--- --- 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 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0) [![Model](https://img.shields.io/badge/Model-Qwen--VL-orange.svg)](https://huggingface.co/Qwen/Qwen-VL) [![LoRA](https://img.shields.io/badge/LoRA-Rank%208-green.svg)](#) [![Version](https://img.shields.io/badge/Version-2.0-purple.svg)](#version-history) --- ### ⚡ Current Version Stats | Metric | Value | |--------|-------| | **Training Images** | 80 anime-style backgrounds generated by seedream-4 | | **LoRA Rank** | 8 | | **Training Steps** | 600 | | **Base Model** | Qwen-Image-Edit | | **Trigger Word** | `anim3bg` | | **Recommended Weight** | 0.7 - 0.9 |
--- ## 📊 Version History
Version 2.0 (Current) #### Training Configuration - **Dataset:** 80 images - **Steps:** 600 - **LoRA Rank:** 8 - **Learning Rate:** 0.0001 - **Batch Size:** 1
Version 1.0 #### Training Configuration - **Dataset:** 80 images - **Steps:** 1000 - **LoRA Rank:** 16 - **Learning Rate:** 0.0004 - **Batch Size:** 1
--- ## 💬 Trigger Word **Required**: `anim3bg` --- ## 🚀 Usage **Simple Prompts:** ``` anim3bg, [location] ``` **Detailed Prompts:** ``` anim3bg, [detailed location description], [time], [weather], no characters, high quality ``` --- ## 📥 Download & Use 1. Download `anime_background_lora.safetensors` 2. Load in Wavespeed AI or compatible platform 3. Set weight to 0.8 4. Include `anim3bg` trigger word in prompt 5. Generate ---
## 📸 Example Outputs: Before vs After Comparison (Simple Prompts) *Testing how well the LoRA performs with minimal prompt detail* #### **Example 1:**
❌ Without LoRA

Prompt: anime-style background, library

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, library

--- #### **Example 2:**
❌ Without LoRA

Prompt: anime-style background, beachside café

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, beachside café

--- #### **Example 3:**
❌ Without LoRA

Prompt: anime-style background, city street at night

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, city street at night

--- #### **Example 4:**
❌ Without LoRA

Prompt: anime-style background, forest path

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, forest path

--- #### **Example 5:**
❌ Without LoRA

Prompt: anime-style background, train station

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, train station

--- ## 📸 Example Outputs: Before vs After Comparison (Detailed Prompts) *Testing how well the LoRA handles complex, detailed descriptions* #### **Example 6:**
❌ Without LoRA

Prompt: anime-style background, mountain lake at sunset, dark narrow japanese alley, neon signs with japanese text, vending machines, bicycle parked on street

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, mountain lake at sunset, dark narrow japanese alley, neon signs with japanese text, vending machines, bicycle parked on street

--- #### **Example 7:**
❌ Without LoRA

Prompt: anime-style background, empty amusement park at sunrise, blanketed in snow, Ferris wheel half-frozen, colorful lights faintly glowing through the frost

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, empty amusement park at sunrise, blanketed in snow, Ferris wheel half-frozen, colorful lights faintly glowing through the frost

--- #### **Example 8:**
❌ Without LoRA

Prompt: anime-style background, japanese shrine festival at night, festival food and game stalls on both sides, glowing orange paper lanterns overhead, fireworks in night sky

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, japanese shrine festival at night, festival food and game stalls on both sides, glowing orange paper lanterns overhead, fireworks in night sky

--- #### **Example 9:**
❌ Without LoRA

Prompt: anime-style background, beach with white cloud, seagulls flying, beach umbrellas and chairs in distance, white foam on turquoise water

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, beach with white cloud, seagulls flying, beach umbrellas and chairs in distance, white foam on turquoise water

--- #### **Example 10:**
❌ Without LoRA

Prompt: anime-style background, mystical waterfall scene at night, cascading down rocky cliff face, rainbow visible in the mist, turquoise blue pool at base

✅ With LoRA (trigger word 'anim3bg')

Prompt: anim3bg, mystical waterfall scene at night, cascading down rocky cliff face, rainbow visible in the mist, turquoise blue pool at base

--- # 🎨 [LoRA 名称] — Cinematic Anime Backgrounds (Outdoor) ## 📄 简介 此 LoRA 专注于生成**户外电影感(cinematic)动漫背景**:夕阳 / 金色时刻、画面感强的云与光线、宁静的日本郊区街景。 **触发词(Trigger)**: `shinkaibg` --- ## 🧪 对比实验说明 展示相同 seed、相同 prompt 在 **未应用 LoRA / 应用 LoRA** 下的差异;以及室内提示的过拟合测试(模型偏向户外氛围)。 --- ## 🔁 Same Seed — Same Prompt 比较(示例) **Prompt 示例:** `a quiet japanese street at sunset, cinematic lighting, painterly detail, shinkaibg` | Prompt | Without LoRA | With LoRA | |--------|--------------|----------| | same seed, same prompt | ![no_lora_street](images/same_seed_prompt_without_lora.png)
*Without LoRA* | ![with_lora_street](images/same_seed_prompt_with_lora.png)
*With LoRA (`shinkaibg` applied)* | **说明:** LoRA 版本通常会加入更柔和的光线过渡、更丰富的天空细节与更强的电影感(色彩与景深)。 --- ## 🏠 室内过拟合测试(Indoor Overfitting Test) **测试说明:** 使用相同 seed,四张图展示 `simple prompt` / `detailed prompt`,并比较有无 LoRA 的输出差异。 **Prompts 示例:** - **simple:** `a cozy living room` - **detailed:** `a cozy japanese-style living room with warm sunlight through window, tatami floor, shinkaibg` | Type | Without LoRA | With LoRA | |------|--------------|----------| | Simple prompt | ![simple_no](images/indoor_simple_without_lora.png)
*No LoRA* | ![simple_yes](images/indoor_simple_with_lora.png)
*With LoRA — 注意可能出现室外光效 / 窗外景观* | | Detailed prompt | ![detailed_no](images/indoor_detailed_without_lora.png)
*No LoRA* | ![detailed_yes](images/indoor_detailed_with_lora.png)
*With LoRA — 需更详尽描述以避免户外化 bias* | **观察:** 若只使用简单室内 prompt,模型往往会“套用”户外 / 日落光线(显示过拟合)。详细 prompt 能降低但不完全消除该现象。 --- ## ⚙️ 推荐参数(示例) - LoRA 权重 (strength): `0.7` - `0.9` - CFG Scale: `7` - `9` - Sampler: `DPM++ 2M Karras`(示例) - Resolution: `1024×1024` --- ## 🧾 标签与数据集信息 - 数据集规模:`~[填数字]` 张户外风景图 - 主要风格标签:`shinkaibg`, `cinematic`, `sunset`, `painterly` - 注意:训练集中强烈偏向户外场景,导致 indoor prompt bias(详见上方测试) --- ## 🔍 使用示例(Prompt)