<h1 id="z-image-aio-collection">🚀 Z-Image AIO Collection</h1><h2 id="base-and-turbo-*-all-in-one-*-bilingual-text-*-qwen3-4b">⚡ Base &amp; Turbo • All-in-One • Bilingual Text • Qwen3-4B</h2><hr /><h1 id="important:-requires-comfyui-v0.11.0+">⚠️ IMPORTANT: Requires ComfyUI v0.11.0+</h1><p>📥 <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI">Download ComfyUI</a></p><hr /><h1 id="what-is-z-image-aio">✨ What is Z-Image AIO?</h1><p>Z-Image AIO is an <strong>All-in-One repackage</strong> of Alibaba Tongyi Lab's 6B parameter image generation models.</p><p><strong>Everything integrated:</strong></p><ul><li><p>✅ VAE already built-in</p></li><li><p>✅ Qwen3-4B Text Encoder integrated</p></li><li><p>✅ Just download and generate!</p></li></ul><hr /><h1 id="available-versions">🎯 Available Versions</h1><hr /><h2 id="z-image-turbo-aio-(8-steps-*-cfg-1.0)">🔥 Z-Image-Turbo-AIO (8 Steps • CFG 1.0)</h2><p><strong>Ultra-fast generation for production &amp; daily use</strong></p><hr /><h3 id="nvfp4-aio-(7.8-gb)">⚫ NVFP4-AIO (7.8 GB) 🆕</h3><pre><code>🎯 ONLY for NVIDIA Blackwell GPUs (RTX 50xx)!
⚡ Maximum speed optimized
💾 Smallest file size
🚀 FP4 precision - blazing fast
</code></pre><p><strong>Perfect for:</strong> RTX 5070, 5080, 5090 owners who want maximum speed</p><hr /><h3 id="fp8-aio-(10-gb)-recommended">🟡 FP8-AIO (10 GB) ⭐ RECOMMENDED</h3><pre><code>✅ Best balance of size &amp; quality
✅ Works on 8GB VRAM
✅ Fast downloads
✅ Ideal for most users
</code></pre><p><strong>Perfect for:</strong> Daily use, testing, RTX 3060/4060/4070</p><hr /><h3 id="fp16-aio-(20-gb)">🔵 FP16-AIO (20 GB)</h3><pre><code>💾 Same file size as BF16
🔄 ComfyUI auto-casts to BF16 for compute
⚠️ Does NOT enable FP16 compute mode
📦 Alternative download option
</code></pre><p><strong>Note:</strong> Z-Image does not support FP16 compute - activation values exceed FP16's max range, causing NaN/black images. Weights are cast to BF16 during inference regardless of file format.</p><p><strong>Perfect for:</strong> Alternative to BF16 download (identical inference behavior)</p><hr /><h3 id="bf16-aio-(20-gb)-recommended-for-full-precision">🌟 BF16-AIO (20 GB) ⭐ RECOMMENDED FOR FULL PRECISION</h3><pre><code>✅ BFloat16 full precision
✅ Absolute best quality
✅ Professional projects
✅ Also works on 8GB VRAM
</code></pre><p><strong>Perfect for:</strong> Professional work, maximum quality</p><hr /><h2 id="z-image-base-aio-(28-50-steps-*-cfg-3-5)">🎨 Z-Image-Base-AIO (28-50 Steps • CFG 3-5)</h2><p><strong>Full creative control for pros &amp; LoRA training</strong></p><hr /><h3 id="fp8-aio-(10-gb)">🟡 FP8-AIO (10 GB)</h3><pre><code>✅ Efficient for daily use
✅ Full CFG control
✅ Negative prompts supported
✅ 8GB VRAM compatible
