Source: https://docs.sogni.ai/models/moss-image/

# MOSS-Image SFT

MOSS-Image SFT is OpenMOSS's uncensored 5-billion-parameter text-to-image model. It supports English and Chinese captions, tag prompts, and experimental JSON layouts that describe object positions and text placement. Sogni hosts the finished SFT generation checkpoint with INT8 ConvRot weights, a Qwen3.5 9B text encoder, and the FLUX.2 VAE.

SFT means supervised fine-tuned: this is a generation model. Training and the separate Base checkpoint are not part of the hosted workflow.

Both the standard and 8-Step presets support uncensored generation, including private mature-theme creativity.

## [#](https://docs.sogni.ai/models/moss-image/#select-the-model)Select the model

| Surface | Selector |
| --- | --- |
| Model picker | MOSS-Image SFT |
| Direct SDK project | `moss_image_sft_int8_convrot` |
| Hosted `generate_image` | `moss-image` |
| Creative Agent CLI | `-m moss_image_sft_int8_convrot` |

MOSS-Image accepts text only. It does not accept guide images, photos, identity references, or masks. Changing a JSON layout creates a new image; it is not an instruction edit of an uploaded image. Use an image-editing model when an existing person, object, or photograph must guide the result.

## [#](https://docs.sogni.ai/models/moss-image/#generation-settings)Generation settings

| Setting | Default | Supported controls |
| --- | --- | --- |
| Canvas | 1024 × 1024 | 512–2048 pixels per edge, multiples of 16 |
| Steps | 50 | 50–100 |
| Guidance | 3 | 3–6 |
| Sampler | Euler | `euler` |
| Scheduler | SGM Uniform | `sgm_uniform` |
| Flow shift | 5 | Fixed by the workflow |
| Skip-layer guidance | Negative branch, layer 22 | Fixed by the workflow |
| Negative prompt | Empty | Optional text |

This is a full-step model. Keep its 50-step default when evaluating quality. The author recommends approximately 1–2 megapixels; 512 × 512 and 2048 × 2048 are extrapolated sizes. Larger canvases and more steps increase the quoted cost. The app or SDK resolves those settings before pricing the project.

## [#](https://docs.sogni.ai/models/moss-image/#eight-step-preset)Eight-step preset

Choose **MOSS-Image 8-Step** for the DMD recipe by lodestones. It uses the same SFT checkpoint with the pinned distillation adapter at strength 1.0.

| Surface | Eight-step selector |
| --- | --- |
| Direct SDK project | `moss_image_sft_dmd8_int8_convrot` |
| Hosted `generate_image` | `moss-image-8step` |
| Creative Agent CLI | `-m moss_image_sft_dmd8_int8_convrot` |

Use exactly **8 steps, guidance 1, Euler and SGM Uniform**. Flow shift remains 5; the negative prompt and skip-layer guidance are inactive. Canvas bounds are the same 512–2048 pixels per edge. Both presets run on CUDA 13 workers with at least 24 GB VRAM; the RTX 4090 passed the 1024² and 2048² tests.

In matched RTX 4090 lab runs with weights loaded, 1024² took a median 3.32 seconds at eight steps and 23.79 seconds at fifty steps. Cold startup includes downloads and model loading. DMD changes composition and texture, even with the same seed; inspect fine detail, gradients and exact writing before choosing a result.

The [DMD author](https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd) labels the adapter work in progress. Sogni uses the recommended `lora.safetensors` at strength 1.0; the author's alternative `lora_latest.safetensors` is not this preset.

## [#](https://docs.sogni.ai/models/moss-image/#personal-loras)Personal LoRAs

Import a compatible MOSS adapter through **My LoRAs** and select it on either MOSS preset. Ready imports are shared across the two presets. On the eight-step preset, the personal adapter follows the built-in DMD adapter; its strength control changes only your personal adapter. Results can vary with the checkpoint used to train an adapter. The image default is strength 1.0; use the entry's published range and prompt requirements.

See [Personal LoRA API](https://docs.sogni.ai/api-reference/personal-loras/) for account requirements, import validation, owned IDs and ordered strength arrays. This feature imports existing adapters; it does not train a model.

## [#](https://docs.sogni.ai/models/moss-image/#write-a-prompt)Write a prompt

Use a complete caption: subject, scene, action, camera, materials, lighting and the exact visible text. Keep requested writing inside quotes. Tag-style prompts are also supported.

```
A clean product photograph of a cobalt-blue ceramic mug on pale stone.
The mug reads "MORNING TIDE" in white uppercase letters. Soft window light
from the left, visible glaze texture, a crisp handle, and a gentle shadow.
```

For JSON layouts, the author uses `summary` and `items` together with `bbox_format: "yxyx"` and `bbox_coord_space: "abs_centralized"`. Bounding boxes are `[y_min, x_min, y_max, x_max]` in pixels relative to the canvas center. At 1024 × 1024, the canvas runs from −512 to +512 on each axis. Keep the whole layout in the positive prompt. Layout control is experimental, so inspect the result for text accuracy, composition and object counts.

## [#](https://docs.sogni.ai/models/moss-image/#source-and-license)Source and license

MOSS-Image SFT is released under Apache 2.0. See the [official SFT model card](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT), [author's guide](https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md), and [ComfyUI weights](https://huggingface.co/Comfy-Org/MOSS-Image).