Arynwood Technology

Lorelei Noble
  • 0.0 (0)
  • Technology

What happens when you train an AI on your own artwork? I'm finding out by teaching a machine to recognize my personal art style and create new images in that style. This is the real-time story of that process - the experiments, the bugs, and what works.

Episodes

  1. Digital Hybridization Techniques

    1d ago

    Digital Hybridization Techniques

    Episode Description: Can you truly lock a character’s identity into an AI model without losing artistic soul? Join Lorelei Noble, lead researcher at Arynwood Technology, for a deep dive into the world of Digital Hybridization—the synthesis of manual human technique and local machine learning.In this episode, we pull back the curtain on the ArynCore MCP to explore how artists can reclaim creative sovereignty from cloud subscriptions. We break down the exact technical pipeline used to create the "Arynwood Artifact," a professional-grade character LoRA trained entirely on consumer-grade hardware.What’s inside this deep dive: The Power of Data Discipline: Why Stage 3 manual captioning beats "luck" every time. We discuss isolating identity from style to prevent "bleeding" and the importance of the "Aaronwood" trigger word [Conversation History].The Tagging Engine: An exploration of the WD14 Tagger and its various architectures—from the speed-critical ViT to the high-accuracy SwinV2—and how it serves as the foundation for your dataset metadata.SDXL Architecture Exposed: Why your characters look better in SDXL. We discuss the dual text encoder system (CLIP ViT-L and OpenCLIP BiG-G), the massive 2048-dimensional embeddings, and fine-grained 32-pixel resolution bucketing.The 12GB VRAM Survival Guide: Practical strategies for home training, including text encoder output caching—where we move heavy encoders to the CPU to free up space for the U-Net—and managing size conditioning embeddings for perfect composition.The BYOM Philosophy: Introducing the Design Studio, a community-focused, lightweight UI that allows you to "Bring Your Own Model" and plug it into a local-first creative ecosystem [Conversation History].Whether you’re a developer building agentic flows or an artist looking to master your own style, this episode provides the technical roadmap to move from "prompt engineer" to "model architect."Built on Open Source: All research and software discussed are released under Creative Commons 4.0 (CC BY 4.0) and MIT licenses to support the global creator community.

    39 min
  2. Distilling Your Artistic Identity into Local AI

    1d ago

    Distilling Your Artistic Identity into Local AI

    This podcast provides a comprehensive exploration of Lorelei Noble’s work at Arynwood Technology, focusing on the intersection of human artistry and local machine learning within the ArynCore MCP ecosystem. It details how creators can leverage open-source software and local hardware to avoid costly subscriptions while pushing the boundaries of "Digital Hybridization"—a process that blends manual brushstrokes from tools like Procreate with trained SDXL LoRAs. The discussion covers several key areas of your research and development: Training at Home on Limited GPUs: The episode breaks down the methodology for training LoRAs, highlighting specific configurations like Network Dimension 32 and Alpha 16 that ensure stable convergence and style capture without needing industrial-grade hardware.The "Arynwood Artifact" and Portability: A major focus is your core research finding: Ecosystem Portability. The podcast explains how a stylistic identity trained on one SDXL derivative (like Juggernaut XL) can successfully transfer to others (like Dream Shaper XL Turbo) without retraining, though it notes the architectural barriers when attempting to move these artifacts to legacy systems like SD 1.5.The ArynCore MCP Hub: Listeners will get a tour of your "local-first AI desktop hub," which unifies diverse tools into a single interface. This includes the Design Center for visual assembly and the Workflow Builder for chaining LLMs with video and audio tools.Creative Audio & Video Tools: The "fun" side of the project is highlighted through the Studio sidecars. This includes using Demucs v4 for high-quality stem separation (vocals, drums, bass), RVC v3 for vocal timbre conversion with pitch shifting, and SadTalker for generating talking-head videos.Open-Source Philosophy: Finally, the podcast underscores the community-driven mission of Arynwood, where all research artifacts and software pipelines are released under Creative Commons 4.0 (CC BY 4.0) and MIT licenses to foster collaboration and local resource utilization.

    44 min
  3. Technical Overview of LoRA Training

    2d ago

    Technical Overview of LoRA Training

    How do you convert your design aesthetic into an algorithm? Artist and developer Lorelei Noble is making machine learning accessible to artists by developing a local desktop application. Let's take a deep dive into the technical elements of her academic study, a review of the software pipeline she developed for this research, and take a philosophical moment to ponder the validity of AI art. The Research Artifact: The study is a formal, citable artifact archived on Zenodo and OpenAIRE with a dedicated DOI. It investigates whether a style trained on one SDXL derivative (like Juggernaut XL) functions on another (like Dream Shaper XL Turbo) without retraining.The Training Pipeline (Aryn Core MCP): A "local first" Python backend (FastAPI) that manages background training jobs using Kohya SS and tracks project states in SQLite.Dataset Configuration:Standardized Hyperparameters:The "Structural Gap": The research explicitly identifies that while portability is high within SDXL, it "architecturally can't transfer to SD 1.5," marking a clear generational divide in LoRA technology.Artistic Result: The "Arynwood" LoRA functions as an abstract color layer, allowing the base model to handle structural "roundness" while the LoRA overlays hand-drawn pencil strokes, high contrast, and the signature "dingy overlay" aesthetic

    21 min

  4. 5d ago

    Training AI with Procreate drawings

    Teaching AI to paint like me. A podcast about training LoRAs, building custom models, and discovering what my art looks like through a machine's eyes.

    21 min
  • 4 Episodes

About

What happens when you train an AI on your own artwork? I'm finding out by teaching a machine to recognize my personal art style and create new images in that style. This is the real-time story of that process - the experiments, the bugs, and what works.

Information

  • Creator
    Lorelei Noble
  • Years Active
    2K
  • Episodes
    4
  • Rating
    Clean
  • Copyright
    © Lorelei Noble
  • Show Website
    Arynwood Technology