ComfyUI-PJiT runs Linum's P-JiT 21K pixel-space image model on stock ComfyUI samplers
ComfyUI-PJiT by PedroMarinhoDev adds native nodes for Linum's P-JiT 21K, a 2B-parameter text-to-image model that draws in pixel space with no VAE. It plugs into ComfyUI's own KSampler and memory manager, with bf16 and int8 ConvRot weights from about 4 GB VRAM.
What it does
Load P-JiT Model and Load P-JiT Text Encoder expose the DiT as a MODEL and Qwen3.5-4B as a CLIP for stock CLIP Text Encode. Extra nodes cover empty image latents, a prompt upsampler, APG guidance, Linum's scheduler and H100-matching noise so a seed can match Linum's samples. A refiner workflow runs a low-denoise image-to-image pass.
Worth knowing
Node pack is GPL-3.0; P-JiT 21K weights and Linum's code are Apache-2.0. Linum calls it a research checkpoint with rough edges, trained for about $21K. Needs ComfyUI v0.44+. Prefer 50–80 word captions, cfg 15 and one of the five trained sizes.
Try it yourself
A quiet Scandinavian reading nook at golden hour, linen curtains, a stack of paperbacks on a low oak table, soft dust in the light, shot on 50mm, natural colour, no text.