Complete Z-Image Base vs Turbo comparison: 4 prompts x 2 models (8 images)

README.md

@@ -1,10 +1,25 @@
+---
+license: apache-2.0
+language:
+- en
+- zh
+tags:
+- z-image
+- text-to-image
+- diffusion
+- comparison
+- benchmark
+---
+
 # ⚡ Z-Image Base vs Turbo 效果对比
 
+使用相同的 prompt 和 seed，对比 [Z-Image (Base)](https://hf.co/Tongyi-MAI/Z-Image) 和 [Z-Image-Turbo](https://hf.co/Tongyi-MAI/Z-Image-Turbo) 的生成效果。
+
 ## 📊 模型信息对比
 
 | 特性 | Z-Image (Base) | Z-Image-Turbo |
 |------|:-:|:-:|
-| 模型链接 | [Tongyi-MAI/Z-Image](https://hf.co/Tongyi-MAI/Z-Image) | [Tongyi-MAI/Z-Image-Turbo](https://hf.co/Tongyi-MAI/Z-Image-Turbo) |
+| 模型 | [Tongyi-MAI/Z-Image](https://hf.co/Tongyi-MAI/Z-Image) | [Tongyi-MAI/Z-Image-Turbo](https://hf.co/Tongyi-MAI/Z-Image-Turbo) |
 | 参数量 | 6B | 6B (蒸馏版) |
 | 架构 | S3-DiT (单流 DiT) | S3-DiT (单流 DiT) |
 | 推理步数 | 28-50 步 | **8 步** |
@@ -14,95 +29,91 @@
 | 视觉质量 | ⭐⭐⭐ 高 | ⭐⭐⭐⭐ 非常高 |
 | RL 强化学习 | ❌ | ✅ (DMDR) |
 | 可微调 (LoRA等) | ✅ | ❌ |
-| 蒸馏技术 | — | Decoupled-DMD |
-| VRAM 需求 | ~16GB+ | **≤16GB** (消费级) |
-| H800 推理延迟 | ~数秒 | **<1秒** (亚秒级) |
+| VRAM 需求 | ~16GB+ | **≤16GB** |
 
-## 🖼️ Z-Image-Turbo 生成效果 (8步, 无CFG)
+## 🖼️ 并排对比
 
 ### Prompt 1: 🐆 写实摄影 / 野生动物
 > A majestic snow leopard resting on a rocky mountain ledge at golden hour, with dramatic clouds and snow-capped peaks in the background, photorealistic, 8K, National Geographic style
 
-![Turbo - Wildlife](turbo_1_wildlife.png)
+| Z-Image Base (30 steps, CFG=4.0) | Z-Image Turbo (8 steps, no CFG) |
+|:---:|:---:|
+| ![Base](base_1_wildlife.png) | ![Turbo](turbo_1_wildlife.png) |
 
 ---
 
 ### Prompt 2: 👩 人像 / 中国风
 > A beautiful young woman in traditional Chinese hanfu, sitting in a bamboo garden with cherry blossoms falling, soft natural lighting, portrait photography, shallow depth of field
 
-![Turbo - Portrait](turbo_2_portrait.png)
+| Z-Image Base (30 steps, CFG=4.0) | Z-Image Turbo (8 steps, no CFG) |
+|:---:|:---:|
+| ![Base](base_2_portrait.png) | ![Turbo](turbo_2_portrait.png) |
 
 ---
 
 ### Prompt 3: 🌃 赛博朋克 / 电影感
 > A cyberpunk city street at night, neon signs in Chinese and English, rain-soaked reflections on the road, a lone figure with an umbrella walking away, cinematic atmosphere, blade runner style
 
-![Turbo - Cyberpunk](turbo_3_cyberpunk.png)
+| Z-Image Base (30 steps, CFG=4.0) | Z-Image Turbo (8 steps, no CFG) |
+|:---:|:---:|
+| ![Base](base_3_cyberpunk.png) | ![Turbo](turbo_3_cyberpunk.png) |
 
 ---
 
 ### Prompt 4: ☕ 插画 / 水彩风
 > A cozy café interior with warm lighting, a steaming cup of latte art on a wooden table, bookshelves in the background, watercolor painting style, soft pastel colors
 
-![Turbo - Illustration](turbo_4_illustration.png)
+| Z-Image Base (30 steps, CFG=4.0) | Z-Image Turbo (8 steps, no CFG) |
+|:---:|:---:|
+| ![Base](base_4_illustration.png) | ![Turbo](turbo_4_illustration.png) |
 
