This repackaging makes it straightforward to run the MOSS-Image family inside ComfyUI so you can generate layout-aware, text-rich images locally without cloud APIs. It bundles both Base and instruction-tuned SFT checkpoints plus compatible text encoder and VAE builds so users can switch checkpoints or run lower-VRAM int8 variants in the same ComfyUI graph.
What Sets It Apart
- ComfyUI-ready file layout and templates: supplies int8/bf16 diffusion checkpoints (Base and SFT), Qwen3.5 text encoder builds, and FLUX.2 VAEs so the model runs natively in ComfyUI templates. This means less manual conversion and fewer compatibility headaches when integrating MOSS-Image into ComfyUI workflows.
- Native layout and long-prompt support preserved: the underlying MOSS-Image pipeline supports JSON-style bounding-box layouts, bilingual (English/Chinese) text rendering, and prompts up to ~4K tokens, so complex scene composition and on-image typography remain available in the ComfyUI environment.
- Instruction-tuned SFT checkpoint included: templates default to the SFT checkpoint for cleaner, layout-consistent outputs while the Base checkpoint is available for fine-tuning and LoRA work, enabling both end-user generation and researcher experimentation.
- Licensing and portability: weights are distributed under Apache-2.0 and the package includes both high-precision and low-VRAM variants, so you can run models locally on different GPU setups without per-image credits.
Who It's For and Tradeoffs
Great fit if you want to run MOSS-Image locally in a visual node-based builder (ComfyUI), need explicit layout control or on-image text rendering, or want ready-made int8/bf16 checkpoints for low-VRAM experimentation. Look elsewhere if you need an out-of-the-box web service, turnkey GUI beyond ComfyUI, or minimal setup for non-technical users—this repackaging reduces friction but still expects familiarity with ComfyUI and model files. Also note that the ComfyUI templates require version 0.41.3 or later and that high-fidelity bf16 runs need more VRAM than int8 variants.
