flo-fi

AI-generated 3D character brand for social media. Fully agentic pipeline — images, video, voice, lip sync — all CLI/API, zero GUI.

LoRA training is deferred. v1 and v2 both failed (see data/experiment_log.jsonl for post-mortems); v3 is on hold until we have ~100 diverse frames. Midjourney is the primary generator right now — see PROGRESS.md for current state. When resuming LoRA work, use the /lora-training skill.

Quick Start

./shared/scripts/start_comfyui.sh                              # ComfyUI at localhost:8188
./shared/scripts/mission_control.py generate-local --scene desert-sunset --seed 66603
./shared/scripts/mission_control.py generate-local --list-scenes  # 7 scene presets

What It Does

• Image generation — ComfyUI + RealCartoon-Pony V3 + custom Flo LoRA via mission_control.py
• Scene presets — 7 locked scenes with Flo's face block (--scene)
• LoRA training — SimpleTuner on Mac Mini, training data in tools/ComfyUI/training/flo/
• Video pipeline (Phase 3) — Grok Aurora → ElevenLabs → PixVerse → TikTok-ready video

Project Structure

shared/scripts/
  mission_control.py       # Generation CLI (--scene, --prompt, --lora, --model, etc.)
  start_comfyui.sh         # ComfyUI launcher
tools/ComfyUI/
  models/checkpoints/      # RealCartoon-Pony V3, Juggernaut-XL v9
  models/loras/            # flo_character_v1.safetensors (custom trained)
  training/flo/            # Curated training images (mj_, leo_, comfy_ prefixes)
  input/                   # Reference images
  output/                  # Generated images
config/
  config.json              # SimpleTuner training config
  multidatabackend.json    # Training data backend config
data/generation_logs/
  generations.jsonl        # Every generation logged with full params
docs/
  ROADMAP.md               # Phase progress and where we're headed
  STYLE_GUIDE.md           # Prompts, settings, anti-patterns
  COMFYUI_OPS.md           # Detailed ComfyUI operations
stories/                   # Story ideas and episode scripts
tutorials/                 # External tutorials (Daisy Studios workflow)

Manual Generation (ComfyUI Web UI)

For hands-on exploration without the CLI.

./shared/scripts/start_comfyui.sh    # leave running in a terminal tab

Open localhost:8188 → folder icon → Load → pick a workflow:

Template	What it does
shared/workflows/flo_basic.json	Flo's locked prompt + RealCartoon-Pony
shared/workflows/flo_with_reference.json	Same + IPAdapter face reference (v20_04)

Edit the prompt in the green CLIP nodes, change the seed, hit Queue Prompt.

Development

# Python venv (SimpleTuner, training tools)
source .venv/bin/activate

# LoRA training (runs on Mac Mini)
ssh eriks-mac-mini.local
cd /Users/eriksjaastad/Flo-Fi && source .venv/bin/activate && simpletuner train

Reference

Run pt info -p flo-fi for models, APIs, seeds, and infrastructure details.