AI-generated promotional art and videos for shelter pets.
My wife volunteers at a local pet adoption clinic, and they were having trouble getting pets noticed. I was also looking for a real-world project to test out an end-to-end AI stack — Ollama for local LLM inference, Replicate for FLUX LoRA training, and Flask for tying it together. This was the result: a pipeline that takes a pet's adoption listing + photos, trains a custom model, and generates stylized artwork of that specific animal. A companion video maker turns pet photos into promotional videos with music and text overlays.
Sample outputs from real shelter pets:
| Jackson | Honey | Yellow | Koda |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
Each pet's LoRA was trained on 5-20 real photos, then used with style LoRAs (Pixar, Ghibli, etc.) to generate these.
The video maker takes pet photos and creates MP4 videos with:
- Ken Burns zoom/pan animation effects
- AI-generated background music via Suno AI
- Text overlays (pet introduction + adoption info)
- Configurable video length and image ordering
Four steps, orchestrated through a Flask web UI:
1. GATHER DATA Paste the pet's adoption listing. Ollama (llama3) extracts
name, breed, age, personality traits, writes an adoption
story, and generates image prompts.
2. TRAIN A LORA Upload 5-20 photos of the pet. The app trains a FLUX LoRA
on Replicate so the AI learns what this specific pet looks like.
(~15-20 min on a GPU-T4)
3. GENERATE ART Using the trained LoRA + style LoRAs (Pixar, Ghibli, Lego,
hand-painted miniature), generates images of the pet in
various scenarios based on extracted personality traits.
4. CREATE MORE A separate page lets you generate additional images with custom
prompts, different styles, and aspect ratios.
Everything is tracked in a per-pet JSON config, and you get an email with the results when it's done.
- Create Pet LoRA - Paste the adoption description, upload photos, hit go. A live terminal shows progress via WebSocket.
- Pet Directory - Browse all processed pets, view their AI art and submission data.
- Make New Images - Pick a pet, write a custom prompt, choose a style LoRA (Pixar, Ghibli, etc.), and generate more images.
app.py Flask + SocketIO web server
0_run_all.py Pipeline orchestrator
1_gather_pet_data.py Ollama-powered data extraction & story generation
2_train_a_lora.py Replicate FLUX LoRA training
3_create_images_of_pet.py Automated image generation from trained model
4_create_additional_images.py On-demand image generation with custom prompts
utilities/
ollama_utils.py Local LLM management (install, serve, query)
replicate_utils.py Replicate API (model creation, training, inference)
gmail_utils.py Email notifications with image attachments
fileio_utils.py Temporary file hosting for training uploads
file_zip_utils.py Image packaging for upload
archive_utils.py Disk space management & output archiving
templates/
base.html Bootstrap 5 layout with SocketIO progress
create_lora.html Pet submission form with live progress terminal
create_images.html Custom image generation with style controls
view_pets.html Gallery view of processed pets
- Python 3.10+
- Ollama installed locally
- A Replicate account (for LoRA training + image generation)
- (Optional) Gmail app password for email notifications
git clone https://github.com/tillo13/pet-adoption-ai.git
cd pet-adoption-ai
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Pull the local LLM for data extraction
ollama pull llama3
# Configure your API keys
cp .env.example .env
# Edit .env with your Replicate API token and other credentials# Start the web UI
python app.py
# Opens at http://localhost:5001Or run the pipeline directly from the command line:
# Edit GLOBAL_VARIABLES.py with the pet description, then:
python 0_run_all.pyEdit GLOBAL_VARIABLES.py to customize:
| Setting | Default | Description |
|---|---|---|
STEPS |
1000 | LoRA training steps |
LORA_RANK |
16 | LoRA rank (higher = more capacity) |
RESOLUTION |
"512, 768, 1024" | Training resolutions |
NUMBER_OF_FACTS |
5 | Encouraging facts / image prompts to generate |
MODE |
"DEVELOPMENT" | Set to "PRODUCTION" for public models + HuggingFace push |
EMAIL_ON_COMPLETION |
True | Send email when generation finishes |
Running on Replicate's GPU-T4 hardware:
- LoRA training: ~$1-2 per pet (15-20 min at ~$0.00055/sec)
- Image generation: ~$0.01-0.03 per image
Processing 7 pets with ~130 total generated images cost roughly $10-15.
This project processed 7 real shelter animals:
| Pet | Type | Images Generated |
|---|---|---|
| Jackson | Dog | 6 |
| Marina | Dog | 9 |
| Yellow | Cat | 35 |
| Koda | Cat | 10 |
| Mister White | Cat | 8 |
| Maybelline | Cat | 39 |
| Honey | Dog | 25 |
The video_maker/ directory contains a separate pipeline for creating promotional pet videos from photos.
Five sequential steps:
1. PREPARE IMAGES Converts photos to JPEG, corrects orientation, resizes to
1280x720 with blurred background fill for portrait images.
2. CREATE MOVIE Generates an MP4 from the prepared images with configurable
duration per image. Optionally generates background music
via Suno AI, or uses a provided MP3 file.
3. APPLY ZOOMPAN Adds Ken Burns-style zoom and pan animation effects to
make the slideshow feel cinematic.
4. ADD TEXT Overlays introduction text (first 5 seconds) and adoption
info text (last 5 seconds) with styled fonts and shadows.
5. CLEAN UP Removes temporary files and intermediate video segments.
video_maker/
run_all.py Pipeline orchestrator (runs steps 1-5 in order)
1_prepare_images.py Image conversion, EXIF orientation, resize
2_create_movie.py FFmpeg video creation + Suno AI music integration
3_apply_zoompan.py Ken Burns zoom/pan effects via FFmpeg
4_add_text.py Text overlays with FFmpeg drawtext filters
5_clean_up.py Temporary file cleanup
GLOBAL_VARIABLES.py Song prompt, text overlays, video settings
testsuno.py Suno AI API connectivity test
- Python 3.8+
- FFmpeg installed and on PATH
- Pillow (
pip install Pillow) - (Optional) Suno AI account for AI-generated music
- (Optional) Local suno-api npm server for music generation
cd video_maker
# 1. Put pet photos in initial_images/
# 2. Edit GLOBAL_VARIABLES.py with pet name, adoption info, and song settings
# 3. Run the pipeline:
python run_all.py
# Output video will be in created_videos/Edit video_maker/GLOBAL_VARIABLES.py:
| Setting | Description |
|---|---|
SONG_PROMPT |
Text prompt for Suno AI music generation |
SONG_TO_USE |
Path to an existing MP3 (leave empty to auto-generate) |
FIRST_5_SECOND_TEXT |
Text shown at the start of the video |
LAST_5_SECONDS_TEXT |
Text shown at the end (e.g., adoption clinic address) |
DEFAULT_VIDEO_LENGTH |
Target video length in seconds |
randomize_images |
Shuffle image order (True/False) |
AI Art Pipeline:
- Local LLM: Ollama + llama3 for text extraction and prompt generation
- Image training: Replicate API for FLUX LoRA fine-tuning
- Image generation: Replicate API with trained LoRA + style LoRAs
- Web UI: Flask + Flask-SocketIO + Bootstrap 5
- File hosting: file.io for ephemeral training image uploads
- Notifications: Gmail SMTP with optional GCP Secret Manager
Video Maker:
- Video processing: FFmpeg (video creation, zoom/pan, text overlays)
- Image handling: Pillow (orientation correction, resize, blur backgrounds)
- Music: Suno AI API for background music generation
- Orchestration: Sequential Python scripts via subprocess
MIT