Mike Boord

Step 1 — The dataset: 12 reference photos, each with its own caption file

Every image.jpg ships with a matching image.txt. The caption always leads with the trigger token "Prim the cat" so the model binds the look to that phrase, then describes the rest of the frame so it doesn't bake the couch or the floor into the character. Three of the twelve, verbatim:

Prim_1.txt

Prim the cat, a long-haired brown-and-white cat, is sitting next to a small soccer goal net inside a room. Prim's dark brown fur covers the face, ears, and tail, with white fur on the chest and front legs. The setting is indoors on a wood floor, with some crates and furniture in the background.

Prim_2.txt

Prim the cat, a fluffy brown and white cat, is being held by a person in a blue football jersey. The cat has a mix of brown and white fur, with darker brown patches on the face and ears. In the background, there are household appliances and electronics, giving the scene a casual, homey feel.

Prim_6.txt

Prim the cat is sleeping inside an open suitcase, nestled among clothes and belongings. Prim's brown and white fur spills over the edges of the suitcase, with Prim lying sprawled out in total relaxation. The scene conveys comfort and casual disarray.

Step 2 — Training: the LoRA learns the character

Step 3 — Text-to-image: trigger word + any prompt

Prim the cat on stage at a sold-out arena, holding a microphone, dramatic purple concert lighting, crowd of thousands

Prim the cat wearing a sombrero, playing an acoustic guitar on a sunlit Mexican plaza, warm golden hour

Prim the cat sitting behind a late-night talk-show desk, New York skyline window behind, studio lighting

Prim the cat in an astronaut harness floating in orbit above Earth, hard sunlight, lens flare

More range — same LoRA, eight different worlds

Animated tests — LoRA-driven character into motion

Step 4 — Two LoRAs at once: Hook LoRAs for a true two-shot

A single LoRA gives you one character. The moment a scene needs two specific trained characters in the same frame, stock Flux falls apart — both faces drift toward whichever LoRA loaded last. Hook LoRAs solve it: each character LoRA is wrapped in a Create Hook LoRA node and bound to its own Set CLIP Hooks conditioning branch with a region mask, so two trained identities hold in one render.

Before After

Before After

Before After

Before After

Drag the gold sliders on any property to wipe between the before photo and the AI-rendered after. No re-shoots, no manual compositing — the LoRA hallucinates the brand's flake coating onto the existing floor geometry.

Delivered to Garage Force · May 2026

8 of the 11 properties delivered in the May 2026 batch — mixed landscape / vertical / square aspect ratios for cross-platform marketing. Audio enabled on every clip.

Step 3: Compare multiple LoRAs on one render — Flux XY-Plot LoRA Testing · FLUX_XY-PLOT_LORA_TESTING.json — pick every LoRA checkpoint I want to evaluate (different characters, different step counts, different brand variants), pin one prompt, and the graph renders them all into a single grid image so the differences are obvious side-by-side instead of buried in a folder of separate outputs. This is how I pick the winning checkpoint before delivery.

Example images: MyAIForce — Mastering LoRA Testing in ComfyUI (used to illustrate the technique; my FLUX_XY-PLOT_LORA_TESTING.json applies the same idea to Flux with the GGUF / FP8 model paths).

Step 2a: What you get — AI-Toolkit web UI Screenshot: Ostris

The actual Toolkit interface — a Next.js dashboard at http://<pod>:8675 showing live training progress, GPU telemetry, dataset management, and checkpoint downloads. AI-Toolkit by Ostris; my contribution is the install pipeline below.

#!/usr/bin/env bash
# AI-Toolkit 1-Click Install or Restart for RunPod V2

# Fast path: existing install → start the UI and exit
if [[ -d "$REPO_DIR" && -f "$REPO_DIR/venv/bin/activate" ]]; then
  source "$REPO_DIR/venv/bin/activate"
  source "$NVM_DIR/nvm.sh"; nvm use 22
  cd "$REPO_DIR/ui" && npm run start
  exit 0
fi

# 0. GPU detection → right Torch wheel (Blackwell sm_120 vs older)
echo "Which GPU is this pod running?"
echo "  [1] RTX-5000-series (Blackwell, sm_120) or newer"
echo "  [2] Anything older (Ada, Hopper, Ampere, etc.)"
read -rp "Enter 1 or 2: " GPU_CHOICE
case "$GPU_CHOICE" in
  1) CUDA_STREAM="cu128" ;;
  2) CUDA_STREAM="cu126" ;;
esac

# 1. System packages, 2. shallow git clone w/ retry,
# 3. venv + Torch (right CUDA stream), 4. uv pip install
#    with pip+constraints fallback if uv fails,
# 5. nvm + Node 22, 6. npm build + start UI

# Total time: fresh pod → UI live in <3 minutes