</code></pre><p><strong>Perfect for:</strong> Daily work with full control</p><hr /><h3 id="fp16-aio-(20-gb)">🔵 FP16-AIO (20 GB)</h3><pre><code>💾 Same file size as BF16
🔄 ComfyUI auto-casts to BF16 for compute
⚠️ Does NOT enable FP16 compute mode
📦 Alternative download option
</code></pre><p><strong>Note:</strong> See technical explanation in FAQ below.</p><p><strong>Perfect for:</strong> Alternative to BF16 download (identical inference behavior)</p><hr /><h3 id="bf16-aio-(20-gb)-recommended-for-full-precision">🌟 BF16-AIO (20 GB) ⭐ RECOMMENDED FOR FULL PRECISION</h3><pre><code>✅ Maximum quality
✅ Ideal for LoRA training
✅ Professional projects
✅ Highest precision
</code></pre><p><strong>Perfect for:</strong> LoRA training, professional work</p><hr /><h1 id="turbo-vs-base-when-to-use">🆚 Turbo vs Base - When to Use?</h1><hr /><h2 id="use-turbo-when:">⚡ Use TURBO when:</h2><pre><code>⚡ Speed is priority → 8 steps = 3-10 seconds
📸 Production workflows → Consistent high quality
💾 Quick iterations → Rapid prototyping
🎯 Simple prompts → Less complex scenes
</code></pre><hr /><h2 id="use-base-when:">🎨 Use BASE when:</h2><pre><code>🎨 Creative exploration → Higher diversity
🔧 LoRA/ControlNet dev → Undistilled foundation
📝 Complex prompting → Full CFG control
🚫 Negative prompts needed → Remove unwanted elements
</code></pre><hr /><h1 id="recommended-settings">⚙️ Recommended Settings</h1><hr /><h2 id="turbo-settings-(incl.-nvfp4)">⚡ Turbo Settings (incl. NVFP4)</h2><pre><code>📊 Steps: 8
🎚️ CFG: 1.0 (don't change!)
🎲 Sampler: res_multistep OR euler_ancestral
📈 Scheduler: simple OR beta
📐 Resolution: 1920×1088 (recommended)
🚫 Negative Prompt: ❌ Not used!
</code></pre><hr /><h2 id="base-settings">🎨 Base Settings</h2><pre><code>📊 Steps: 28-50
🎚️ CFG: 3.0-5.0 (start with 4.0)
🎲 Sampler: euler ⭐ OR dpmpp_2m
📈 Scheduler: normal ⭐ OR karras
📐 Resolution: 512×512 to 2048×2048
🚫 Negative Prompt: ✅ Fully supported!
</code></pre><hr /><h1 id="quick-overview">📊 Quick Overview</h1><hr /><h2 id="turbo-versions">Turbo Versions</h2><pre><code>⚫ NVFP4  │ 7.8 GB  │ RTX 50xx only  │ Max Speed 🆕
🟡 FP8   │ 10 GB   │ 8GB VRAM       │ Recommended ⭐
🔵 FP16  │ 20 GB   │ → BF16 compute │ See FAQ ⚠️
🌟 BF16  │ 20 GB   │ 8GB VRAM       │ Max Quality ⭐
</code></pre><hr /><h2 id="base-versions">Base Versions</h2><pre><code>🟡 FP8   │ 10 GB   │ 8GB VRAM       │ Efficient
🔵 FP16  │ 20 GB   │ → BF16 compute │ See FAQ ⚠️
🌟 BF16  │ 20 GB   │ 8GB VRAM       │ LoRA Training ⭐
</code></pre><hr /><h1 id="prompting-guide">💡 Prompting Guide</h1><hr /><h2 id="good-example:">✅ Good Example:</h2><pre><code>Professional food photography of artisan breakfast plate. 
Golden poached eggs on sourdough toast, crispy bacon, fresh 
avocado slices. Morning sunlight creating warm glow. Shallow 
depth of field, magazine-quality presentation.
