joshwolk/jwen
joshwolk/jwen is a 3.1 billion parameter Qwen 2.5 3B model, fully fine-tuned and preference-tuned by joshwolk, specifically designed to generate Apple Notes-style face drawings. It predicts pen tool, color, position, and timing token by token, producing drawings stroke by stroke in a natural order. This model excels at generating unique, doodle-like facial illustrations, offering a distinct approach to image generation through sequential stroke prediction.
Loading preview...
Overview
joshwolk/jwen is a specialized 3.1 billion parameter model based on Qwen 2.5 3B, developed by joshwolk. It is uniquely fine-tuned to generate drawings in the style of Apple Notes doodles, specifically focusing on faces. The model predicts drawing elements like pen tool, color, position, and timing token by token, allowing it to render images stroke by stroke, mimicking human drawing order. This approach enables the creation of distinct, doodle-like facial illustrations.
Key Capabilities
- Stroke-by-Stroke Generation: Generates drawings sequentially, predicting each pen stroke's attributes (tool, color, position, timing).
- Apple Notes Style: Produces illustrations that resemble doodles made in Apple Notes, including various pen tools (pen, pencil, marker, crayon, watercolor, fountain pen).
- Grammar-Constrained Sampling: Utilizes a grammar-constrained sampler to ensure valid and coherent stroke sequences, preventing malformed outputs.
- Preference-Tuned: Enhanced through preference tuning on 660 human verdicts, refining its drawing style.
- Customizable Output: Allows control over character groups, color palettes, ink types, and temperature settings for varied drawing styles.
Good for
- Generating Unique Facial Doodles: Ideal for creating stylized, doodle-like face drawings with a distinct aesthetic.
- Exploring Sequential Image Generation: Useful for researchers and developers interested in token-by-token image synthesis based on drawing strokes.
- Creative Applications: Suitable for projects requiring programmatically generated, hand-drawn-style illustrations, particularly for faces.
Limitations
- Specialized Scope: Primarily draws faces and little else, reflecting one person's drawing style.
- Feature Placement: While plausible, feature placement (eyes, nose, mouth) can be loose rather than precisely accurate.
- PencilKit Rendering: The true visual fidelity of the ink tools is best observed through Apple's PencilKit renderer, as preview renders are stand-ins.