<div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/kWVfLtxJk8c" start="0"></iframe></div>A 1-click, 4-step upscaler built around NVIDIA's PiD (Pixel Diffusion) model on the Flux1 backbone. Feed it any image and it intelligently regenerates real, fine detail at 4× resolution — then supersamples it back down to a razor-sharp final image. An automatic captioner keeps the upscale faithful to your image's content, so you get crisp detail without hallucinated junk.<h2 id="what-makes-this-workflow-different">✨ What makes this workflow different</h2><ul><li>Works with ANY aspect ratio — 16:9, 9:16, square, anything. The input is auto-normalized to the model's native 1024px long-edge (e.g. a 1280×720 image becomes 1024×576), which is the #1 cause of broken/green-tinted PiD results. This is handled for you.</li><li>Auto-prompting via Florence-2 — a vision model reads your image and writes a detailed caption automatically, guiding the upscaler to preserve the actual content. No manual prompting needed.</li><li>Locked to the model's true 4× regime — PiD <code>1024→4096</code> is a fixed 4× model; this workflow targets exactly that for maximum sharpness (no blur from under/over-scaling).</li><li>Supersampled final output — the 4096px result is Lanczos-downscaled back to 1024px, baking all 4× of generated detail into a clean, antialiased image with zero quality loss. You also get the full 4096px version saved.</li><li>Side-by-side comparison — a built-in slider compares your original against the result.</li><li>Fast — only 4 sampling steps (LCM).</li></ul><h2 id="how-to-use">🔧 How to use</h2><ol><li>Install the 3 model files (links + folder layout below).</li><li>Load the workflow, drop your image into the Load Image node.</li><li>Hit Queue. That's it.</li></ol>Outputs:<ul><li>Full 4× image (e.g. 4096×2304) — saved via the first Save node.</li><li>Supersampled 1024 image — crisp, detail-packed, saved via the second Save node.</li></ul>Tip: The PiD model is brightest/cleanest on near-square and landscape images. Extreme portrait crops can show mild color shifts — that's a known characteristic of the current distilled model, not the workflow.<h2 id="downloads-links-below:">📥 Downloads Links Below:</h2>Gemma 2b: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/PixelDiT/tree/main/text_encoders">https://huggingface.co/Comfy-Org/PixelDiT/tree/main/text_encoders</a> PiD models: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/PixelDiT/tree/main/diffusion_models">https://huggingface.co/Comfy-Org/PixelDiT/tree/main/diffusion_models</a> VAE: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/vae">https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/vae</a>📂 ComfyUI/├── 📂 models/│ ├── 📂 text_encoders/│ │ └── gemma_2_2b_it_elm_bf16.safetensors│ ├── 📂 vae/│ │ └── ae.safetensors│ └── 📂 diffusion_models/│ └── pid_flux1_1024_to_4096_4step_bf16.safetensors<h2 id="required-custom-nodes">📋 Required custom nodes</h2><ul><li>ComfyUI-Florence2 (auto-captioning)</li><li>ComfyUI-Custom-Scripts (pythongosssss — ShowText / MathExpression)</li><li>ComfyUI-easy-use</li><li>rgthree-comfy (Image Comparer)</li><li>ComfyUI_Swwan (GetImageSizeAndCount)</li></ul>Requires a recent ComfyUI build with PixelDiT / PiD support. ~16GB VRAM recommended for the full 4096px pass.<a target="_blank" rel="ugc" href="https://www.youtube.com/@AiMotionStudio">https://www.youtube.com/@AiMotionStudio</a>

Files with SHA256: 8206ee6537c6b5a991f1267940ac151af0038c8d4b9c4b5913c4ceb3469d34fd