</code></pre><hr /><h2 id="bad-example:">❌ Bad Example:</h2><pre><code>breakfast, eggs, bacon, toast, food, morning, plate
</code></pre><hr /><h2 id="tips">📝 Tips</h2><p><strong>DO:</strong></p><ul><li><p>✅ Use natural language</p></li><li><p>✅ Be detailed (100-300 words)</p></li><li><p>✅ Describe lighting &amp; mood</p></li><li><p>✅ Specify camera angle</p></li><li><p>✅ English OR Chinese (or both!)</p></li></ul><p><strong>DON'T:</strong></p><ul><li><p>❌ Tag-style prompts (tag1, tag2, tag3)</p></li><li><p>❌ Very short prompts (under 50 words)</p></li><li><p>❌ Negative prompts with Turbo</p></li></ul><hr /><h1 id="bilingual-text-rendering">🌐 Bilingual Text Rendering</h1><hr /><h2 id="english:">English:</h2><pre><code>Neon sign reading "OPEN 24/7" in bright blue letters 
above entrance. Modern sans-serif font, glowing effect.
</code></pre><hr /><h2 id=":">中文:</h2><pre><code>Traditional tea house entrance with sign reading 
"古韵茶坊" in elegant gold Chinese calligraphy.
</code></pre><hr /><h2 id="both:">Both:</h2><pre><code>Modern cafe with bilingual sign. "Morning Brew" in 
white script above, "晨曦咖啡" in Chinese below.
</code></pre><hr /><h1 id="installation">📥 Installation</h1><hr /><h2 id="step-1:-download">Step 1: Download</h2><p>Choose your version based on:</p><ul><li><p><strong>GPU:</strong> RTX 50xx → NVFP4 possible</p></li><li><p><strong>VRAM:</strong> 8GB → FP8 recommended</p></li><li><p><strong>Purpose:</strong> LoRA Training → Base BF16</p></li></ul><hr /><h2 id="step-2:-place-file">Step 2: Place File</h2><pre><code>ComfyUI/models/checkpoints/
└── Z-Image-Turbo-FP8-AIO.safetensors
</code></pre><hr /><h2 id="step-3:-load-and-generate">Step 3: Load &amp; Generate</h2><ol><li><p>Open ComfyUI (v0.11.0+!)</p></li><li><p>Use "Load Checkpoint" node</p></li><li><p>Select your AIO version</p></li><li><p>Generate!</p></li></ol><p><strong>No separate VAE or Text Encoder needed!</strong></p><hr /><h1 id="credits">🙏 Credits</h1><hr /><h2 id="original-model">Original Model</h2><pre><code>👨‍💻 Developer: Tongyi Lab (Alibaba Group)
🏗️ Architecture: Single-Stream DiT (6B parameters)
📜 License: Apache 2.0
</code></pre><hr /><h2 id="links">Links</h2><p>🔗 <strong>Z-Image Base:</strong> <a target="_blank" rel="ugc" href="https://huggingface.co/Tongyi-MAI/Z-Image">https://huggingface.co/Tongyi-MAI/Z-Image</a></p><p>🔗 <strong>Z-Image Turbo:</strong> <a target="_blank" rel="ugc" href="https://huggingface.co/Tongyi-MAI/Z-Image-Turbo">https://huggingface.co/Tongyi-MAI/Z-Image-Turbo</a></p><p>🧠 <strong>Text Encoder:</strong> <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen3-4B">https://huggingface.co/Qwen/Qwen3-4B</a></p><hr /><h1 id="version-history">📈 Version History</h1><hr /><h2 id="v2.2-fp16-clarification">v2.2 - FP16 Clarification</h2><pre><code>📝 Updated FP16 descriptions for technical accuracy
⚠️ Clarified: FP16 weights ≠ FP16 compute
🔄 FP16 files are cast to BF16 during inference
</code></pre><hr /><h2 id="v2.1-nvfp4-release">v2.1 - NVFP4 Release 🆕</h2><pre><code>➕ Z-Image-Turbo-NVFP4-AIO (7.8GB)
⚡ Optimized for NVIDIA Blackwell (RTX 50xx)
🚀 Maximum speed generation
</code></pre><hr /><h2 id="v2.0-base-aio-release">v2.0 - Base AIO Release</h2><pre><code>➕ Z-Image-Base-BF16-AIO
