MOSS Toolkit

Runs ComfyUI's MOSS-Image DiT with an fp32 residual stream, modulation, norms, embedders and head around the block GEMMs (bf16 or INT8). No weight is reloaded…

Image-to-image for MOSS-Image: a LoRA loader, a text encode that puts the page through Qwen3.5's vision tower, and a MODEL patch that feeds the page as a clean VAE reference latent on RoPE's reserved r axis. Works under KSampler and the MOSS Spectrum KSampler. Needs the moss-image node.

MOSS i2i

model

Runs ComfyUI's MOSS-Image DiT with an fp32 residual stream, modulation, norms, embedders and head around the block GEMMs (bf16 or INT8). No weight is reloaded, so it works with dynamic VRAM, LoRAs and the INT8 files.

MOSS-Image fp32 Stream

Inputs (1)

Outputs (1)