MOSS-Image depth control LoRA

MOSS-Image depth control LoRA

Lets a grayscale depth map steer the layout of images made with MOSS-Image-SFT in ComfyUI, while your prompt sets the subject and style. For image makers who want to keep camera depth fixed.

Made by thedeoxen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MOSS-Image depth control LoRA is an add-on for the MOSS-Image-SFT image model. It reads a grayscale depth map in which close objects are drawn white, and it keeps the picture's layout in step with that map. Your text prompt then decides the subject, materials and style.

It behaves like a ControlNet but is built as an edit adapter. It is a separate community add-on, different from the color and depth adapter that MOSS provides officially.

Who it suits: ComfyUI users who generate images with MOSS-Image-SFT and want control over near and far layout.

Industry | Who uses it | How they use it | Example
Design | Artists | Redraw a scene in another style while keeping its camera depth | Keep the layout of a photo in a new style
Product | Product designers | Hold a near subject in front while changing materials | Show a product on a clear foreground
Interior | Interior designers | Keep room layouts while restyling them | Try new finishes on the same room layout

If you do not use ComfyUI or MOSS-Image-SFT, you can skip it.

Where to use it:

• ComfyUI, with the included workflow
• On top of MOSS-Image-SFT, the base model it was built for

Use case | Who | What they get back
Restyling a scene | Artists | A new look at the same camera depth
Foreground and background control | Designers | A near subject kept apart from the background
Layout-driven images | Product and interior teams | Pictures that follow a clear near subject

How to set it up:

• ComfyUI with MOSS-Image-SFT loaded
• UNet file: moss_image_sft_int8_convrot.safetensors
• Text encoder file: qwen3.5_9b_w4a8_gptq.safetensors, set to the moss_image type
• VAE file: flux2-vae_small.safetensors
• The LoRA file moss_image_sft_depth_control_lora.safetensors from this repository
• The included ComfyUI workflow file, moss_image_depth_control.json

1. Load MOSS-Image-SFT with the three files above.
2. Attach the LoRA to the UNet with its strength at 1.0.
3. Make a depth map and feed it into the Text Encode MOSS-Image Edit node as its only reference picture.
4. In both the positive and negative encoders, choose ai-toolkit as the edit format.

pip install -U huggingface_hub
hf download thedeoxen/MOSS-Image-depth-controlnet

How to use it: Write a prompt that names the subject and style, and keep close surfaces white in the depth map. The bundled workflow does the rest. It loads a photo, produces the map with Depth Anything V2, then runs sampling. No trigger word is needed.

Sample with 30 steps and a CFG of 4, and leave the LoRA strength at 1.0.

Setting | Value
Sampler | euler with simple
Edit format | ai-toolkit

Example: You load a photo of a room, let the workflow build its depth map, and write a prompt describing the same room as a cozy cabin. The picture keeps the room's near-to-far layout while the prompt sets the materials and style.

For newcomers and experts: If you are new, begin with the bundled workflow and leave its settings as they come. Make sure close surfaces are white, since a flipped map works against the LoRA.

For experts, linear layers carry an adapter of rank 64, and convolution layers carry rank 16. Training treated it as an edit adapter, so the control image serves as the reference while the prompt describes the output. The shared checkpoint comes from step 5000, and kv cache was on during training.

Depth Anything V2 builds a grayscale depth map
The map goes in as the reference image
The prompt sets subject and style
MOSS-Image-SFT with the LoRA samples the picture

Good to know: Wide flat areas in the depth map give the prompt more freedom over texture. If the control feels stiff, try lowering the strength toward 0.7, since 1.0 is the trained value. It suits cases where overall camera depth matters more than matching every edge.

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need a ComfyUI setup that can run MOSS-Image-SFT.
• Data it reaches: downloading reaches Hugging Face. The workflow works on the photo you load.
• Not for: depth maps drawn the other way round, with near objects dark, which work against it.

On Hugging Face, 13 November 2026

• Task: image to image
• Library: diffusers
• Downloads in the last 30 days: 2,011
• Likes: 6
• License: apache-2.0
• Languages: en
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: diffusers, moss-image, controlnet, lora, comfyui, depth, image-to-image

Links

• Model: https://huggingface.co/thedeoxen/MOSS-Image-depth-controlnet
• Files: https://huggingface.co/thedeoxen/MOSS-Image-depth-controlnet/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download thedeoxen/MOSS-Image-depth-controlnet

Then follow the usage in its model card. Nothing in this kit runs by itself.