popopo-99/zy-cinematic-realism

Turn simple scene ideas into restrained, physically believable cinematic AIGC prompts.

What it solves

This project provides a structured cinematic visual workflow for AI image generation. It solves the problem of generic, "advertisement-like" AI images by shifting the focus from simple keyword lists to a professional filmmaking approach. Instead of relying on vague "cinematic" tags, it forces the user to define the narrative moment, camera positioning, lighting, and composition before translating those decisions into model-specific prompts.

How it works

The system operates as a "Skill" for LLMs (like ChatGPT or Codex), using a multi-stage pipeline:

  1. Scene Master: Establishes a single source of truth for the scene's logic (characters, action, location, lighting).
  2. Creative Grammar: Applies professional cinematography and directing styles (including a library of 38 directors' visual signatures) to determine the visual execution.
  3. Model Compiler: Translates the Scene Master and grammar into native prompts optimized for specific models (e.g., Midjourney V8.2, GPT Image 2, Seedream 5.0 Pro, Nano Banana).
  4. Result Repair: Uses a "Prompt Doctor" to diagnose and fix specific visual failures without rewriting the entire scene.

Who it’s for

Digital artists, AI filmmakers, and prompt engineers who want precise, narrative-driven control over their AI-generated imagery and want to avoid the "AI look" in favor of authentic cinematic realism.

Highlights

  • Model-Agnostic Translation: Convert a single visual plan into native prompts for multiple different image models.
  • Director Signatures: A system that decomposes a director's style into four axes: lighting/contrast, color/exposure, lens/camera, and composition/spatial signatures.
  • Continuity Bible: Tools to lock characters, wardrobe, and props across multiple shots to maintain visual consistency.
  • Prompt Doctor: A targeted repair system that modifies only specific variables (like camera angle) while preserving the rest of the scene.
  • Task-Based Routing: A Model Router that recommends the best model path based on the specific task (e.g., character design vs. prop modification).

Related

  • Project
    bentoml/BentoDiffusion

    A collection of example projects for self-hosting and deploying various Stable Diffusion family models using the BentoML framework.

  • Project
    cure-lab/MMA-Diffusion

    A framework for evaluating the security of Text-to-Image models by using multi-modal adversarial attacks to bypass prompt filters and safety checkers to generate NSFW content.

  • Project
    Shilin-LU/MACE

    MACE is a finetuning framework for diffusion models that enables the mass erasure of up to 100 unwanted concepts, such as celebrities or explicit content, while preserving unrelated model knowledge.

  • Project
    wzj177/ecommerce-image-suite

    An AI-powered toolkit for e-commerce sellers to automatically analyze product photos and generate a complete set of professional marketing images across various styles and platforms.