Kandoo/MOSS-Image-Impressionism-LoRA

MOSS-Image Impressionism LoRA

A big thank you to RunDiffusion - without them none of these experiments would be possible.

An always-on Impressionism style LoRA for MOSS-Image by KandooAI.

There is no trigger word. Load the LoRA and any prompt turns into an impressionist oil painting: soft broken brushwork, vibrating light and color, painted reflections. Modern subjects like neon-lit city streets, astronauts or rocket launches work just as well as classic motifs.

None of the showcase prompts contain style words like "painting" or "impressionist". The style comes entirely from the LoRA.

Trained on MOSS-Image Base. It also works on MOSS-Image SFT, which is what the showcase images were made with.

Usage

File: moss-image-impressionism-lora.safetensors

Strength: 1.0 is the default. Motifs that are already "painterly" (landscapes, light, historic scenes) look great at 1.0. Modern or very photographic prompts pull harder toward a photo look. For those, go up to 1.4 - 1.6.

Prompting: Write normal, descriptive prompts about the scene, the light and the people in it. If there is any text in the image (signs, posters, shop names), write it out in quotes in the prompt, e.g. a red awning with the words "CAFE DES ARTISTES" in gold letters. Otherwise MOSS renders gibberish.

ComfyUI

Load the model with UNETLoader, then LoraLoaderModelOnly with this LoRA, then SkipLayerGuidanceDiTSimple, then KSampler. Text encoder: Qwen 3.5 9B (CLIPLoader, type moss_image). VAE: FLUX.2 VAE.

Example workflows (ComfyUI API format, can be dragged into ComfyUI) are in workflows/.

	MOSS-Image SFT	MOSS-Image Base
Sampler / scheduler	euler / simple	res_multistep / simple
Steps	50	30
CFG	3.0	4.0
Skip Layer Guidance	layer 22, 0.0 - 1.0	layer 22, 0.0 - 0.5
Negative prompt	empty	worst quality, bad anatomy, watermark, jpeg artifact

The showcase images were rendered at 1728x1344 / 1344x1728 on SFT with LoRA strength 1.0. The exact prompts and seeds are in showcase_prompts.txt.

Base vs. LoRA

Same prompt, same seed, MOSS-Image Base without the LoRA (left) and with the LoRA at strength 1.0 (right). These were plain test prompts without any style words. All 20 comparisons are in images/comparison/, prompts and seeds in comparison_prompts.txt.

Even a flat vector illustration from the base model turns into a painted park scene.

Training

Trainer	Fizgig (headless, --family moss_image)
Base model	MOSS-Image Base, quantized to NF4 during training
Dataset	51 public-domain (CC0) impressionist paintings from the Art Institute of Chicago
Captions	short, content only (e.g. "two haystacks in a field at sunset"), no style words, no trigger
Resolution	1024, aspect ratio buckets
Network	LoRA rank 16, alpha 16, all linear layers in all 42 blocks
Optimizer	AdamW, learning rate 5e-5 constant, max grad norm 1.0, batch size 1
Timesteps	30 % of samples drawn from the low-noise (clean) end
Steps	25 epochs, 1275 steps. This is the epoch-25 checkpoint; epoch 30 started to drift back toward the base look.
Hardware	NVIDIA RTX A4500

The training images are works by Claude Monet, Pierre-Auguste Renoir, Camille Pissarro, Childe Hassam, Alfred Sisley, Berthe Morisot, John Henry Twachtman, Edgar Degas, Mary Cassatt, William Merritt Chase, Joaquín Sorolla, Edward Henry Potthast, Robert William Vonnoh and Henry Ward Ranger, all released under CC0 by the Art Institute of Chicago.

License

The LoRA is released under Apache-2.0, the same license as MOSS-Image.

Created by KandooAI.