ThreadGens

ThreadGens is a local Java pipeline for turning Reddit-style threads or X-style posts/replies into 9:16 short-form video assets.

The production path now integrates all three quality layers:

• P0 originality/integrity — five content-aware formats, deterministic + semantic novelty, source-clean fictional social rendering, timed visual states, motion, and transitions.
• P1 variation/provenance — timed captions, voice pools/delivery presets, persistent identity recency, and AI/manual provenance/disclosure metadata.
• P2 pre-publish audit — final finished-video repetition analysis across approved output history with PASS/WARN/BLOCK gating.

The P2 gate evaluates actual content and presentation repetition. It does not add classifier-evasion noise, spoof platform metadata, or fabricate human upload behavior.

Production flow

Prompt / manual script
        ↓
P0 content + semantic novelty
        ↓
Reddit/X rendering + P1 identity rotation
        ↓
P1 TTS voice planning + narration sidecars
        ↓
P1 captions + timed visual scenes
        ↓
Dynamic final MP4 + provenance metadata
        ↓
P2 finished-output fingerprint
        ↓
PASS / WARN / BLOCK
        ↓
Approved publish history

P0 generation history, P1 identity history, and P2 approved-publish history are deliberately separate.

Windows setup

Run:

setup_windows.bat

Setup checks/installs Java 21, FFmpeg/FFprobe, Kokoro, Piper, Ollama, and the local models used by the production pipeline:

llama3.1:8b
nomic-embed-text

Then run:

run_ai_windows.bat

Batch generation:

batch_create_videos_windows.bat

The batch launcher rotates all five content formats by target slot, rotates the four high-end Kokoro voices (af_heart, af_bella, af_nicole, bf_emma) by video, and offsets the next format from recent approved publish history.

Build

javac -d out src/redditTxtToImg/*.java

or:

gradle build

Production entrypoints

Recommended compatibility-friendly CLI:

java -cp out redditTxtToImg.CheckedRunner data/comments.txt output --platform reddit

Direct final production entrypoint:

java -cp out redditTxtToImg.P2Entrypoint data/comments.txt output --platform reddit

Normal CheckedRunner, the GUI compatibility path, Runner, OpImageVideoSafeRunner, Gradle application, runnable JAR, and Windows batch flow route through P2Entrypoint.

P2Entrypoint invokes the complete P0/P1 pipeline first and only audits after finished video/provenance output exists.

P0-only orchestration remains available for deliberate development/debug runs:

java -cp out redditTxtToImg.P0Entrypoint data/comments.txt output --platform reddit

Explicit raw renderer/debug path:

java -cp out redditTxtToImg.RawCheckedRunner data/comments.txt output --platform reddit

Direct renderer classes remain available for development:

java -cp out redditTxtToImg.RedditScreenshotGenerator data/comments.txt output
java -cp out redditTxtToImg.XThreadGenerator data/comments.txt output

Raw rendering bypasses P0/P1/P2 orchestration, but renderer-source integrity still prevents fabricated engagement and X verification.

Supported platforms:

• reddit
• x
• twitter as an alias for x

Local AI generation

Reddit example:

java -cp out redditTxtToImg.CheckedRunner --platform reddit --auto \
  --post-title "Finish this story in the comments" \
  --topic "I found a locked box behind my dryer." \
  --count 10 \
  --llm-model llama3.1:8b

X example:

java -cp out redditTxtToImg.CheckedRunner --platform x --auto \
  --post-title "Wrong answers only" \
  --topic "Why is there a shopping cart in my living room?" \
  --count 10

For X, --post-title is a hidden reply-style instruction and --topic is the visible original post.

P0 formats and novelty

Recommended:

--format auto

Explicit formats:

thread_story
confession
debate
best_answers
escalating_conversation

Each format also has four narrative substyles. Use --format-variant auto (recommended) to rotate them from generation history, or pass a concrete substyle such as witness_chain, private_note, skeptical_qa, editor_picks, or calm_to_weird.

