The goal of this project is to train a generative neural network on photographs of my dog, Yogi, to transform him into a standalone character. The resulting photo series places this character within a unified narrative—the story of a single journey, told through a sequence of generated frames.
Example source images
All photographs were taken by me. When selecting the photos, I strove to maintain a diversity of angles, poses, environmental variations, and compositional choices. As a result, I compiled a dataset of 39 photographs, each sized 1080×1080 px.
Education
For training, Stable Diffusion XL was utilized with the DreamBooth LoRA method.The dataset comprised 39 original photographs of my dog. All images were cropped to a square format. Automatic textual descriptions were generated for the photographs using BLIP, after which a unified character identifier—TOK dog—was added to each.The model was trained on Stable Diffusion XL 1.0 for 500 steps at a resolution of 512 × 512 px. The result was a standalone LoRA file containing the learned characteristics of the character.The LoRA training and initial test generations were performed in Google Colab. Due to GPU limitations, the final stages of generation and series refinement were moved to Kaggle, where the already trained model pytorch_lora_weights.safetensors was used.
Although perfectly replicating the animal’s appearance from the original for the model proved challenging, overall it successfully retained the character’s key identifiable features: coat color, muzzle shape, ears, proportions, and general aesthetic. Furthermore, the dog was placed in scenarios and locations not present in the original dataset.Visually, the series is unified by the aesthetic of amateur analog photography: soft natural light, muted tones, and the feel of candid shots taken during a journey. While the different images vary in composition, lighting, the character’s emotional state, and frame scale, they maintain a consistent style and narrative continuity.
Final Images
Example Prompt: «small cute happy TOK dog on the back seat of a car, looking out the window during a road trip to the station, analog film photography, 35 mm film, candid documentary style, natural light, slightly faded colors, grainy, cinematic mood»
Prompt Example: «small cute happy TOK dog running along a narrow street in a small town, surrounded by low houses and quiet facades, analog film photography, 35 mm film, candid documentary style, natural daylight, slightly faded colors, film grain, cinematic travel atmosphere»
Prompt example: «small cute sleepy TOK dog curled up asleep on a neatly made bed in a stylish comfortable hotel room, photographed from a little farther away, calm and peaceful end of the journey, analog film photography, 35 mm film, candid photo, soft warm light, slightly faded colors, film grain, cinematic atmosphere»
Conclusion
The result demonstrated that the characteristic traits of a specific character—appearance, coloring, proportions, and overall visual impression—can be successfully transposed into new scenes and narrative situations. The trained neural network does not literally copy the original photographs; rather, it retains the dog’s recognizability while placing it in various stages of a fictional journey. Thus, the project became an exploration of how a generative model can transform a real, yet complex, subject into a distinct visual character and, based on that character, create entirely new storylines.
Software and Tools Used
Google Colab — environment setup, DreamBooth LoRA training, and initial test generations.
Kaggle Notebooks — final generation of the series after model training.
Stable Diffusion XL — the base generative model.
Hugging Face — model uploading and weight management.
ChatGPT — project concept development, storyline creation, and prompt refinement