MOSS-Image in ComfyUI: SFT vs Turbo setup and tests

This tutorial compares MOSS-Image SFT with its optional Turbo LoRA in a ComfyUI image generation workflow. The supplied setup uses SFT INT8 ConvRot with Prompt Enhancer disabled. Regular sampling runs at 25 steps and CFG 2.5; Turbo uses eight steps and CFG 1. The speaker distinguishes these presets from the official reference settings of 50 steps and CFG 3.

Setup covers the checkpoint, Qwen3.5 9B text encoder, FLUX.2 Small VAE and Vantage-NodesV2 controls. The metadata specifies a ComfyUI nightly build for the depth workflow and the latest public release for the version without depth support. Depth guidance also needs the Color/Depth Control LoRA, Lotus model and a separate Lotus VAE.

Tests cover rainy portraits, long chalkboard passages, lighting, a conference poster and a gym layout with 13 regions. The presenter generally prefers regular SFT for natural textures and softer light, while Turbo sometimes follows placement and palette instructions more closely. These are judgments about the demonstrated outputs, rather than a universal ranking.

The bounding box editor assigns descriptions and palettes to individual regions. Lotus supplies depth maps for reusing a reference composition, but the workflow does not preserve identity or clothing. Its separate color-map input remains unconnected.