[Doc] Add Z-Image Base support docs (and turbo tweaks)

Add documentation and examples for Z-Image Base model alongside existing
Z-Image Turbo support.

Changes:
- Add z_image_examples.py with Base vs Turbo comparison examples
- Update quickstart.md with comments distinguishing Base and Turbo
- Update supported_models.md to list both Z-Image variants
- Update text_to_image README with usage examples and comparison table
- Fix negative_prompt to be passed in prompt dict (not sampling params)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
Signed-off-by: dougbtv <dosmith@redhat.com>

Author: dougbtv

docs/getting_started/quickstart.md

@@ -39,6 +39,8 @@ Text-to-image generation quickstart with vLLM-Omni:
 from vllm_omni.entrypoints.omni import Omni
 
 if __name__ == "__main__":
+    # Use Z-Image-Turbo for fast inference (8 steps)
+    # or Z-Image (Base) for higher quality (28-50 steps with CFG support)
     omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
     prompt = "a cup of coffee on the table"
     outputs = omni.generate(prompt)
@@ -59,8 +61,9 @@ You can pass a list of prompts and wait for them to process altogether, shown be
 from vllm_omni.entrypoints.omni import Omni
 
 if __name__ == "__main__":
+    # For batch inference with Z-Image models
     omni = Omni(
-        model="Tongyi-MAI/Z-Image-Turbo",
+        model="Tongyi-MAI/Z-Image-Turbo",  # or "Tongyi-MAI/Z-Image" for Base
         # stage_configs_path="./stage-config.yaml",  # See below
     )
     prompts = [
@@ -93,7 +96,11 @@ For more usages, please refer to [offline inference](../user_guide/examples/offl
 Text-to-image generation quickstart with vLLM-Omni:
 
 ```bash
+# Fast inference with Turbo (8 steps, no CFG)
 vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091
+
+# Or use Base model for higher quality (50 steps, CFG support)
+# vllm serve Tongyi-MAI/Z-Image --omni --port 8091
 ```
 
 ```bash

docs/models/supported_models.md

@@ -26,7 +26,7 @@ th {
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
 | `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
 | `GlmImagePipeline` | GLM-Image | `zai-org/GLM-Image` |
-|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
+|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image` (Base), `Tongyi-MAI/Z-Image-Turbo` |
 | `WanPipeline` | Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` |
 | `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` |
@@ -60,7 +60,7 @@ th {
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
 | `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2511 | `Qwen/Qwen-Image-Edit-2511` |
-|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
+|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image` (Base), `Tongyi-MAI/Z-Image-Turbo` |
 |`LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` |
 |`Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` |
 |`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` |

examples/offline_inference/text_to_image/README.md

@@ -1,8 +1,9 @@
 # Text-To-Image
 
-This folder provides several entrypoints for experimenting with `Qwen/Qwen-Image` `Qwen/Qwen-Image-2512` `Tongyi-MAI/Z-Image-Turbo` using vLLM-Omni:
+This folder provides several entrypoints for experimenting with `Qwen/Qwen-Image`, `Qwen/Qwen-Image-2512`, `Tongyi-MAI/Z-Image` (Base), and `Tongyi-MAI/Z-Image-Turbo` using vLLM-Omni:
 
 - `text_to_image.py`: command-line script for single image generation with advanced options.
+- `z_image_examples.py`: comparison examples showing Z-Image Base vs Turbo usage.
 - `web_demo.py`: lightweight Gradio UI for interactive prompt/seed/CFG exploration.
 
 Note that when you pass in multiple independent prompts, they will be processed sequentially. Batching requests is currently not supported.
@@ -74,17 +75,33 @@ if __name__ == "__main__":
 