 ---
 
 ## 📋 生成参数
-| 参数 | Turbo 设置 | Base 推荐设置 |
-|------|-----------|-------------|
+
+| 参数 | Base 设置 | Turbo 设置 |
+|------|-----------|-----------|
 | 分辨率 | 1024×1024 | 1024×1024 |
 | 种子 (Seed) | 42 | 42 |
-| 推理步数 | 8 | 30 (推荐28-50) |
-| Guidance Scale | 0.0 | 4.0 (推荐3.0-5.0) |
-| CFG Normalization | — | False |
-| 负向提示词 | 不支持 | 支持 |
-
-## 🎮 在线交互对比
+| 推理步数 | 30 | 8 |
+| Guidance Scale | 4.0 | 0.0 |
+| 负向提示词 | "" (空) | 不支持 |
+| 推理后端 | fal-ai | fal-ai |
 
-直接在线体验两个模型的效果差异：
+## 🎮 在线体验
 
 | 模型 | 在线 Demo | Playground |
 |------|-----------|-----------|
-| Z-Image (Base) | [🤗 Space](https://huggingface.co/spaces/Tongyi-MAI/Z-Image) | [🎮 Playground](https://hf.co/playground?modelId=Tongyi-MAI/Z-Image) |
-| Z-Image-Turbo | [🤗 Space](https://huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo) | [🎮 Playground](https://hf.co/playground?modelId=Tongyi-MAI/Z-Image-Turbo) |
+| Z-Image Base | [🤗 Space](https://huggingface.co/spaces/Tongyi-MAI/Z-Image) | [🎮 Playground](https://hf.co/playground?modelId=Tongyi-MAI/Z-Image) |
+| Z-Image Turbo | [🤗 Space](https://huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo) | [🎮 Playground](https://hf.co/playground?modelId=Tongyi-MAI/Z-Image-Turbo) |
 
-## 📝 总结对比
+## 📝 总结
 
-### 🏆 各维度赢家
+### 🏆 各维度对比
 
-| 维度 | 赢家 | 原因 |
+| 维度 | 赢家 | 说明 |
 |------|------|------|
-| **推理速度** | 🥇 **Turbo** | 仅需 8 步 (vs Base 28-50步)，H800 上亚秒级推理 |
-| **视觉质量** | 🥇 **Turbo** | 经过 DMDR (蒸馏+强化学习) 联合优化，细节更丰富、更锐利 |
-| **生成多样性** | 🥇 **Base** | 未蒸馏模型保留完整训练信号，不同 seed 生成差异更大 |
-| **可控性** | 🥇 **Base** | 支持 CFG 和负向提示词，可精细调控生成方向 |
-| **可微调性** | 🥇 **Base** | 未蒸馏，适合作为 LoRA/ControlNet/Fine-tuning 基座 |
-| **部署成本** | 🥇 **Turbo** | 更少计算步数 + 16GB VRAM 即可运行 |
-| **文字渲染** | 🥇 **Turbo** | 中英文双语文字渲染更准确 |
-
-### 💡 使用建议
-
-- **追求速度和质量** → 选 **Turbo** (日常使用、产品级部署、实时生成)
-- **追求多样性和可控性** → 选 **Base** (创意探索、精细调控、复杂场景)
-- **需要二次开发** → 选 **Base** (LoRA 训练、ControlNet、下游微调)
-- **资源受限** → 选 **Turbo** (16GB 消费级显卡可运行)
-
-### 🔬 技术亮点
-
-**Z-Image-Turbo** 的核心加速技术:
-1. **Decoupled-DMD**: 将 CFG 增强和分布匹配解耦优化的蒸馏算法
-2. **DMDR**: 将 DMD 蒸馏与强化学习 (RL) 协同融合
-   - RL 释放 DMD 的潜力 🚀
-   - DMD 有效正则化 RL ⚖️
-
-**Z-Image (Base)** 的核心优势:
-1. **完整训练信号**: 未蒸馏，保留全部生成能力
-2. **负向提示控制**: 高保真地响应负向提示词
-3. **美学多样性**: 覆盖超写实、电影级数字艺术、动漫、插画等多种风格
-
-## 📜 参考论文
-
-- [Z-Image: An Efficient Image Generation Foundation Model](https://arxiv.org/abs/2511.22699)
-- [Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield](https://arxiv.org/abs/2511.22677)
-- [Distribution Matching Distillation Meets Reinforcement Learning](https://arxiv.org/abs/2511.13649)
+| **推理速度** | 🥇 Turbo | 8 步 vs 30-50 步，H800 上亚秒级推理 |
+| **视觉质量** | 🥇 Turbo | DMDR 优化，细节更丰富锐利 |
+| **生成多样性** | 🥇 Base | 未蒸馏，不同 seed 差异更大 |
+| **可控性** | 🥇 Base | 支持 CFG + 负向提示词 |
+| **可微调** | 🥇 Base | LoRA / ControlNet 基座 |
+| **部署成本** | 🥇 Turbo | 16GB VRAM 即可 |
+| **文字渲染** | 🥇 Turbo | 中英双语更准确 |
+
+### 💡 选择建议
+
+| 场景 | 推荐 |
+|------|------|
+| 日常使用、快速出图 | **Turbo** |
+| 产品级部署、实时生成 | **Turbo** |
+| 创意探索、多样性需求 | **Base** |
+| 精细控制、负向提示 | **Base** |
+| LoRA/ControlNet 开发 | **Base** |
+| 消费级显卡 (≤16GB) | **Turbo** |
+
+## 📜 参考
+
+- [Z-Image Paper](https://arxiv.org/abs/2511.22699)
+- [Decoupled-DMD](https://arxiv.org/abs/2511.22677)
+- [DMDR](https://arxiv.org/abs/2511.13649)

