Introduction

If your MOSS-Image SFT anime renders keep coming out with that glassy sheen, the rosy blush and soft airbrushed shading, this one's for you. Developer bag3labo released un-GPT v1 on GitHub on Nov 2 (release at 10:53 UTC). It's an edit LoRA for MOSS-Image SFT that re-samples a finished image toward a flatter, more hand-drawn anime look. Line work gets more defined, and the character, pose and framing mostly stay the same.

It's open weights under Apache-2.0 (the same license as MOSS-Image), a 149 MB ungpt_v1.safetensors file. It runs in ComfyUI with no custom nodes, using the core Text Encode MOSS-Image Edit node. Two ready-made workflows come in the repo.

How it works

The key point: un-GPT is a second pass, not a style LoRA you load during text-to-image.

  • Generate with MOSS-Image SFT as usual
  • Feed that image back through MOSS-Image's edit mode with the LoRA loaded and one fixed instruction
  • The re-sample starts from empty noise and is guided by your input image, so denoise stays at 1.0. It isn't a low-denoise img2img trick.

The author tested exactly that. Loading the LoRA during plain text-to-image barely changes anything, because training used prior regularization on every other step to keep plain text-to-image stable.

The numbers

bag3labo scored outputs with a CLIP-embedding classifier trained to separate gpt-image-2 images from hand-drawn anime art. They froze it before any LoRA output existed, and it reached AUC 0.99 on held-out images. Across 24 test prompts (one seed each), median score (lower = less GPT-like):

  • Text-to-image, no LoRA: 0.927 (3 of 24 under 0.5)
  • Text-to-image with the LoRA loaded: 0.909 (5 of 24)
  • Re-sampled without the LoRA: 0.952 (4 of 24)
  • Re-sampled with the LoRA: 0.452 (13 of 24)

Every one of the 24 images scored lower after the re-sample. The character stayed the same in 24 of 24 (CCIP check), and the pose held in 22 of 23 where a pose could be detected (DWPose). The author points out that the classifier never saw MOSS images, so use the scores as a guide and judge with your own eyes.

Set it up in ComfyUI

Update ComfyUI to a build that includes Text Encode MOSS-Image Edit, then put the files here:

  • models/diffusion_models/ → moss_image_sft_bf16.safetensors (10.2 GB, Comfy-Org/MOSS-Image)
  • models/text_encoders/ → qwen3.5_9b_bf16.safetensors (19.3 GB)
  • models/vae/ → flux2-vae.safetensors (336 MB)
  • models/loras/ → ungpt_v1.safetensors (149 MB, from the v1.0 release)

Grab the repo for the workflows:

git clone https://github.com/bag3labo/moss-image-ungpt

Then drag one of these into ComfyUI:

  • moss_sft_t2i_then_ungpt.json: prompt to final image in one queue, saving both before and after
  • ungpt_resample_image.json: you already have a MOSS SFT image, so load it and queue (an example input PNG is included)

API-format versions for scripting sit in workflows/api/.

The graph, node by node

The re-sample workflow is short enough to rebuild by hand:

  • UNETLoader (MOSS SFT bf16) → LoraLoaderModelOnly (ungpt_v1, strength 1.0) → SkipLayerGuidanceDiTSimple (single_layers = 22, double_layers empty)
  • CLIPLoader (Qwen3.5 9B, type moss_image) and VAELoader (Flux2 VAE)
  • LoadImage → ImageScaleToTotalPixels (lanczos, about 1 MP, sides in multiples of 16) → GetImageSize
  • Positive: TextEncodeMOSSImageEdit with the fixed instruction below and edit format ai-toolkit (kv_cache off)
  • Negative: TextEncodeMOSSImageEdit with an empty prompt and the same image attached
  • EmptyFlux2LatentImage at the input size → KSampler (50 steps, CFG 3, euler, sgm_uniform, denoise 1.0) → VAEDecode → SaveImage

The instruction the results were measured with:

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

On an RTX PRO 6000, the author measured about 20 s for text-to-image and 46 s for the re-sample, using about 28 GB of VRAM. Smaller cards and quantized MOSS files weren't tested.

Caveats

  • Results vary. 13 of 24 images dropped under 0.5, but on 4 of them the score barely moved (0.04 or less). Compare before and after on your own images.
  • Small details can drift. Hair, faces and expressions can shift a little, so check any text in the image too.
  • It adds a look, not an artist. It pushes toward a generic hand-drawn style and doesn't recreate any particular artist.
  • gpt-image-2 inputs are shakier. On 36 gpt-image-2 test images the median score fell from 0.890 to 0.013, but the pose held in only 83%, with stronger colour casts and more face/framing drift.
  • Untested so far: MOSS-Image Base, multiple seeds per prompt, other strengths or samplers.
  • It's still AI art. The author asks you to keep labeling outputs as AI-generated. The score isn't an AI detector, and the LoRA hasn't been tested against one.
  • Brand new: one developer, zero stars when I checked.

Original Source

Conclusion

MOSS-Image SFT gives you great characters, but the glossy polish can make every render look the same. un-GPT is a tidy fix: one LoRA, one fixed instruction and one extra pass, and you get flatter shading and crisper lines with your character still intact. Keep the original, run the re-sample, and pick your favourite. 🎨✨

—Aurelia ♡