 ## Local CLI Usage
 
+### Z-Image Turbo (Fast Inference)
 ```bash
 python text_to_image.py \
   --model Tongyi-MAI/Z-Image-Turbo \
   --prompt "a cup of coffee on the table" \
   --seed 42 \
-  --cfg_scale 4.0 \
   --num_images_per_prompt 1 \
-  --num_inference_steps 50 \
+  --num_inference_steps 8 \
+  --guidance_scale 0.0 \
   --height 1024 \
   --width 1024 \
-  --output outputs/coffee.png
+  --output outputs/coffee_turbo.png
+```
+
+### Z-Image Base (High Quality with CFG)
+```bash
+python text_to_image.py \
+  --model Tongyi-MAI/Z-Image \
+  --prompt "a cup of coffee on the table" \
+  --negative_prompt "blurry, low quality, distorted" \
+  --seed 42 \
+  --num_images_per_prompt 1 \
+  --num_inference_steps 50 \
+  --guidance_scale 4.0 \
+  --height 1280 \
+  --width 720 \
+  --output outputs/coffee_base.png
 ```
 
 Key arguments:
@@ -103,6 +120,26 @@ Key arguments:
 
 > ℹ️ If you encounter OOM errors, try using `--vae_use_slicing` and `--vae_use_tiling` to reduce memory usage.
 
+## Z-Image Base vs Turbo Comparison
+
+For detailed comparison and usage examples of both Z-Image variants, see:
+
+```bash
+python z_image_examples.py --example all
+```
+
+Key differences:
+
+| Feature | Z-Image Base | Z-Image Turbo |
+|---------|--------------|---------------|
+| Model | `Tongyi-MAI/Z-Image` | `Tongyi-MAI/Z-Image-Turbo` |
+| Inference Steps | 28-50 (default: 50) | 8 |
+| CFG Support | ✅ Yes (guidance_scale 3.0-5.0) | ❌ Must use 0.0 |
+| Negative Prompts | ✅ Supported | ❌ Not supported |
+| Fine-tunable | ✅ Yes | ❌ No (distilled) |
+| Scheduler Shift | 6.0 | 3.0 |
+| Best For | High quality, fine-tuning | Fast iteration, speed |
+
 > ℹ️ Qwen-Image currently publishes best-effort presets at `1328x1328`, `1664x928`, `928x1664`, `1472x1140`, `1140x1472`, `1584x1056`, and `1056x1584`. Adjust `--height/--width` accordingly for the most reliable outcomes.
 
 ## Web UI Demo

examples/offline_inference/text_to_image/text_to_image.py

@@ -21,7 +21,7 @@ def parse_args() -> argparse.Namespace:
         "--model",
         default="Qwen/Qwen-Image",
         help="Diffusion model name or local path. Supported models: "
-        "Qwen/Qwen-Image, Tongyi-MAI/Z-Image-Turbo, Qwen/Qwen-Image-2512",
+        "Qwen/Qwen-Image, Tongyi-MAI/Z-Image (Base), Tongyi-MAI/Z-Image-Turbo, Qwen/Qwen-Image-2512",
     )
     parser.add_argument("--prompt", default="a cup of coffee on the table", help="Text prompt for image generation.")
     parser.add_argument(

examples/offline_inference/text_to_image/z_image_examples.py

@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Z-Image Base vs Turbo Comparison Examples
+
+This script demonstrates the differences between Z-Image Base and Z-Image Turbo models.
+
+Key Differences:
+- Z-Image Base: Foundation model with full CFG support, fine-tunable, 28-50 steps
+- Z-Image Turbo: Distilled model optimized for speed, 8 steps, guidance_scale must be 0.0
+"""
+
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+
+def z_image_base_example():
+    """
+    Z-Image Base - High Quality Generation
+
+    Features:
+    - Full CFG support (guidance_scale 3.0-5.0)
+    - Negative prompts work
+    - Fine-tunable
+    - 28-50 inference steps (default 50)
+    - Scheduler shift: 6.0
+    """
+    print("\n=== Z-Image Base (High Quality) ===")
+
+    omni_base = Omni(model="Tongyi-MAI/Z-Image")
+
+    outputs_base = omni_base.generate(
+        {
+            "prompt": "a majestic mountain landscape at sunset, detailed, photorealistic",
+            "negative_prompt": "blurry, low quality, distorted, oversaturated",
+        },
+        OmniDiffusionSamplingParams(
+            height=1280,
+            width=720,
+            num_inference_steps=50,
+            guidance_scale=4.0,
+            seed=42,
+        ),
