Big thanks to RunDiffusion (https://www.rundiffusion.com) – without them none of these experiments would be possible.
Also on Hugging Face: https://huggingface.co/Kandoo/MOSS-Image-Impressionism-LoRA
This is an always-on Impressionism LoRA for MOSS-Image. There’s no trigger word. Just load it and whatever you prompt turns into an impressionist oil painting, with soft broken brushwork, vibrating light and painted reflections. It works for classic stuff like harbors, gardens and Parisian cafés, but also for modern things like neon streets, astronauts or rocket launches. None of the showcase prompts have words like “painting” or “impressionist” in them. The style comes entirely from the LoRA.
I trained it on MOSS-Image Base and it works on both Base and SFT. The showcase images are made with SFT. Strength 1.0 is a good default and already looks great for landscapes, light and anything historic. Very modern or photographic prompts pull harder toward a photo look, so for those just go up to 1.4–1.6.
MOSS likes long, descriptive prompts, so describe the scene, the light and what people are doing. If there’s any text in the image like signs, posters or shop names, write it out in quotes in your prompt, otherwise you get gibberish.
Settings I used. SFT: euler / simple, 50 steps, CFG 3, SkipLayerGuidance on layer 22 (0–1), empty negative. Base: res_multistep / simple, 30 steps, CFG 4, SkipLayerGuidance on layer 22 (0–0.5), negative “worst quality, bad anatomy, watermark, jpeg artifact”. ComfyUI workflows and all prompts with seeds are in the Hugging Face repo.
Trained with Fizgig on 51 public domain (CC0) impressionist paintings from the Art Institute of Chicago. Short content-only captions, rank 16, AdamW 5e-5 constant, 1024px, 25 epochs. Epoch 30 started drifting back toward the base look, so this is epoch 25.
Have fun with it!