THE MOSS-IMAGE FIELD NOTES
Compare OpenMOSS MOSS-Image Base and SFT, choose Diffusers or ComfyUI component files, and understand the checkpoint and settings used by this online studio.
Updated October 24, 2026 · Independent guide · Official sources linked belowMOSS-Image-Base is OpenMOSS’s base text-to-image checkpoint. MOSS-Image-SFT is a supervised fine-tuned derivative with the same architecture and inference pipeline. For the text and layout cases on this site, start with SFT; choose Base when a training project or adapter specifically expects it. Match the download format to your runtime.
| Question | MOSS-Image-Base | MOSS-Image-SFT |
|---|---|---|
| What is it? | The base checkpoint in the model family. | A supervised fine-tuned model derived from Base. |
| Input formats | Natural-language prompts, tag sequences and structured layouts. | The same documented input families. |
| Practical starting point | Use when a training recipe or adapter was built for Base. | Use to follow this site’s hosted generation cases. |
| Published comparison | Reference output in the model cards. | The publisher’s selected whiteboard example shows clearer text grouping; that is not a universal benchmark. |
| Official model card | Base downloads and instructions | SFT downloads and instructions |
OpenMOSS lists a 5B diffusion backbone and Apache 2.0 licensing for the model releases. The complete inference pipeline also uses a text encoder and VAE. Neither the parameter count nor a license label guarantees that a generated image will follow every instruction or be suitable for a particular commercial use. Read the source model card and review your output.
For prompt construction, use the real prompt examples. For checkpoint-specific adapters, check the LoRA compatibility list before changing models.
Use the Comfy-Org repackaged files: one diffusion model, one text encoder and one VAE. These go into separate ComfyUI model folders. A single diffusion safetensors file is not the complete pipeline.
Exact filenames and model foldersUse the original Base or SFT repository with its transformer, text encoder, tokenizer, VAE and configuration. This matches the Python pipeline or OpenMOSS HF-directory plugin, not the native component selectors.
HF-directory plugin setupThe workflow download chooser provides separate graphs for these routes. Import a graph that matches the files you downloaded. Changing one loader inside an old plugin graph does not convert the rest of its sampling chain into a native workflow.
The Comfy-Org listing includes BF16 and INT8 diffusion models and several encoder variants. Precision and quantization affect hardware support and memory use, and can affect adapter compatibility. Begin with the files selected by a matched graph. Do not rename one precision variant to impersonate another filename.
The published standalone Base/SFT reference uses 50 steps, guidance 3, shift 5, flow-matching Euler and negative-branch layer skipping at layer 22. The current pinned native ComfyUI template uses 25 steps, CFG 2.5 and er_sde / simple. Follow the route-specific settings table, not a mixture of the two.
VRAM depends on the full pipeline, precision, offload, canvas size and batch size. A model file’s disk size is not a measured minimum VRAM requirement. This site has not benchmarked those local configurations; it publishes real hosted results with their actual API settings instead.
Community “turbo” or few-step adapters are separate configurations. Their author’s checkpoint, strength, precision and sampler requirements take precedence over a generic 50-step recipe. See the MOSS-Image turbo LoRA explanation.
This independent studio calls the MOSI API with model moss-image-sft-1.0, 50 inference steps and guidance 3. Standard output has a 1024-pixel long edge; the layout canvas is 1024 × 1024. These are this service’s current settings, not a claim that every MOSS-Image host uses the same configuration.
Try the real CITRUS product-label input, then inspect the original image and observed differences.
Use this promptThe button fills the prompt and settings on the homepage. It does not generate or deduct credits. Local Base/SFT selection and custom LoRA uploads are separate from this hosted service.
Checked October 24, 2026 against the Base model card, SFT model card, OpenMOSS ComfyUI instructions and Comfy-Org model listing. For local inference, continue to the matching workflow download; for layout input, use the layout prompt guide.