---
license: apache-2.0
base_model: OpenMOSS-Team/MOSS-Image-SFT
base_model_relation: adapter
language:
- en
pipeline_tag: image-to-image
tags:
- moss-image
- controlnet
- lora
- comfyui
- depth
library_name: diffusers
---

# MOSS-Image-SFT — Depth Control LoRA

## Short description

A depth Control LoRA for **MOSS-Image-SFT**. Feed a grayscale depth map, near is white. The model follows distance and layout, and the prompt sets the subject and style.

This is a ControlNet-style **edit LoRA**, not the official MOSS color/depth concat adapter. Load it with a LoRA loader and pass one control image through **Text Encode MOSS-Image Edit**.

Selected checkpoint: training step **5000**.

## Files

- **Weights:** [moss_image_sft_depth_control_lora.safetensors](https://huggingface.co/thedeoxen/MOSS-Image-depth-controlnet/blob/main/moss_image_sft_depth_control_lora.safetensors)
- **ComfyUI workflow:** [moss_image_depth_control.json](https://huggingface.co/thedeoxen/MOSS-Image-depth-controlnet/blob/main/moss_image_depth_control.json)
- **Repo:** [thedeoxen/MOSS-Image-depth-controlnet](https://huggingface.co/thedeoxen/MOSS-Image-depth-controlnet)

## Examples

Each preview is one combined image, left to right: **control**, then three prompts.

| Control → Results |
| --- |
| ![](images/figure.jpg) |
| ![](images/stag.jpg) |
| ![](images/espresso.jpg) |
| ![](images/ramen.jpg) |
| ![](images/subway.jpg) |


## Extended description

Follows near-to-far layout from a grayscale depth map. Appearance, materials, clothing and the world come from the text prompt.

The LoRA is rank 64 on linear layers and rank 16 on convolutions. It was trained as an edit adapter: the control image is the reference, and the prompt describes the picture to generate.

## How to use

1. Load **MOSS-Image-SFT**: UNet `moss_image_sft_int8_convrot.safetensors`, text encoder `qwen3.5_9b_w4a8_gptq.safetensors` with type `moss_image`, VAE `flux2-vae_small.safetensors`.
2. Load this LoRA on the UNet at strength **1.0**.
3. Build the control map and pass it as the single reference image in **Text Encode MOSS-Image Edit**.
4. Set **edit format** to `ai-toolkit` on both the positive and the negative encode. This LoRA was trained with kv cache on.
5. Sample with **30 steps**, **CFG 4**, `euler` + `simple`.

No trigger word. The control image sets the structure. The prompt sets identity, materials, clothing and the world.

## ComfyUI

The workflow in this repo loads a photo, builds the control map, encodes it with the prompt, and samples.

The included workflow builds the map with **Depth Anything V2** (`depth_anything_v2_vitl`). Near surfaces are white, the background falls to black, as in the training controls.

Also uses the built-in **Text Encode MOSS-Image Edit** node. Set the edit format to `ai-toolkit`.

### Base model

Trained on and recommended with **[OpenMOSS-Team/MOSS-Image-SFT](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT)**.

Typical settings from the included workflow: **30 steps**, **CFG 4**, `euler` + `simple`, LoRA strength **1.0**.

Negative prompt used in training samples:

```
nsfw, worst quality, bad anatomy, watermark, jpeg artifact
```

### Example prompt

```
Dramatic low angle, face and torso huge. A chunky voxel person built from primary-color bricks, one brick hand reaching toward the camera, the other arm thrown out. White gallery, soft studio light, sharp detail.
```

## What it does

- Follows near-to-far layout from a grayscale depth map.
- Lets the prompt define identity, materials, clothing and style.
- Works as a single-image control adapter for MOSS-Image-SFT.

## Tips

- Keep near = white. An inverted map fights the LoRA.
- Large flat depth regions leave the prompt more room for texture.
- Use it when the camera depth matters more than every edge.
- Lower the LoRA strength toward 0.7 if the control is too rigid. Strength 1.0 is the trained value.

## Use cases

- Rebuilding a scene at the same camera depth in a new style.
- Holding foreground and background apart while changing materials.
- Product and interior layouts with a clear near subject.