Upload 4 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+pixel-art-result.png filter=lfs diff=lfs merge=lfs -text

pixel_art_lo_ra_training_readme.md

@@ -0,0 +1,143 @@
+---
+language: en
+library_name: diffusers
+pipeline_tag: text-to-image
+license: apache-2.0
+base_model: Tongyi-MAI/Z-Image-Turbo
+tags:
+- lora
+- pixel-art
+- diffusion
+- text-to-image
+- style-adaptation
+---
+
+# 🎨 Pixel Art Character LoRA – Z-Image-Turbo
+
+This repository hosts **LoRA adapter weights** fine-tuned on top of **Tongyi-MAI/Z-Image-Turbo** to improve **pixel art character generation** from text prompts.
+
+The LoRA is optimized for prompts that start with or include:
+
+> **"a pixel art character ..."**
+
+---
+
+## 🚀 Model Description
+
+- **Base model**: `Tongyi-MAI/Z-Image-Turbo`
+- **Fine-tuning method**: LoRA (Low-Rank Adaptation)
+- **Task**: Text-to-image generation
+- **Specialization**: Pixel art characters
+- **Trainable parameters**: ~0.1% of base model
+
+This model does **not** replace the base model. Instead, it injects lightweight LoRA adapters into the transformer layers.
+
+---
+
+## 🧠 Why Pixel Art?
+
+Pixel art differs significantly from natural images:
+- Sharp, discrete edges
+- Limited color palettes
+- Low-resolution spatial structure
+
+Generic diffusion models often blur these characteristics. This LoRA improves:
+- Structural sharpness
+- Style consistency
+- Prompt–image alignment for pixel art descriptions
+
+---
+
+## 🧩 How to Use
+
+```python
+import torch
+from diffusers import DiffusionPipeline
+from peft import PeftModel
+
+pipe = DiffusionPipeline.from_pretrained(
+    "Tongyi-MAI/Z-Image-Turbo",
+    torch_dtype=torch.bfloat16,
+    device_map="cuda"
+)
+
+pipe.transformer = PeftModel.from_pretrained(
+    pipe.transformer,
+    "<your-username>/<repo-name>"
+)
+
+prompt = "a pixel art character with square orange glasses, a chef hat-shaped head and a purple-colored body on a cool background"
+
+image = pipe(prompt).images[0]
+image.save("pixel_art.png")
+```
+
+---
+
+## 🧪 Evaluation
+
+### CLIPScore (Prompt–Image Alignment)
+
+| Model | Normalized CLIPScore (mean ± std) |
+|------|----------------------------------|
+| Base Z-Image-Turbo | **7.834 ± 2.577** |
+| + Pixel Art LoRA | **8.856 ± 2.473** |
+
+➡️ **+1.02 CLIPScore improvement**, indicating stronger alignment with pixel-art-specific prompts.
+
+---
+
+## 📈 Training Details
+
+- **Dataset**: `m1guelpf/nouns`
+- **Image resolution**: 512×512
+- **Epochs**: 1
+- **Optimizer**: AdamW (`lr=1e-4`)
+- **Precision**: bfloat16
+- **Noise scheduler**: DDPM (300 steps)
+
+### LoRA Configuration
+```text
+r = 16
+lora_alpha = 32
+lora_dropout = 0.05
+target_modules = [to_q, to_k, to_v, to_out.0]
+```
+
+---
+
+## 🖼️ Example Prompts
+
+```text
+a pixel art character with a wizard hat and glowing blue eyes
+a pixel art character holding a sword and wearing red armor
+a pixel art character with a robot body and green visor
+```
+
+---
+
+## ⚠️ Limitations
+
+- Optimized primarily for **pixel art characters**
+- May not improve (or may slightly degrade) photorealistic prompts
+- Trained on a relatively small dataset
+
+---
+
+## 🔮 Future Work
+
+- Multi-epoch training with early stopping
+- Broader pixel-art prompt coverage
+- Palette-aware regularization
+
+---
+
+## 📜 License
+
+This LoRA follows the license of the base model **Tongyi-MAI/Z-Image-Turbo**.
+Please check the original repository for full license terms.
+
+---
+
+**Status**: Research / Experimental
+