+    )
+
+    images = outputs_base[0].request_output[0].images
+    images[0].save("z_image_base_output.png")
+    print("Saved to: z_image_base_output.png")
+    print(f"Generated {len(images)} image(s) with 50 steps and CFG=4.0")
+
+
+def z_image_turbo_example():
+    """
+    Z-Image Turbo - Fast Inference
+
+    Features:
+    - Optimized for speed
+    - guidance_scale MUST be 0.0 (no CFG)
+    - Negative prompts not supported
+    - 8 inference steps
+    - Scheduler shift: 3.0
+    """
+    print("\n=== Z-Image Turbo (Fast) ===")
+
+    omni_turbo = Omni(model="Tongyi-MAI/Z-Image-Turbo")
+
+    outputs_turbo = omni_turbo.generate(
+        "a majestic mountain landscape at sunset, detailed, photorealistic",
+        OmniDiffusionSamplingParams(
+            height=1024,
+            width=1024,
+            num_inference_steps=8,
+            guidance_scale=0.0,  # MUST be 0.0 for Turbo!
+            seed=42,
+        ),
+    )
+
+    images = outputs_turbo[0].request_output[0].images
+    images[0].save("z_image_turbo_output.png")
+    print("Saved to: z_image_turbo_output.png")
+    print(f"Generated {len(images)} image(s) with 8 steps (no CFG)")
+
+
+def batch_inference_example():
+    """
+    Batch inference with Z-Image Base
+
+    Note: Batch processing depends on max_batch_size in stage configs.
+    By default, diffusion models process one prompt at a time.
+    """
+    print("\n=== Batch Inference Example ===")
+
+    omni = Omni(model="Tongyi-MAI/Z-Image")
+
+    prompts = [
+        {"prompt": "a cup of coffee on a wooden table", "negative_prompt": "blurry, low quality"},
+        {"prompt": "a cat sleeping on a cozy blanket", "negative_prompt": "blurry, low quality"},
+        {"prompt": "a futuristic city skyline at night", "negative_prompt": "blurry, low quality"},
+    ]
+
+    # Note: These will be processed sequentially unless max_batch_size > 1
+    outputs = omni.generate(
+        prompts,
+        OmniDiffusionSamplingParams(
+            height=1024,
+            width=1024,
+            num_inference_steps=40,
+            guidance_scale=4.0,
+            seed=42,
+        ),
+    )
+
+    for i, output in enumerate(outputs):
+        image = output.request_output[0].images[0]
+        image.save(f"batch_output_{i}.png")
+        print(f"Saved to: batch_output_{i}.png")
+
+
+def recommended_settings():
+    """
+    Print recommended settings for both models
+    """
+    print("\n=== Recommended Settings ===\n")
+
+    print("Z-Image Base (Tongyi-MAI/Z-Image):")
+    print("  - num_inference_steps: 28-50 (default: 50)")
+    print("  - guidance_scale: 3.0-5.0 (default: 4.0)")
+    print("  - negative_prompt: Supported and recommended")
+    print("  - resolution: 1280x720 or 720x1280")
+    print("  - cfg_normalization: False (default)")
+    print("  - Use when: Quality is priority, fine-tuning needed")
+
+    print("\nZ-Image Turbo (Tongyi-MAI/Z-Image-Turbo):")
+    print("  - num_inference_steps: 8")
+    print("  - guidance_scale: 0.0 (REQUIRED)")
+    print("  - negative_prompt: Not supported")
+    print("  - resolution: 1024x1024")
+    print("  - Use when: Speed is priority, quick iterations")
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(description="Z-Image Base vs Turbo comparison examples")
+    parser.add_argument(
+        "--example",
+        choices=["base", "turbo", "batch", "all"],
+        default="all",
+        help="Which example to run (default: all)",
+    )
+
+    args = parser.parse_args()
+
+    recommended_settings()
+
+    if args.example in ("base", "all"):
+        z_image_base_example()
+
+    if args.example in ("turbo", "all"):
+        z_image_turbo_example()
+
+    if args.example in ("batch", "all"):
+        batch_inference_example()
+
+    print("\nDone!")