Two-stage MOSS-Image-Base workflow: 52+20 steps, 4MP-to-6MP latent handoff at 40%
Ciclosporin · reddit · 2026-11-06
A Reddit user shares a complete two-stage ComfyUI workflow for the MOSS-Image-Base diffusion model, with full parameters and a downloadable workflow JSON.
Core idea: Auryg's Krea-2-Two-Stage-Sampler keeps everything in latent space across both stages, handing off denoising at a 40% completion point with automatic latent scaling between stages.
Parameters:
- Stage 1: MOSS-Image-Base (int8 convrot) only — 52 steps, CFG 3.0, euler sampler, simple scheduler
- Stage 2: base + lodestones' 8-steps-dmd2 LoRA — 20 steps, CFG 1.0, dpmpp2m, karras scheduler
- 4MP base resolution, upscaled to 6MP between stages via bislerp
Supporting nodes: ComfyUI-MOSSLayout (BBOX layout) and kijai's ComfyUI-KJNodes (model preview). All 10 example images use seed 1 and aren't cherry-picked; prompts are embedded in the workflow.
More from Multimodal
- AI-crafted fictional game The Last Sticker Shop hides a sticker puzzle — tisch_eins · 2026-11-06
- Frameboy ships inside Viewfinder beta: build scenes frame-by-frame, explore in real time — theworldlabs · 2026-11-06
- World Labs launches Frameboy: real-time exploration of Atlas world model at up to 15 fps — theworldlabs · 2026-11-06
- Kling AI street-walk video looks so real viewers did a double take — TonyaAnne · 2026-11-06
- Kling 4.0 'kissing in the rain' video shows lifelike human interaction — TonyaAnne · 2026-11-06
- Kling 4.0 rain-walking video showcases realistic reflections and cloth physics — TonyaAnne · 2026-11-06