Tired of the MOSS Image SFT's aggravated GPT look? You are at the right place
This is a by-product of coping with bumpy moss image reference training attempts

Re-sample your MOSS Image SFT renders to reduce the glossy, GPT-style look: less sheen and blush, flatter shading, more defined line work. The character and composition stay mostly the same.

This is a second pass, not a normal style LoRA. Generate with MOSS SFT first, then run the image through this LoRA with the instruction below. Loading it during normal text-to-image changes almost nothing.

How to use

No custom nodes are needed. Use a current ComfyUI that includes the Text Encode MOSS-Image Edit node.

1. Put this LoRA in models/loras/ and name the file ungpt_v1.safetensors.

2. Get the MOSS SFT model, text encoder and VAE from Comfy-Org/MOSS-Image.

3. Download a workflow from GitHub and drag it into ComfyUI:

   • moss_sft_t2i_then_ungpt.json: from prompt to final image, both passes in one queue.

   • ungpt_resample_image.json: re-sample an image you already have.

Settings to keep

These are the settings the results below were measured with. The workflows already use them.

• Instruction: Redraw this image as hand-drawn anime art without the AI-generated look. Keep the character, pose, outfit and framing.

• Negative: empty prompt, with the same image attached

• Edit format: ai-toolkit (kv_cache off)

• LoRA strength: 1.0

• Sampling: 50 steps, CFG 3, euler, sgm_uniform, denoise 1.0

• Guidance: Skip Layer Guidance DiT Simple, single_layers 22, double_layers empty

• Size: about 1 MP, sides in multiples of 16, output the same size as the input

The re-sample starts from empty noise and is guided by the input image, so denoise stays at 1.0. It is not a low-denoise img2img pass.

On an RTX PRO 6000, text-to-image took about 20 s and the re-sample about 46 s, with about 28 GB of VRAM in use. Smaller cards and the quantized model files were not tested.

Results

Tested on 24 text-to-image prompts with one seed each. The score comes from a classifier that separates gpt-image-2 images from hand-drawn anime art (lower is less GPT-like, under 0.5 reads as hand-drawn).

• MOSS SFT text-to-image: median 0.927, 3 of 24 under 0.5

• With the LoRA loaded during text-to-image: median 0.909, 5 of 24 under 0.5

• Re-sampled without the LoRA: median 0.952, 4 of 24 under 0.5

• Re-sampled with the LoRA: median 0.452, 13 of 24 under 0.5

The re-sample lowered the score on all 24 images. The same character was kept in 24 of 24 and the pose in 22 of 23 where a pose could be detected. The classifier never saw MOSS images in training, so treat its scores as a guide and judge the images yourself.

Limits

• The effect varies. On 4 of the 24 images the score moved by 0.04 or less. The last sample image is the smallest change in the test.

• Details can change. Expect small changes to hair, faces and expressions, and check any text in your output.

• It adds a look, not an artist. It moves images toward a generic hand-drawn style and does not recover any specific artist's style.

• gpt-image-2 input is less reliable. The median score fell from 0.890 to 0.013 on 36 test images, but the pose was kept in only 83%. Colour casts can get stronger, and framing and faces can drift more than on MOSS images.

• Not tested: the MOSS Image Base model, more than one seed per prompt, other LoRA strengths or samplers, and input from other models.

The classifier score is not an AI detector, and this LoRA has not been tested against one.

Full details, workflows and license (Apache-2.0): github.com/bag3labo/moss-image-ungpt