P-JiT 21K: 2B pixel-space text-to-image model trained for $21K gets native ComfyUI port
Apprehensive_Sky892 · reddit · 2026-11-06
Linum released P-JiT 21K, a 2B-parameter text-to-image model trained for $21K that draws directly in pixels with no VAE (Apache-2.0). PedroMarinho ported it to ComfyUI as native nodes: stock KSampler support (euler, simple, 50 steps, cfg 15), a bf16 file (4.4GB) pixel-identical to fp32, an int8 ConvRot version (2.3GB, 1.8x faster), a Qwen3.5-4B text encoder down to 3.2GB, and it runs from 4GB of free VRAM. Workflows include text-to-image, a prompt upsampler, and a refiner (use denoise 0.1-0.2 — P-JiT's noise is 4x stronger than most models). Verified against Linum's code: identical tokens, noise, schedule and guidance.
More from Multimodal
- World Labs world model demo shows actual moving people in generated scenes — andrew_n_carr · 2026-11-06
- Grimes stars in first AI film to hit theatres, gunning for Best Animated Feature Oscar — TomLikesRobots · 2026-11-06
- AI-generated video now looks like full TV show episodes — Additional_Judge_337 · 2026-11-06
- AI-generated Willem Dafoe Green Goblin shots show how far image models have come — egeberkina · 2026-11-06
- Is LTX 2.5 broken in ComfyUI? Users report eerie, degraded outputs — brocolongo · 2026-11-06
- How to caption datasets for motion LoRA training on Minimax H3? — Ambitious_Corgi5723 · 2026-11-06