ASI v3.2 – Deterministic LoRA Architect & Visual Governance System
System Manifest (Civitai Presentation)
ASI v3.2 is a deterministic visual governance framework designed to eliminate probabilistic entropy in diffusion models. Its primary objective is the eradication of Layer-Bleed (style contamination between subject and background) through strict domain isolation.
The system serves as an Enterprise Dataset Engineering Tool. It enables the generation of ultra-pure, consistent image sets required for training high-stability LoRAs. By mathematically decoupling Physics (Render) from Aesthetics (Style), ASI v3.2 creates "Software-Defined Styles" that achieve reproducibility previously only possible through fixed model weights.
The Multi-Agent Core (Operational Modules)
The architecture is managed by four specialized sub-systems (Agents) ensuring process isolation:
ASI_Discovery (The Eye): Deconstructs user input or uploaded images into atomic components. It identifies domain contamination (e.g., style attributes bleeding into anatomy) and sanitizes the input.
ASI_Architect (The Law): Governs the 18-Layer Enterprise Stack. It enforces hierarchical information placement and Logic-Separation, ensuring background vectors do not corrupt foreground geometry.
ASI_Engineer (The Scripter): Translates the architecture into executable syntax. It calculates deterministic weights (
_ASI-Weight-High/Low) and generates finalized 18-layer scripts for ComfyUI.ASI_Director (The Calibrator): Sets hardware-level render parameters including Temperature, Contrast, Gamma, and Saturation. It acts as the virtual cinematographer.
Module R (The Auditor): A closed-loop refinement system. It audits the render for "Drift" or "Corruption" and commands weight recalibration (±0.05) if the visual intent is not met.
The Index Matrix & Wildcard Logic
The 10 provided Index files constitute the system’s Genetic Library. They are not mere lists; they are mathematically weighted building blocks.
Hierarchical Taxonomy: Wildcards follow the strict
ASI-[Domain]-[Category]-[Descriptor]syntax to prevent cross-token association errors.Deterministic Weighting: Each wildcard contains internal parameters (e.g.,
1.35) to ensure material properties (like the refraction of jade or skin haptics) remain identical across different seeds.Logic-Index (Level 3): This index functions as a physical barrier, preventing the "scent" of a style from altering the structural physics of an object.
Pro-Tips for Advanced Users
1. Self-Expanding Library (Dynamic Growth) The Index system is modular and expandable. To add your own styles or subjects:
Upload an image to the LLM.
Command:
ASI_Discovery, deconstruct this image into layers and generate new ASI-compliant wildcards.The system will extract the visual DNA and format it for your personal library.
2. Optimal Prompt Engineering You do not need to know the index content by heart.
Provide your core intent (e.g., "A mechanical angel") and describe desired effects (e.g., "glowing internal cooling, oil-painting textures").
ASI will scan the Matrices, select the optimal combination of Level 1-9 descriptors, and synthesize a clean, high-detail, 18-layer prompt.
3. Version Control The package includes an independent sub-folder containing various versions of the AI Role. This allows users to choose the specific governance intensity that best fits their hardware or aesthetic requirements.
Workflow Protocol
Initialize: Load the ASI v3.2 Role and the 10 Index files. Trigger:
ASI, initialize v3.2 closed-loop.Intent: Define your target (Subject + Style).
Mapping: ASI assigns components to the 18-Layer Stack.
Execution: Use the generated script and Director-parameters in your render engine.
Audit: Use Module R for iterative refinement.