you," "I’ll send you"): no one replies to comments. The link and source stay in the pinned comment.

Opinion prompt

The question at the end, with two one-word options (in the 1st video, PAGO or LOCAL). The tension appears in frame 0, the options are planted early, and only those (and the headline’s hot word) are highlighted in yellow.

Maxim

The summary line at the end, in antithesis ("Being the owner of your system was never about knowing every line. It’s about knowing what holds what in place."). It comes after a short silence.

Stage

The motor/palco.html page that draws each scene. The renderer takes a snapshot of it for every frame.

Music bed

Quiet music under the voice that gets quieter when the voice speaks and stops during the silence before the maxim.

LUFS

A measure of perceived loudness. −14 LUFS is the volume used as a standard by social platforms.

CTA

"Call to action": the video’s final request. Here it is always a single opinion question answered with one word.

TTS and lip sync

TTS is a voice generated from text. Lip sync is mouth movement matching the speech; rtexjp does not do that, the avatar only changes poses.

Whisper and Playwright

Whisper transcribes the voice and times each word. Playwright opens the stage in a headless browser and takes a snapshot of each frame.

Typography flash

A quick screen with only large text between points, so the video doesn’t feel static.

Block types

tese = claim (the hook, the problem); autoridade = proof with a source; virada = the index that introduces the points; metodo = each point; planta is just the id of the block that plants the options for the final question.

Beat

A section of a block that changes the headline, illustration, and pose on the word marked in em. It makes something change every 2–3 s.

Hero illustration

The large illustrated object in the middle of the screen (broken terminal, map, padlock), taken from the motor/assets/objetos/ library.

1. Label in monospace (≥ 30 px) and, for the points, pills 01–03.
2. Headline in Sora ExtraBold, centered, 2–3 lines, ≥ 96 px, almost white, with one word in light steel blue (yellow only in frame 0).
3. Hero illustration or real screenshot, large, between the headline and avatar.
4. Karaoke captions in their own strip, with the spoken word highlighted.
5. Large avatar, at the bottom and centered, in the beat’s pose; the face stays above 80 % of the frame height.

The background has a spotlight and light grain and alternates between navy and ice-blue paper ("tema": "claro" in the scene), with navy text. During the silence before the maxim, the last frame freezes with the avatar looking thoughtful: the screen is never blank.

Visual quality gate: in two samples per block, QA measures the headline (≥ 96 px), all text (≥ 30 px), occupied area (≥ 55 % of the usable frame), text contrast (≥ 4.5:1), and avatar height (≥ 38 %). Applies to scripts with "versao": "rtexjp/3".

Script and validation

checar checks the structure using the tipo field: it starts with tese, has a virada before the first metodo (2–5 points), a ressalva before the maxima, and the maxima immediately before the final cta. The 2nd example uses tese (hook) → autoridade (proof) → tese (problem + options) → virada → metodo ×3 → ressalva → maxima → cta.

Voice and scene

voz generates one audio file per block, speeds it up to 1.15×, and joins them; cena combines the script with the timing of each word.

Render, mix, and QA

render snapshots the stage frame by frame; mix adds the soundtrack (which stops during the maxim’s silence); qa measures everything.

Voice: inemavox + chatterbox environment

Clone the inematds/inemavox project to ~/projetos/inemavox (the engine calls its tts_direct.py and whisper_daemon.py). Chatterbox TTS and Whisper large-v3 run in a separate Python environment (on Nei’s machine, the conda env chatterbox at ~/miniconda3/envs/chatterbox); if yours is elsewhere, point the variables below to it. If you clone it somewhere else, export RTEXJP_INEMAVOX with the path. Whisper large-v3 uses about 10 GB of GPU memory (NVIDIA), plus chatterbox. Tested on Linux.

git clone https://github.com/inematds/inemavox ~/projetos/inemavox
cd ~/projetos/inemavox && python3 -m venv venv
./venv/bin/pip install -r requirements.txt   # o rtexjp chama venv/bin/python tts_direct.py
# chatterbox + faster-whisper num ambiente próprio (ver o README do inemavox)
export CHATTERBOX_PYTHON=/caminho/do/env/bin/python3   # TTS
export CS_WHISPER_PY=/caminho/do/env/bin/python3       # Whisper

Image: flux2-klein

The inematds/inemaimg server with the flux2-klein model (default http://localhost:8000/generate, or the address in RTEXJP_IMG_URL), plus the avatar reference photo (RTEXJP_AVATAR_REF). Needed only for optional steps 4 and 5 (poses and scene illustrations); Nei’s poses and the illustration library are already included.

Node, ffmpeg, Python

Node 18 or newer, ffmpeg (includes ffprobe), and Python 3 with Pillow and numpy.

git clone https://github.com/inematds/rtexjp
cd rtexjp
npm i
npx playwright install chromium
pip install --user pillow numpy   # ou dentro de um venv

Fields in each block

tipo (order above) · fala (spelled out, no digits) · fala_tts (optional, spelling to help TTS pronounce it correctly) · tela (large text and highlighted word) · cena.comp + cena.params (what the stage draws; params.gatilhos ties a scene change to a word in the speech) · avatar pose · destaque_fala (colored caption words) · fonte (id in fontes) · rotulo and n (point number). Only in maxima: silencio_antes_s (0.5–0.9). Only in cta: hold_s, how long the card stays after the last word (5–9 s; with the speech, the whole card lasts ~10 s). At the top level: gancho.tipo (contrario, perda, estatistica, historia) and gancho.objeto; isca with tipo: "opiniao", pergunta, opcoes, and lado.

Stage components (cena.comp)

Current format: gancho (frame 0 with damaged object), contador (supports decimais), duas (problem + prompt options), indice, mapa, explica, impacto, ressalva, flash, opiniao (question card). From the 1st edition, still on the stage: titulo, frase, pasta, repo, carro, equipe, cta (legacy). The robot mascot is drawn by the stage itself, beside the avatar, and reacts to the scene (question mark, alert, check mark). Poses: bracos, queixo, apontando, segurando, surpreso, mostrando. Each component’s parameters are in the examples edicoes/2026-10-28-china-fecha/roteiro.json and edicoes/2026-10-28-paga-200/roteiro.json.