Long batch controls:

• -FormatSelection per-slot keeps replacement attempts on the target slot's format.
• -FormatSeries thread_story,confession,debate,best_answers,escalating_conversation controls the rotation pool.
• -FormatVariant auto rotates substyles within each format.
• -PublishHistoryFile data\publish_history.jsonl seeds the cooldown from approved output.

Accepted generated scripts are stored in:

data/generation_history.jsonl

P0 performs deterministic duplicate/structure checks followed by local Ollama embedding comparison for substantially reworded versions of the same premise.

Useful controls:

• --history-file PATH
• --history-limit N
• --novelty-threshold N
• --novelty-retries N
• --embedding-model MODEL
• --semantic-threshold 0.86
• --semantic-history-limit N
• --no-novelty
• --no-semantic-novelty

Malformed generation history fails closed when strict novelty validation is active.

P1 captions and visual timing

P1 builds timing from the exact narration sidecar and measured WAV duration.

Caption modes:

--captions off
--captions word
--captions sentence

word mode uses ASS karaoke highlighting. sentence uses timed sentence cues. off removes visible captions while sentence timing can still drive visual scenes.

Useful controls:

• --caption-words N
• --visual-max-scenes N

Timing is deterministic and audio-duration-aligned, but it is estimated rather than phoneme-level forced alignment.

P1 voice variation

Voice selection modes:

single
series
per-slide

Example:

java -cp out redditTxtToImg.CheckedRunner data/comments.txt output \
  --platform reddit \
  --tts kokoro \
  --voice-series "af_heart,af_bella,am_adam" \
  --voice-selection series \
  --series-id example-series \
  --tts-delivery calm \
  --captions word \
  --video --concat-video

Useful controls:

• --voice NAME_OR_PATH
• --voice-series VOICE1,VOICE2
• --voice-selection single|series|per-slide
• --series-id ID
• --tts-delivery natural|calm|energetic|dramatic
• --tts-speed 0.60..1.60
• --tts-language CODE
• --tts-sentence-pause-ms 0..2000

Every successful narration segment writes:

<segment>.txt
<segment>.voice.json

The P2 audit reads these .voice.json files to fingerprint the voice(s) actually rendered, rather than trusting only the launch-time --voice argument.

P1 identity history

Persistent identity history:

data/identity_history.jsonl

P1 rotates names/profile images by recency and avoids same-video duplication where the pools permit it. AI/custom profile images are preferred when available.

The integrated implementation uses a same-process mutex, OS file lock, and atomic replacement. Malformed nonblank identity-history rows fail closed instead of being silently ignored.

Controls:

• --identity-history-file PATH
• --identity-history-limit N
• --no-identity-history

P1 provenance and disclosure

Successful runs can write:

<output>/metadata/<prefix>-provenance.json
<final-video>.provenance.json

The manifest records content origin, platform/format/model, novelty state, voice strategy and actual segment voices, caption timing method, image-generation configuration, synthetic-identity/integrity policies, and SHA-256 artifact hashes.

MP4 files receive standard disclosure tags such as comment, description, artist, and encoded_by.

This is generation provenance/disclosure metadata; it is not a cryptographically signed C2PA credential.

Controls:

• --metadata-dir PATH
• --disclosure TEXT
• --no-provenance-metadata

P2 pre-publish audit

Approved finished-output history:

data/publish_history.jsonl

P2 fingerprints and compares:

• exact script hash;
• lexical content similarity;
• semantic premise similarity;
• final video/clip byte fingerprint;
• rendered social-frame perceptual hashes;
• sampled frames from the completed MP4;
• avatar/author identity-region perceptual hashes;
• actual P1-rendered voice-set signature and TTS engine;
• selected content format;
• normalized narration-segment pacing;
• stable P1 caption/delivery/provenance configuration.