results.json

@@ -8,7 +8,18 @@
     "cfg": 0.0,
     "seed": 42,
     "filename": "turbo_1_wildlife.png",
-    "exists": true
+    "file_size_kb": 1391
+  },
+  {
+    "prompt_id": "1_wildlife",
+    "category": "写实摄影 / 野生动物",
+    "prompt": "A majestic snow leopard resting on a rocky mountain ledge at golden hour, with dramatic clouds and snow-capped peaks in the background, photorealistic, 8K, National Geographic style",
+    "model": "Z-Image (Base)",
+    "steps": 30,
+    "cfg": 4.0,
+    "seed": 42,
+    "filename": "base_1_wildlife.png",
+    "file_size_kb": 1591
   },
   {
     "prompt_id": "2_portrait",
@@ -19,7 +30,18 @@
     "cfg": 0.0,
     "seed": 42,
     "filename": "turbo_2_portrait.png",
-    "exists": true
+    "file_size_kb": 1178
+  },
+  {
+    "prompt_id": "2_portrait",
+    "category": "人像 / 中国风",
+    "prompt": "A beautiful young woman in traditional Chinese hanfu, sitting in a bamboo garden with cherry blossoms falling, soft natural lighting, portrait photography, shallow depth of field",
+    "model": "Z-Image (Base)",
+    "steps": 30,
+    "cfg": 4.0,
+    "seed": 42,
+    "filename": "base_2_portrait.png",
+    "file_size_kb": 1289
   },
   {
     "prompt_id": "3_cyberpunk",
@@ -30,7 +52,18 @@
     "cfg": 0.0,
     "seed": 42,
     "filename": "turbo_3_cyberpunk.png",
-    "exists": true
+    "file_size_kb": 1456
+  },
+  {
+    "prompt_id": "3_cyberpunk",
+    "category": "赛博朋克 / 电影感",
+    "prompt": "A cyberpunk city street at night, neon signs in Chinese and English, rain-soaked reflections on the road, a lone figure with an umbrella walking away, cinematic atmosphere, blade runner style",
+    "model": "Z-Image (Base)",
+    "steps": 30,
+    "cfg": 4.0,
+    "seed": 42,
+    "filename": "base_3_cyberpunk.png",
+    "file_size_kb": 1830
   },
   {
     "prompt_id": "4_illustration",
@@ -41,6 +74,17 @@
     "cfg": 0.0,
     "seed": 42,
     "filename": "turbo_4_illustration.png",
-    "exists": true
+    "file_size_kb": 1410
+  },
+  {
+    "prompt_id": "4_illustration",
+    "category": "插画 / 水彩",
+    "prompt": "A cozy café interior with warm lighting, a steaming cup of latte art on a wooden table, bookshelves in the background, watercolor painting style, soft pastel colors",
+    "model": "Z-Image (Base)",
+    "steps": 30,
+    "cfg": 4.0,
+    "seed": 42,
+    "filename": "base_4_illustration.png",
+    "file_size_kb": 1815
   }
 ]