MOSS-Image Normal Control LoRA

MOSS-Image Normal Control LoRA

Keeps an object's shape and volume from a colour normal map while the prompt changes its material and scene, as a normal control LoRA for MOSS-Image-SFT in ComfyUI, for people who retexture and relight forms.

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: This is a normal-map control LoRA for the MOSS-Image-SFT image model. You give it a colour normal map, and it follows the surface direction and volume while your prompt sets the material and the scene. It is an edit-style LoRA, not the official MOSS colour or depth adapter. The author picked the checkpoint from training step 5000.

Who it suits: People who retexture or relight a fixed form in ComfyUI will find it fits. It also suits product and 3D artists who already have normal maps and want the volume to stay put.

Industry | Who uses it | How they use it | Example
3D art | 3D artists | Change a form's material while its volume stays fixed | The same form in a new material
Product imagery | Marketers | Show a product shape in new materials and scenes | The same product volume in a new setting
Visual effects | Compositors | Relight and retexture a solid form | A form lit in a different scene
Research | Image-model tinkerers | Test normal-map control on MOSS-Image-SFT | Comparing smooth and noisy normals

If you do not use ComfyUI or MOSS-Image-SFT, you can skip it, since it is trained for that base model.

Where to use it:

• ComfyUI, with the workflow file included in the repository
• ComfyUI's stock node for encoding MOSS edit prompts
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Retexture a form | 3D artists | A new surface material on the same volume
Product shots | Marketers | Pictures where the object volume is fixed
Relight a scene | Compositors | Depth and surface facing that stay consistent

How to set it up: You need ComfyUI, the MOSS-Image-SFT files and this LoRA. The author names three base files and a workflow.

• ComfyUI with the BAE Normal Map node the workflow uses
• MOSS-Image-SFT: the UNet moss_image_sft_int8_convrot.safetensors
• A Qwen3.5 text encoder file for MOSS-Image, with the moss_image type chosen when you load it
• The VAE flux2-vae_small.safetensors

1. Install the Hugging Face hub package.
2. Download this repository's files, which hold the LoRA and the workflow.
3. Load the MOSS-Image-SFT UNet, text encoder and VAE.
4. Apply this LoRA to the UNet with a strength of 1.0.

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

How to use it: Make a colour normal map from your photo, then hand it to the Text Encode MOSS-Image Edit node as the only reference image. On the positive and the negative encode alike, switch the edit format to ai-toolkit. Sample with 30 steps, CFG 4, and euler with simple. There is no trigger word.

Tool | What it does
BAE Normal Map node | Builds the colour normal map the LoRA expects
Text Encode MOSS-Image Edit | Takes the control image and your prompt
LoRA loader | Applies this LoRA at strength 1.0
Included workflow | Loads a photo, builds the map, encodes, and samples

Example: You load a photo of a solid object, build its normal map and write a prompt for the same form in brushed metal on a studio table. The result keeps the volume and facing from the map, while the prompt supplies the material and the scene.

For newcomers and experts: A normal map is a colour picture that records which way each part of a surface faces. A good first step is to run the included workflow on one photo with the default settings, then change only the prompt.

Its adapter weights use rank 64 for the linear layers and rank 16 for the convolutions. It learned as an edit adapter. The control picture serves as the starting reference, and the text describes the image you want out. It was trained with the kv cache on, which is why the edit format is ai-toolkit. Feed a camera-space normal map, never a grayscale height map. If the control feels too rigid, lower the strength toward 0.7.

You load a photo into the workflow
The BAE Normal Map node builds a colour map
The map and your prompt go to the encode node
MOSS-Image-SFT with the LoRA samples the new image

Good to know:

• License: apache-2.0, as listed on Hugging Face
• Cost to run: none named
• Data it reaches: the images and prompts you load into ComfyUI, and the source names no web or account access
• Smooth normals give cleaner volumes, while noisy ones add unwanted bumps
• Not for: other base models, since it was trained on and is recommended with MOSS-Image-SFT
• It is not the official MOSS colour or depth adapter

On Hugging Face, 13 November 2026

• Task: image to image
• Library: diffusers
• Downloads in the last 30 days: 1,014
• Likes: 4
• License: apache-2.0
• Languages: en
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: diffusers, moss-image, controlnet, lora, comfyui, normal, image-to-image

Links

• Model: https://huggingface.co/thedeoxen/MOSS-Image-normal-controlnet
• Files: https://huggingface.co/thedeoxen/MOSS-Image-normal-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-normal-controlnet

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