jejunepixels/Hammershoi-Flux3-LoRA

Hammershoi FLUX 3 Klein LoRA

Prompt

A medium view of a woman standing beside a tall window in a quiet room, looking out. Daylight from the window falls across her shoulder and the wall behind her. A bare wall and a plain wooden floor. hmrsh style

Prompt

a row of buildings on a quiet street, hmrsh style

Prompt

an empty room with a door, hmrsh style

Prompt

a woman standing by a window, hmrsh style

Prompt

A wide view of a row of buildings along a quiet street, with no people or traffic. Flat facades with rows of tall windows, a paved road in the foreground, sky above the rooftops. hmrsh style

Prompt

A wide view of an empty room with a single closed door in the far wall and a tall window on the left. Daylight falls across the wooden floorboards. Bare walls, a simple skirting board, no furniture. hmrsh style

Model description

A LoRA fine-tuned on FLUX 3 Image Klein Base to generate images in the quiet, introspective style of Danish painter Vilhelm Hammershøi (1864–1916). Characterized by muted palettes, sparse domestic interiors, solitary figures, and profound stillness.

You don't need to describe the style in your prompt. The gallery prompts contain no style words, only the subject and the trigger phrase. See also the Flux1 version, the Flux2 version and the Ideogram 4 version.

The same prompt and seed in each row, on prompts with no style words.

Trigger words

End your prompt with hmrsh style.

Recommended settings

Parameter	Value
Base model	FLUX 3 Image Klein Base
Steps	30
Guidance	3.5
LoRA scale	1.0
Resolution	1024×1024 (square)

LoRA scale 1.5 moves further from the original image and closer to the muted colours of the paintings, but it can look grainy. All testing used these settings at 1024×1024.

Usage

The base model is gated, so accept its license on the model page and log in with your Hugging Face token first. Install Diffusers from GitHub, along with peft. This is the code the gallery images were made with:

import torch
from diffusers.modular_pipelines import Flux3AutoBlocks

blocks = Flux3AutoBlocks()
blocks.sub_blocks.pop("prompt_upsampler", None)  # send the prompt as typed
pipe = blocks.init_pipeline("black-forest-labs/flux-3-image-klein-base")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

pipe.load_lora_weights(
    "jejunepixels/Hammershoi-Flux3-LoRA",
    weight_name="Hammershoi_flux3_klein.safetensors",
    adapter_name="hammershoi",
)
pipe.set_adapters(["hammershoi"], adapter_weights=[1.0])
pipe.guider = pipe.guider.new(guidance_scale=3.5)

image = pipe(
    prompt="a woman standing by a window, hmrsh style",
    width=1024,
    height=1024,
    num_inference_steps=30,
    generator=torch.Generator("cuda").manual_seed(42),
    output="images",
)[0]
image.save("hammershoi.png")

Prompt tips

Describe the framing and the subject, then add the trigger phrase:

a medium view of a woman in a dark dress seen from behind, hmrsh style

Both short prompts and full scene descriptions work in the gallery. The full descriptions gave cleaner pictures when tested without the LoRA, and Black Forest Labs recommends them for this model.

Limitations and biases

• Testing was small. The final checkpoint was tested on six prompts with one seed, at guidance 3.5 and 1024×1024 only. Other subjects, seeds and image shapes are untested.
• Some setups are untested. This has not been tried with the Klein prompt upsampler, with the distilled FLUX 3 Image Klein checkpoint, or in ComfyUI.
• It can change your subject. It doesn't only change the surface. At scale 1.0 a modern figure can become a woman with pinned-up hair in a dark period dress, even if you didn't ask for that.
• Narrow subject range. It was trained on one painter's work: sparse Danish interiors, landscapes and portraits. People in these paintings are mostly European, often women in dark period dress, so expect results to lean that way unless your prompt says otherwise.
• Captions are not perfect. They were shortened by script from captions that were drafted by Claude and reviewed by hand. The result was spot-checked, with some fixed by hand.

Training details

Base model	black-forest-labs/flux-3-image-klein-base
Training images	288 crops, the same images as jejunepixels/Hammershoi-Ideogram4-Dataset, which lists each source
Captions	Framing and subject only, at most 15 words, no style words
Sources	SMK Open · Wikimedia Commons
Training tool	AI Toolkit
Training hardware	RTX 4090
Trigger phrase	hmrsh style
LoRA rank	32
Steps	2000
Learning rate	0.0003
content_or_style	content
Caption dropout	0.3, leaving only the trigger phrase
Sampler	Flowmatch
dtype	bf16

License

These weights are a fine-tuned derivative of FLUX 3 Image Klein Base, modified by jejunepixels, and are released under the FLUX Kommunity License v.1.0. Any rights to use them are granted directly by Black Forest Labs to you under that license. It permits non-commercial use, and commercial use of outputs by Qualifying Users, subject to its terms. A copy of the license and the required NOTICE are included in this repo.

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

The weights are provided as is, with no warranty of any kind, and with the same disclaimers and limits of liability as Sections 5 and 6 of the license.

This is a community project. It is not an official Black Forest Labs product and has not been endorsed, approved or validated by Black Forest Labs. Images made with this LoRA are AI-generated.

Training images were collected from SMK Open and Wikimedia Commons. The Wikimedia files trace back to museums, auction houses and other sources. Vilhelm Hammershøi died in 1916 and his paintings are in the public domain; check the source of a reproduction before reusing it.

Download model

Download the weights in the Files & versions tab.