An ecosystem grew around the open-source MOSS-Image in just 4 days

Four days after the release of OpenMOSS's open 5B image generation model MOSS-Image (Apache-2.0), the community released the Kromoss fine-tune at 1024px, a 4-step turbo-LoRA, a set of nodes for ComfyUI, and support in Fizgig 8.5.

Compiled by Sergey KostenchukPublished 2026-10-20Updated 2026-10-20

2026-10-20 Multimodal HuggingFace

🎨 An ecosystem grew around the open-source MOSS-Image in just 4 days

Following the release of MOSS-Image (OpenMOSS, 5B, Apache-2.0), the lodestones/Kromoss fine-tune at 1024px and a 4-step turbo-LoRA were released: ~11 s for 1024x1024 on an A100. For ComfyUI, the Spectrum nodes (x2–2.5 sampling), a prompt enhancer, and a control adapter were released.

🌍 The 5B model with strong text rendering and JSON layout gained an ecosystem on par with FLUX/SDXL in days: distillation down to 4–8 steps speeds up iterations in production (banners, posters), and the "model + official prompt compiler" pattern is moving to open-source.

👤 In ComfyUI, it is enough to download Kromoss (int8, 5.4 GB) or the 4-step LoRA. The numbers: 1024px natively, 1168px in ~30 s on an RTX 3060 12GB, ~11 s on an A100 in 4 steps. Custom styles can be trained on 12GB VRAM (Fizgig 8.5). The quality of the distillation is currently confirmed only by paired images.

Source 1: https://huggingface.co/lodestones/Kromoss Source 2: https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps

Sources