|
|
--- |
|
|
language: en |
|
|
tags: |
|
|
- stable-diffusion |
|
|
- sdxl |
|
|
- lora |
|
|
- image-generation |
|
|
- style-transfer |
|
|
- sketch |
|
|
license: creativeml-openrail-m |
|
|
--- |
|
|
|
|
|
# Model Card for SDXL Sketch-Style LoRA |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
|
|
|
This model is a **LoRA fine-tuning module** for **SDXL-based anime and art-style image generation models**. |
|
|
It transforms polished renders into **sketch-like / draft illustration** variants by emphasizing linework, rough shading, and reduced color saturation while preserving composition and subject structure. |
|
|
|
|
|
- **Developed by:** WorthyHuman1 |
|
|
- **Model type:** LoRA (Low-Rank Adaptation) for SDXL |
|
|
- **License:** CreativeML Open RAIL-M |
|
|
- **Finetuned from model:** SDXL 1.0–compatible anime/art checkpoints |
|
|
|
|
|
|
|
|
--- |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
|
|
|
Load this LoRA into any **SDXL-compatible pipeline** that supports LoRA weights to produce sketch-style outputs from text prompts. |
|
|
|
|
|
Common use cases: |
|
|
- Generating draft/concept-art sketches |
|
|
- Exploring stylistic variations |
|
|
- Educational demonstrations of style fine-tuning |
|
|
|
|
|
### Downstream Use |
|
|
|
|
|
Can be combined with: |
|
|
- Other LoRAs (character, pose, lighting) |
|
|
- Custom SDXL pipelines and workflows |
|
|
|
|
|
### Out-of-Scope Use |
|
|
|
|
|
- Photorealistic image generation |
|
|
- High-fidelity full-color illustration |
|
|
- Deceptive or misleading content generation |
|
|
|
|
|
--- |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
- Output quality depends on base model and prompt phrasing. |
|
|
- Strong LoRA weights may oversimplify textures or reduce color fidelity. |
|
|
- Style reflects patterns in the training data and may not generalize uniformly across all art styles. |
|
|
|
|
|
### Recommendations |
|
|
|
|
|
- Experiment with LoRA strengths between `0.6` and `1.0`. |
|
|
- Use sketch-related keywords: `sketch`, `rough lineart`, `pencil drawing`, `unfinished illustration`. |
|
|
- Combine with an appropriate SDXL base model for best results. |
|
|
|
|
|
--- |
|
|
|
|
|
## How to Get Started |
|
|
|
|
|
**Suggested prompt keywords:** |
|
|
sketch_style, rough lineart, pencil drawing, unfinished illustration |
|
|
|
|
|
**Recommended LoRA strength:** `0.6 – 1.0` |
|
|
|
|
|
Compatible with common SDXL tools (Diffusers, AUTOMATIC1111, ComfyUI, etc.). |
|
|
|
|
|
--- |
|
|
|
|
|
## Training Details |
|
|
|
|
|
### Training Data |
|
|
|
|
|
Trained on a curated dataset of sketch-style and draft illustration images, filtered to emphasize line-based structure and reduced rendering polish. Training focused on style transformation rather than subject memorization. |
|
|
|
|
|
### Training Procedure |
|
|
|
|
|
Parameter-efficient fine-tuning using LoRA techniques on SDXL-compatible layers. |
|
|
|
|
|
#### Training Hyperparameters |
|
|
|
|
|
- **Precision / regime:** fp16 mixed precision |
|
|
|
|
|
#### Compute |
|
|
|
|
|
- **Hardware:** NVIDIA GeForce RTX 4050 |
|
|
- **Approximate training time:** 16 hours |
|
|
|
|
|
--- |
|
|
|
|
|
## Evaluation |
|
|
|
|
|
### Testing Data, Factors & Metrics |
|
|
|
|
|
Evaluation performed qualitatively across multiple SDXL base models and diverse prompts covering characters, portraits, and scenes. Assessment prioritized visual coherence, stylistic consistency, and prompt adherence via side-by-side comparisons. |
|
|
|
|
|
### Results |
|
|
|
|
|
Produces consistent sketch-like outputs while preserving core composition across a variety of prompts when paired with suitable base models. |
|
|
|
|
|
--- |
|
|
|
|
|
## Environmental Impact |
|
|
|
|
|
- **Hardware:** NVIDIA GeForce RTX 4050 |
|
|
- **Hours used:** ~16 hours |
|
|
- **Compute region:** local / on-prem |
|
|
|
|
|
(Estimate only) |
|
|
|
|
|
--- |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Model Architecture and Objective |
|
|
|
|
|
- Architecture: SDXL LoRA |
|
|
- Objective: Style adaptation toward sketch/draft aesthetics |
|
|
|
|
|
### Software |
|
|
|
|
|
- SDXL-compatible inference/training tools |
|
|
- LoRA training framework (PyTorch-based) |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Card Authors |
|
|
|
|
|
- WorthyHuman1 |
|
|
|
|
|
## Model Card Contact |
|
|
|
|
|
- Hugging Face: `@WorthyHuman1` |
|
|
|
