Flux3D Dataset Pipeline / Pipeline Dataset Flux3D

Français	English
Application Windows native pour préparer et contrôler un dataset d'images d'assets compatible Flux/Trellis2. Elle analyse les images, score la qualité, vérifie l'aptitude Trellis2, génère captions/prompts/tags et produit des rapports exploitables.	Native Windows application for preparing and reviewing an image asset dataset compatible with Flux/Trellis2. It analyzes images, scores quality, checks Trellis2 suitability, generates captions/prompts/tags, and writes usable reports.
Ce n'est pas une application web. L'interface est Tkinter, lancée localement. Le watermark produit est UnicorSoundEngine.	This is not a web app. The UI is Tkinter and runs locally. The product watermark is UnicorSoundEngine.

Features / Fonctionnalités

Français	English
Analyse qualité: résolution, netteté, bruit, exposition, contraste.	Quality analysis: resolution, sharpness, noise, exposure, contrast.
Évaluation Trellis2: fond, lisibilité sujet, proportions, continuité surface.	Trellis2 evaluation: background, subject readability, proportions, surface continuity.
Passes modèles locales: YOLO, SmolVLM2, Ministral 3B vision.	Local model passes: YOLO, SmolVLM2, Ministral 3B vision.
Verdicts opérateur: OK, À REVOIR, REJETÉ.	Operator verdicts: OK, REVIEW, REJECT.
Exports JSON, CSV, Markdown et sidecars texte.	JSON, CSV, Markdown, and text sidecar exports.
Correction Flux Klein locale optionnelle, sans dépendance obligatoire à ComfyUI.	Optional local Flux Klein correction, without requiring ComfyUI.

Public Repository Status / Statut dépôt public

Français	English
Cible actuellement supportée: pipeline Python + application Windows native Tkinter.	Current supported target: Python pipeline + native Tkinter Windows app.
backend/ fournit une API FastAPI locale minimale reliée au même batch runner, avec inférence modèle désactivée par défaut.	backend/ provides a minimal local FastAPI API wired to the same batch runner, with model inference disabled by default.
Le backend n'est pas une promesse cloud, frontend ou web app; l'usage principal reste desktop.	The backend is not a cloud, frontend, or web-app promise; desktop remains the primary workflow.
Les poids de modèles, bases SQLite, fichiers .env et sorties générées restent locaux et ne sont pas distribués.	Model weights, SQLite databases, .env files, and generated outputs stay local and are not distributed.
Licence publique: MIT.	Public license: MIT.

Quick Setup / Installation rapide

Core/offline setup without GPU or model weights:

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade pip
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -m unittest discover -s tests

Optional desktop model runtimes:

.\.venv\Scripts\python.exe -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
.\.venv\Scripts\python.exe -m pip install -r requirements-models.txt
.\.venv\Scripts\python.exe scripts\setup_desktop_models.py

Run / Lancer

Source launch:

.\.venv\Scripts\python.exe scripts\run_desktop_app.py

Windows launcher:

.\scripts\run_desktop_app.bat

Package build:

.\.venv\Scripts\python.exe scripts\package_desktop_app.py --check-only
.\scripts\package_desktop_app.bat
.\.venv\Scripts\python.exe scripts\package_desktop_app.py --verify-dist

scripts\package_desktop_app.py is the canonical build source. Flux3DDesktop.spec is a generated, ignored PyInstaller artifact and must not be treated as source.

Packaged executable:

$env:FLUX3D_PROJECT_ROOT = '<repo-root>'
.\dist\Flux3DDesktop\Flux3DDesktop.exe

Optional local API:

.\.venv\Scripts\python.exe -m pip install -r backend\requirements.txt
.\.venv\Scripts\python.exe -m uvicorn backend.main:app --host 127.0.0.1 --port 8000

The API is local-first. POST /api/batches/{id}/start runs the same batch pipeline as the desktop workflow and does not call Ollama, Trellis servers, or cloud services by default.

Dependency Files / Fichiers de dépendances

File	Use
requirements.txt	Core CPU/offline pipeline and public CI; no GPU or model weights.
backend/requirements.txt	Optional local FastAPI API tests and runtime.
requirements-models.txt	Optional heavy desktop model runtimes, packaging, and Flux Klein correction; not installed by public CI.

pyproject.toml is intentionally deferred for now. Model licenses and upstream restrictions are tracked in docs/model_assets.md.

Runtime Configuration

config.yaml is the CLI/pipeline source of truth for:

• input.supported_formats and input.max_file_size_mb
• output.report_json, output.report_csv, output.report_md, output.captioned_dir
• scoring.quality_weight, scoring.trellis2_weight, scoring.verdict.accept, scoring.verdict.review
• captioning.llm_provider, captioning.max_tokens, captioning.temperature

CLI values override config only for --input, --output, --config, --no-recursive, --no-captioned, --quality-weight, and --trellis2-weight. captioning.llm_provider currently supports kimi only; missing API keys still fall back to the offline mock captioner.

Default Models / Modèles par défaut

Role	Model	Local path
Detector	YOLO11n	.models/yolo/yolo11n.pt
Vision scorer	SmolVLM2 500M Video Instruct	.models/hf-local/HuggingFaceTB--SmolVLM2-500M-Video-Instruct
Caption verifier	Ministral 3B Instruct vision	.models/hf-local/mistralai--Ministral-3-3B-Instruct-2512
Optional correction	Flux Klein 4B local folder	.models/flux/FLUX.2-klein-4B

Validation / Vérification

.\.venv\Scripts\python.exe scripts\test_desktop_models.py
.\.venv\Scripts\python.exe -m unittest discover -s tests
.\.venv\Scripts\python.exe -m unittest discover -s tests -p "test_backend*.py"

Expected current baseline:

3 model roles ready
77 tests OK

Documentation

Full bilingual documentation:

docs/flux3d_desktop_pipeline.md

Model asset policy:

docs/model_assets.md

Public-release roadmap:

docs/public_release_roadmap.md

Watermark reference:

docs/assets/unicorsoundengine-watermark.jpg

The complete documentation covers installation, operator workflow, model orchestration, memory unloading, prompt guardrails, output formats, Flux Klein correction, packaging, reproducibility, troubleshooting, and file policy.