MOSS-Image Base vs SFT — Which Model Should You Use?

The short answer

Choose SFT for direct image generation. Choose Base when you want the un-post-trained foundation checkpoint for evaluation, LoRA training, or your own tuning work.

SFT uses the same architecture and inference pipeline as Base. It is a supervised fine-tuned descendant rather than a separate model family.

OpenMOSS shows SFT producing clearer text grouping and a more organized whiteboard layout in its published comparison. That is useful evidence for that example, not a promise that SFT wins every prompt. Test the scenes, lettering, and layouts that matter to your work.

Base vs SFT at a glance

Area	MOSS-Image Base	MOSS-Image SFT
Text rendering	Supports English, Chinese, line breaks, and text inside images; results still need proofreading.	Supports the same formats; the official comparison shows clearer text grouping in its example.
JSON layout	Supports JSON regions for text placement and object arrangement.	Uses the same architecture and pipeline and supports the same JSON layout format.
License	Apache 2.0.	Apache 2.0.
Download	Separate Base repository on Hugging Face; suitable as a foundation for LoRA or post-training work.	Separate SFT repository on Hugging Face; derived from Base and ready for direct inference.
ComfyUI setup	Load the Base weights with a matching native or HF-directory workflow.	Load the SFT weights with the same model family pipeline and a matching workflow.
Best fit	Controlled evaluation, LoRA training, research, and custom post-training.	Prompt-to-image creation, text-led graphics, and hosted or local generation.

JSON layout control is shared by both checkpoints. It is not an SFT-only feature.

How to switch Base and SFT in ComfyUI

Keep the workflow type consistent. With native ComfyUI support, place the selected checkpoint components in the documented model folders and use the native MOSS-Image workflow. With the optional Hugging Face directory plugin, point MOSSImageHFLoader at the complete local Base or SFT directory containing transformer/, text_encoder/, tokenizer/, and vae/.

Do not move between native and plugin workflows by changing only the loader node. The official ComfyUI guide says their loader nodes are not interchangeable and the native path uses a different timestep multiplier. Import the matching workflow, then compare checkpoints with the same prompt, dimensions, seed, sampler, and step count.

Which model should you use?

Start with SFT if you want to generate finished images now, especially when text grouping and organized layouts matter. The online studio on this site uses the hosted SFT endpoint, so it is the fastest way to test prompts without managing weights or a GPU.

Start with Base if your goal is to train a LoRA, benchmark the foundation checkpoint, or build your own post-training process. Base retains the pre-post-training model behavior and requires careful inference settings, which makes it useful for customization rather than the simplest default.

Use the two studio links below to open the matching model state. SFT can generate online; Base opens its local-install path because this site's hosted fal.ai endpoint serves SFT.

Official sources

Checkpoint details, licenses, and workflows checked October 24, 2026.

MOSS-Image Base model card MOSS-Image SFT model card MOSS-Image Base user guide Official ComfyUI guide