➕ Z-Image-Base-FP16-AIO
➕ Z-Image-Base-FP8-AIO
🔄 ComfyUI v0.11.0+ support
📝 Qwen3-4B Text Encoder
</code></pre><hr /><h2 id="v1.1-fp16-added">v1.1 - FP16 Added</h2><pre><code>➕ Z-Image-Turbo-FP16-AIO
🔧 Wider GPU compatibility
</code></pre><hr /><h2 id="v1.0-initial-release">v1.0 - Initial Release</h2><pre><code>✅ Z-Image-Turbo-FP8-AIO
✅ Z-Image-Turbo-BF16-AIO
✅ Integrated VAE + Text Encoder
</code></pre><hr /><h1 id="faq">❓ FAQ</h1><hr /><p><strong>Q: Which version should I choose?</strong></p><pre><code>RTX 50xx + Speed → NVFP4 🆕
Most users       → Turbo FP8 ⭐
Full precision   → BF16 ⭐
LoRA Training    → Base BF16
</code></pre><hr /><p><strong>Q: Turbo or Base?</strong></p><pre><code>Fast &amp; simple    → Turbo ⚡
Full control     → Base 🎨
</code></pre><hr /><p><strong>Q: Will NVFP4 work on my RTX 4090?</strong></p><p>❌ <strong>No!</strong> NVFP4 is <strong>only for RTX 50xx</strong> (Blackwell architecture).</p><p>Use FP8 instead for RTX 40xx and older.</p><hr /><p><strong>Q: Do I need separate VAE/Text Encoder?</strong></p><p>❌ <strong>No!</strong> Everything is already integrated.</p><p>Just Load Checkpoint and go!</p><hr /><p><strong>Q: Works on 8GB VRAM?</strong></p><p>✅ <strong>Yes!</strong> All versions work on 8GB VRAM.</p><p>(NVFP4 requires RTX 50xx regardless of VRAM)</p><hr /><h2 id="q:-what-about-fp16-for-older-gpus-(rtx-20003000)">⚠️ Q: What about FP16 for older GPUs (RTX 2000/3000)?</h2><p><strong>Important technical clarification:</strong></p><p>Z-Image <strong>does NOT support FP16 compute type</strong>. Here's why:</p><pre><code>📊 Technical reason:
- FP16 max value: ~65,504
- BF16 max value: ~3.39e+38 (same as FP32)
- Z-Image's activation values exceed FP16's range
- Result: Overflow → NaN → Black images
</code></pre><p><strong>What actually happens:</strong></p><ul><li><p>ComfyUI automatically casts weights to BF16 for computation</p></li><li><p>You can see this in logs: "model weight dtype X, manual cast: torch.bfloat16"</p></li><li><p>"Weight dtype" (file format) ≠ "Compute dtype" (actual calculation)</p></li></ul><p><strong>For RTX 20xx users (no native BF16):</strong></p><ul><li><p>BF16 is emulated via FP32 = slower but works</p></li><li><p>There is no way to run Z-Image in true FP16 compute</p></li><li><p>FP8 with CPU offload may be a better option for limited VRAM</p></li></ul><p><strong>TL;DR:</strong> FP16 and BF16 files behave identically during inference. Choose based on download preference, not GPU compatibility.</p><hr /><h1 id="get-started-now!">🚀 Get Started Now!</h1><p><strong>Download → Load Checkpoint → Generate!</strong></p><p><strong>Recommended versions:</strong></p><ul><li><p>🟡 <strong>FP8</strong> for most users (best size/quality balance)</p></li><li><p>🌟 <strong>BF16</strong> for maximum quality</p></li><li><p>⚫ <strong>NVFP4</strong> for RTX 50xx speed</p></li></ul><p>All versions work on <strong>8GB VRAM</strong></p><hr /><p><strong>Happy generating! 🎨</strong></p>

Z-Image-Turbo/Base-AIO

<p><strong>Z-Image-Base-AIO-FP8</strong></p>

zImageTurboBaseAIO_zImageBaseFP8.safetensors

Z-Image-Base-FP8

Z-Image-Turbo-NVFP4-AIO

Files with SHA256: 69117dc79354cabc04286ebbc49df0528f154fc1ede983896b8cc91c37233949