Descriptions:
Veteran AI walks through PiD — Pixel Diffusion Decoder — a new image upscaling model that operates natively in pixel space rather than the latent space used by standard diffusion workflows. The video opens by contrasting PiD with NVIDIA’s PixelDiT to ground the core concept: most image generation pipelines encode into latent space for efficiency, then decode with a VAE, which can compress or smear fine textures, text edges, and small objects. PiD bypasses that bottleneck by running diffusion directly on pixels, enabling genuine detail reconstruction rather than interpolation.
The practical tutorial covers running PiD inside ComfyUI, which now officially supports the model via the Comfy-Org repository on Hugging Face. Viewers learn to select the correct model variant for their base (Flux one or SD3) and target resolution (e.g., 1024→4096 for Flux), configure the Gemma 2 text encoder with PixelDiT CLIP type, and set up two critical nodes: Context Window (size 2048, overlap 512, pyramid fusion) for tiling the large output, and PiD Conditioning for controlling the degrade_sigma parameter, which determines how aggressively the model restores degraded content. Sampling runs LCM at four steps with CFG=1, making generation fast but with a characteristic plastic-skin texture on some images.
The video also demonstrates workflows on RunningHub, an online ComfyUI platform. Key limitations noted: PiD currently only accepts Flux and SD3 latents, making it a specialized tool rather than a universal upscaler, and the LCM sampling style is not always appropriate for photorealistic subjects.
📺 Source: Veteran AI · Published May 29, 2026
🏷️ Format: Tutorial Demo
