LoRA Black Stories - Image Generation

About the project

LoRA adapter training project to replicate the characteristic visual aesthetic of "Black Stories" and generate new stories/scenes with visual consistency. The project includes meticulous dataset curation, autocaptioning system to strengthen style, trigger word definition for precise style control through prompts, and multi-resolution training (512-1024) to improve generalization. It includes a rigorous aesthetic QA process reviewing composition, silhouettes, contrast, contours, and coherence between samples to ensure stable and reproducible aesthetics. Trained on Black Forest Labs FLUX models (flux-schnell and flux-dev).

Technologies

Features

• LoRA adapter training on FLUX (Black Forest Labs)
• Base models: black-forest-labs/flux-schnell and flux-dev
• Curated dataset with autocaptioning for consistency
• Trigger word system ("TOK") for prompt-based control
• Multi-resolution training (512-1024px)
• Aesthetic QA: composition, silhouettes, contrast, and contours
• Inference pipeline on managed services (Replicate)
• Integration with ComfyUI/Automatic1111 for testing
• Conditionable and reproducible generation
• Stable visual style without depending on original assets

Technical challenges

• Dataset curation representative of target style
• Development of coherent autocaptioning system
• Balance between style fidelity and generative capability
• Trigger word optimization for precise control
• Exhaustive QA of visual consistency between generations
• Generalization to different resolutions and compositions

Learnings

• LoRA adapter architecture and training on FLUX
• Diffusion model fine-tuning techniques (Black Forest Labs)
• Dataset curation for visual style control
• Image generation pipeline integration
• Qualitative evaluation of aesthetic consistency
• Deployment on Replicate for managed inference

Screenshots

Vista 1

Vista 2