README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen-VL
tags:
  - lora
  - qwen-image
  - anime
  - background
  - visual-novel
language:
  - en
pipeline_tag: text-to-image
library_name: peft

Anime Visual Novel Backgrounds LoRA

LoRA fine-tuned on Qwen-Image for generating consistent anime-style backgrounds for visual novels.

Model Details

• Base Model: Qwen-Image-Edit
• LoRA Rank: 8
• Training Steps: 600
• Dataset Size: 80 anime backgrounds
• Specialization: Visual novel backgrounds with time/weather control

Example Outputs

Before vs After - See the Difference 🎨

This LoRA dramatically improves consistency, style, and detail in anime background generation.

---

Beach Shore - Clear Sky
❌ Base Model (No LoRA)	✅ With LoRA ('anim3bg' style)

Prompt: anim3bg, beach shore with ocean waves, seagulls flying in bright blue sky, beach umbrellas and chairs in distance, white foam on turquoise water, clear sunny day with white clouds, no people

---

Empty Classroom - Sunset
❌ Base Model (No LoRA)	✅ With LoRA ('anim3bg' style)

Prompt: anim3bg, empty classroom at sunset, student desks and chairs arranged in rows, large windows with flowing curtains, green chalkboard, no people

---

Park Pathway - Autumn
❌ Base Model (No LoRA)	✅ With LoRA ('anim3bg' style)

Prompt: anim3bg, park pathway in autumn with falling maple leaves, ornate street lamp, wooden bench, no people

---

Train Interior - Rainy Night
❌ Base Model (No LoRA)	✅ With LoRA ('anim3bg' style)

Prompt: anim3bg, train interior at night, empty passenger car with teal seats, overhead handrails, windows showing rainy night outside with city lights, no people

---

Trigger Word

Required: anim3bg

Usage

Basic Prompt Structure

anim3bg, [location], [time], [weather], [specific elements], no people

Time of Day Keywords

• sunrise / early morning - Soft pink/orange tones, gentle lighting
• morning - Bright morning light, fresh atmosphere
• day / midday - Bright daylight, strong shadows
• afternoon - Warm afternoon light, golden tones
• sunset / golden hour - Golden orange lighting, warm glow
• evening / dusk - Post-sunset, blue hour tones
• night - Dark with artificial lighting, street lamps
• late night - Very dark, minimal lighting, quiet atmosphere

Weather Keywords

• clear sky / clear - No weather effects, good visibility
• cloudy / overcast - Gray sky, diffused lighting
• partly cloudy - Mixed conditions, scattered clouds
• rainy / light rain - Rain effects, wet surfaces with reflections
• heavy rain / stormy - Intense rain, dramatic atmosphere
• foggy / misty - Reduced visibility, mysterious atmosphere
• snowy / snow falling - Snow effects, white coverage
• windy - Movement suggestion, leaves blowing

Lighting Keywords

Use these for better control:

• warm lighting - Orange/yellow tones
• soft lighting - Gentle, diffused light
• dramatic lighting - Strong contrast, shadows
• golden hour lighting - Sunset/sunrise warm glow
• streaming light / light beams - Rays of light visible
• reflections - Light bouncing off surfaces
• glowing - Artificial lights illuminated
• atmospheric lighting - Mood-setting light

Recommended Settings

For Best Results:

• Always include trigger word anim3bg at the start
• Use natural language descriptions (not tags)
• Be specific about location and architectural details
• Include time of day and weather conditions
• End with "no people"

Prompt Writing Tips

✅ DO:

• Use descriptive adjectives: "dramatic light reflections", "warm golden lighting"
• Include spatial details: "left wall", "right side", "in center"
• Describe materials: "polished wooden floor", "brick pathways"
• Add atmosphere: "nostalgic", "peaceful", "moody"
• Be specific about colors: "orange and yellow lighting", "pink and orange sky"

❌ DON'T:

• Use structured tags: location:school, time:sunset
• Use abbreviations: bg, loc
• Be too generic: just "school" instead of "japanese school hallway interior"
• Skip the trigger word anim3bg
• Include character-related terms

Example of Good vs Bad Prompts:

❌ Bad (Too Generic):

anim3bg, school, sunset

✅ Good (Detailed & Natural):

anim3bg, school courtyard at sunset, with cherry blossom trees in full bloom, decorative fountain in center, brick pathway, no people

## Training Details

- **Training Data**: Custom curated dataset of anime backgrounds
- **Locations Covered**: 
  - School environments (classrooms, hallways, courtyards, rooftops)
  - Urban settings (streets, stations, shopping districts)
  - Nature scenes (beaches, parks, forests, riversides)
  - Indoor locations (cafes, libraries, homes)
- **Variations**: Multiple times of day and weather conditions per location
- **Optimizer**: AdamW
- **Learning Rate**: 1e-5
- **Scheduler**: Cosine with warmup
- **Caption Style**: Natural language descriptions

## Limitations

- Best results with locations similar to training data
- May require multiple generations for perfect consistency
- Not designed for character generation
- Works best with detailed, natural language prompts

## Version History

- **v1.0** (2025-10-30): Initial release
  - 80 training images
  - LoRA Rank 16
  - 1000 training steps

- **v1.0** (2025-11-04):
  - 80 training images
  - LoRA Rank 8
  - 600 training steps

## License

Apache 2.0

## Acknowledgments

Trained using Qwen-Image-Edit base model. Special thanks to the Qwen team for the excellent base model.

## Tips for Integration in Visual Novels

1. **Generate variations**: Create 2-3 versions of each scene
2. **Maintain logs**: Keep track of prompts and seeds used
3. **Post-processing**: Minor touch-ups in Photoshop/Krita if needed
4. **Consistent naming**: Use organized file naming (e.g., `ch01_scene03_school_morning.png`)
5. **Test in context**: View backgrounds with character sprites before finalizing
6. **Backup everything**: Save prompts, seeds, and original generations

**Enjoy creating beautiful anime backgrounds for your visual novel!** 🎨✨

---