P2 outputs:

PASS
WARN
BLOCK

Default production behavior is block mode. A BLOCK writes the audit report but returns failure before a caller can promote the video as publish-ready.

Default P2 settings:

publishAuditEnabled=true
publishAuditMode=block
publishHistoryFile=data/publish_history.jsonl
publishHistoryLimit=100
publishAuditWarnThreshold=58
publishAuditBlockThreshold=78

P2 controls:

• --publish-audit
• --no-publish-audit
• --publish-history PATH
• --publish-history-limit N
• --publish-audit-warn N
• --publish-audit-threshold N
• --publish-audit-mode block|warn
• --audit-report PATH
• --publish-metadata PATH

The history read, semantic comparison, decision, and approved-history append are one serialized transaction using an in-process fair lock plus an OS file lock.

Malformed publish history fails closed.

Dynamic video

Video generation requires FFmpeg and FFprobe.

java -cp out redditTxtToImg.CheckedRunner --platform x --auto \
  --topic "Why is there a shopping cart in my living room?" \
  --count 10 \
  --tts kokoro \
  --voice af_heart \
  --video --concat-video

The video pipeline uses narration-aligned visual states rather than looping one unchanged screenshot. P0/P1 scene timing, continuous motion, captions, and format-specific transitions are combined into conventional H.264/AAC/yuv420p output with +faststart.

OP image generation

OP image generation is optional.

Local ComfyUI:

--image-mode comfyui

Existing local image:

--image-mode local --op-image path/to/image.png

Default ComfyUI endpoint:

http://127.0.0.1:8188

Default checkpoint:

RealVisXL_V5.0_fp32.safetensors

Batch behavior

Batch jobs execute through OpImageVideoSafeRunner, which routes through P2. The default Windows launcher serializes Ollama generation while completed generations render in parallel.

Each successful video enters approved publish history before the next job, so later videos in the same batch are compared against earlier approved videos.

The batch script copies a final MP4 into final_videos only after the Java command exits successfully. A P2 BLOCK therefore prevents promotion while retaining the job render and audit report for diagnosis.

Default runtime files

output/script/generated_comments.txt
output/*.png
output/audio/*.wav
output/audio/*.txt
output/audio/*.voice.json
output/video/*.mp4
output/video/final.mp4
output/video/final.mp4.provenance.json
output/video/publish_audit.json
<output>/metadata/<prefix>-provenance.json

data/generation_history.jsonl
data/identity_history.jsonl
data/publish_history.jsonl

History/runtime files are git-ignored.

Validation

The integrated GitHub Actions workflow validates on Ubuntu:

• complete Java compilation;
• P0 deterministic novelty/history tests;
• real P0 FFmpeg/ffprobe video tests for all five formats;
• P1 caption, voice, identity and provenance unit tests;
• P1 FFmpeg/libass caption/provenance integration;
• profile-image regressions;
• strict identity-history corruption handling;
• stable P1→P2 provenance signatures;
• P2 deterministic scoring/history tests;
• P2 semantic hard-block behavior;
• in-process and cross-process publish-history locking;
• real-media visual/identity/audio fingerprinting;
• actual P1 voice-sidecar ingestion;
• a full production PASS then duplicate-BLOCK scenario;
• Reddit/X production, raw compatibility, OP-image and failure-path regressions.

Windows validates:

• complete Java compilation;
• P1 runtime smoke tests;
• P2 runtime smoke tests;
• integrated P1/P2 boundary tests;
• profile-image regression tests;
• PowerShell launcher parsing.

Detailed architecture:

• P0_IMPLEMENTATION.md
• P1_IMPLEMENTATION.md
• P2_IMPLEMENTATION.md
• INTEGRATION_STATUS.md

Runnable JAR

gradle jar
java -jar build/libs/ThreadGens-0.7.0-p0-p1-p2-integrated.jar data/comments.txt output