| --- |
| license: mit |
| datasets: |
| - SubMaroon/danbooru-lineart |
| base_model: |
| - cagliostrolab/animagine-xl-3.0 |
| --- |
| |
| # Experimental ControlNet (Low Quality / Research Prototype) |
|
|
| > **Experimental model. Low quality. Not intended for production use.** |
| > This ControlNet was trained as a research experiment to explore line-based conditioning and colorization behavior in SDXL anime models. |
|
|
| --- |
|
|
| ## Model Summary |
|
|
| This repository contains an **experimental ControlNet for SDXL**, trained on anime-style images. |
| The model is **not stable**, shows **inconsistent color behavior**, and should be treated as a **research prototype** rather than a finished or polished solution. |
|
|
| The goal of this experiment was to understand: |
| - How SDXL ControlNet learns **colorization from line-based conditioning** |
| - How different conditioning types (Canny vs Lineart) affect **color consistency** |
|
|
| --- |
|
|
| ## Base Model |
|
|
| - **Base model:** `cagliostrolab/animagine-xl-3.0` |
| - **Architecture:** ControlNet SDXL |
| - **Training framework:** 🤗 Diffusers |
| - **Precision:** `bf16` |
|
|
| --- |
|
|
| ## Conditioning Type |
|
|
| - Primary conditioning: **Lineart / Canny-like edges** |
| - Backgrounds are mostly white |
| - Line quality varies (mostly clean, some noisy samples) |
|
|
| > Important limitation: |
| > Lineart / Canny **does not contain color information**, which leads to unstable and drifting color predictions. |
|
|
| --- |
|
|
| ## Dataset |
|
|
| - Size: ~**14,000 image pairs** |
| - Format: |
| - Original image (color) |
| - Conditioning image (lineart / canny) |
| - Prompt (caption) |
|
|
| ### Known dataset issues |
| - Some lineart images are **noisy or inconsistent** |
| - Images are resized to square resolution (possible cropping artifacts) |
| - No explicit color supervision |
| - No palette or region-level color constraints |
|
|
| --- |
|
|
| ## Training Configuration |
|
|
| Typical training setup: |
|
|
| ```bash |
| resolution: 768 |
| train_batch_size: 2 |
| gradient_accumulation_steps: 2 |
| effective_batch_size: 4 |
| learning_rate: 2e-5 |
| lr_scheduler: cosine |
| max_train_steps: 6000–8000 |
| mixed_precision: bf16 |
| |
