<h1 id="z-image-turbo-fp8-hires-workflow-(low-vram-optimized)">🚀 Z-Image Turbo FP8 Hires Workflow (Low VRAM Optimized)</h1>This is a high-efficiency ComfyUI workflow designed specifically for Low VRAM users. By utilizing FP8 Quantized Models and Latent Upscale technology, it generates high-resolution images (1024x1792) rapidly while maintaining minimal resource usage.<h2 id="key-features">✨ Key Features</h2><ul><li>Extreme Low VRAM Usage: Full FP8 pipeline (Model &amp; Text Encoder) to drastically reduce memory footprint.</li><li>Lightning Fast: Optimized for Turbo models and efficient sampling steps.</li><li>Hires Fix Pipeline: Utilizes <code>Latent Upscale</code> + <code>2nd Pass KSampler</code> to ensure crisp details without heavy VRAM cost.</li><li>AuraFlow Architecture: Optimized using the <code>ModelSamplingAuraFlow</code> node.</li></ul><hr /><h2 id="models-required-and-downloads">📂 Models Required &amp; Downloads</h2>To ensure the workflow functions correctly, please download the following models and place them in your respective ComfyUI folders:<h3 id="1.-unet-model-(place-in-modelsunet)">1. UNet Model (Place in <code>models/unet/</code>)</h3><ul><li>File Name: <code>z-image-turbo-fp8-e4m3fn.safetensors</code></li><li>Download: <a target="_blank" rel="ugc" href="https://huggingface.co/T5B/Z-Image-Turbo-FP8/blob/main/z-image-turbo-fp8-e4m3fn.safetensors">HuggingFace - Z-Image-Turbo-FP8</a></li></ul><h3 id="2.-clip-text-encoder-(place-in-modelsclip)">2. CLIP / Text Encoder (Place in <code>models/clip/</code>)</h3><ul><li>File Name: <code>qwen3-4b-fp8-scaled.safetensors</code></li><li>Download: <a target="_blank" rel="ugc" href="https://huggingface.co/jiangchengchengNLP/qwen3-4b-fp8-scaled">HuggingFace - Qwen3-4B-FP8</a></li></ul><hr /><h2 id="key-settings-and-configuration">⚙️ Key Settings &amp; Configuration</h2>This workflow operates on a 2-Pass system. Please adhere to the following settings for the best results:<h3 id="phase-1:-base-generation">🔹 Phase 1: Base Generation</h3><ul><li>Latent Size: Generates at a lower initial resolution (e.g., 512x896) to save compute resources.</li></ul><h3 id="phase-2:-latent-upscale">🔹 Phase 2: Latent Upscale</h3><ul><li>Upscale Method: Uses <code>LatentUpscaleBy</code>.</li><li>Scale Factor: Default is <code>2</code> (resulting in a final output of 1024x1792).</li></ul><h3 id="phase-3:-hires-fix-(refiner)">🔹 Phase 3: Hires Fix (Refiner)</h3>This step is crucial for image clarity and detail:<ul><li>Sampler: <code>res_multistep</code> (Highly Recommended).</li><li>Denoise: Recommended range <code>0.5 - 0.6</code>.<ul><li><code>&lt; 0.5</code>: Changes are minimal; the image may remain slightly blurry.</li><li><code>&gt; 0.6</code>: Adds more detail, but setting this too high may alter the image structure or cause hallucinations.</li></ul></li></ul><hr /><h2 id="performance-benchmark">📊 Performance Benchmark</h2>Data based on actual testing:GPUOutput ResolutionTimeNVIDIA RTX 5070 Ti1024 x 17928 ~ 9 sec<hr /><h2 id="usage-tips">📝 Usage Tips</h2><ol><li>Memory Management: If you are extremely limited on VRAM, ensure no other large models are loaded in the background.</li><li>Prompting: Since this uses the Qwen text encoder, it has strong natural language understanding. Detailed, sentence-based prompts work very well.</li><li>Troubleshooting: If you notice the image details breaking or looking "burnt," try slightly lowering the <code>denoise</code> value in the second KSampler.</li></ol>

Files with SHA256: e26df5cce44c682340c1339552d19349a053ceb034c034909ff0af1a9a4ce2e1