victor/Krea-2-LoRA-magritte

Krea 2 LoRA — Magritte (mgrtt style)

A style LoRA that teaches Krea 2 the surrealist painting style of René Magritte: clean illusionistic rendering, muted palettes, flat skies with cumulus clouds, and impossible juxtapositions of everyday objects. Built in the spirit of the official Krea 2 LoRAs collection.

Trigger phrase: mgrtt style — put it at the start of every prompt.

Want to make your own? This LoRA was produced end to end (dataset research, training on fal, and even this model card) by one small open-source recipe. It is genuinely simple and cheap: roughly 30 images, ~1000 steps, about $3 on fal. Point it at any artist or style with the skill here: gary149/train-krea2-lora.

Gallery

All previews below were generated with this LoRA on Krea 2 Turbo (8 steps, guidance 0.0); each image shows its prompt.

Prompt

mgrtt style, a green apple floating in front of a man's face, blue sky with white clouds

Prompt

mgrtt style, men in bowler hats and dark coats hovering in a pale sky above city rooftops

Prompt

mgrtt style, a man in a suit whose head is a glowing light bulb, standing in a field under a cloudy sky

Prompt

mgrtt style, a giant goldfish floating above an empty subway platform at dusk

Prompt

mgrtt style, a vintage rotary telephone carved entirely from grey stone resting on a beach, blue sky with clouds

Prompt

mgrtt style, an open umbrella with a glass of water balanced on top, against a flat grey sky

Prompt

mgrtt style, a wooden door standing alone in a misty forest, an identical forest visible through the open doorway

Prompt

mgrtt style, a room in warm daylight with a window revealing a deep starry night sky

Model details

• Developed by: @victor
• Model type: LoRA adapter for a text-to-image diffusion transformer
• Base model (inference): krea/Krea-2-Turbo
• Trained on: krea/Krea-2-Raw — Krea's recommended workflow is to train on RAW and run on Turbo.
• Trigger phrase: mgrtt style
• Rank / alpha: 32 / 32
• File: magritte.safetensors (~469 MB, fp32 LoRA weights)
• License: Krea 2 Community License (inherited from the base model)

How to get started

Option A — fal (tested; loads the native weights directly)

These weights are in Krea's native checkpoint key format, so the fal-ai/krea-2/turbo/lora endpoint loads them as-is. This is exactly how the gallery above was produced.

import fal_client

result = fal_client.subscribe("fal-ai/krea-2/turbo/lora", arguments={
    "prompt": "mgrtt style, a steam train emerging from a fireplace in a quiet room",
    "loras": [{
        "path": "https://huggingface.co/victor/Krea-2-LoRA-magritte/resolve/main/magritte.safetensors",
        "scale": 1.0,
    }],
    "image_size": "square_hd",
})
print(result["images"][0]["url"])

Lower scale to ~0.8 for subtler stylization and more compositional freedom.

Option B — diffusers

import torch
from diffusers import Krea2Pipeline  # pip install git+https://github.com/huggingface/diffusers.git

pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Turbo", torch_dtype=torch.bfloat16).to("cuda")
pipe.transformer.load_lora_adapter("victor/Krea-2-LoRA-magritte", weight_name="magritte.safetensors")
pipe.transformer.set_adapters("default", weights=1.0)
image = pipe("mgrtt style, a green apple floating before a man's face, blue cloudy sky",
             num_inference_steps=8, guidance_scale=0.0).images[0]

Note: these weights use Krea's native module key names (base_model.model.blocks…) rather than the diffusers naming (transformer.blocks…) used by Krea's first-party LoRAs. The fal path above is verified; loading in diffusers may require a key remap depending on your diffusers version. If load_lora_adapter errors on unexpected keys, use Option A or convert the keys first.

Uses

• Direct use: generate Magritte-style surrealist images by prompting mgrtt style, <scene>. Works best when you describe a concrete scene/object juxtaposition.
• Downstream use: combine with other Krea 2 LoRAs (the fal endpoint accepts up to 3), or use as a starting point for further fine-tuning.
• Out-of-scope use: this is a style LoRA, not a character/likeness model. It does not reproduce a specific person, and it should not be used to misrepresent the work as authentic Magritte paintings or to forge attribution.

Training details

• Training data: 32 René Magritte paintings from his signature surrealist period (1926–1967), sourced from WikiArt at full resolution, kept at native ~1.5–2K (aspect-preserved, sRGB, /16-aligned, metadata stripped). Candidates were curated for genuineness, period, and scan quality, then selected to maximize motif diversity (bowler-hat men, skies/clouds, windows & easels, petrified objects, fragmented figures, day/night plays, fire, curtains, birds, the word-image paradox).
• Captions: each image is captioned mgrtt style, <concrete content description> with no medium/style words in the body, so the painterly look binds to the trigger phrase rather than to generic tokens.
• Procedure: trained with the fal krea-2-trainer on krea/Krea-2-Raw. Rank 32, alpha 32, 1000 steps, learning rate 5e-4 (AdamW), fp8 trainer. Targets attention (q/k/v/o, gate), MLP, and text-fusion projection layers.
• Sizes/times: output LoRA ~469 MB; 1000-step run on fal.

Technical specifications

Krea 2 is a ~12B dense diffusion transformer (MMDiT) with a Qwen-Image VAE and a Qwen3-VL text encoder. RAW is the undistilled base used for training; Turbo is the 8-step distilled checkpoint used for inference. LoRAs trained on RAW are designed to express strongly on Turbo.

Bias, risks, and limitations

• Copyright: René Magritte died in 1967; his works remain under copyright (until ~2037 in life+70 jurisdictions). This adapter was trained on reproductions for personal/research style transfer. Generated images imitate a style, which is generally not protected, but do not present outputs as genuine Magritte works, and review your local rules before commercial use.
• Memorization: with only 32 images and 1000 steps, prompts that closely describe a specific training painting can produce near-reproductions of that composition. For novel results, prompt with scenes that are not in Magritte's catalogue and/or lower the LoRA scale.
• Inherited behavior: all biases, safety limitations, and the acceptable-use policy of the krea/Krea-2 base models apply here.

Recommendations

Use a distinctive scene description, keep scale between 0.7–1.0, and rely on the base model's safety filtering when deploying.

Acknowledgements

Built on Krea 2 by Krea AI. Trained via fal. Source imagery from WikiArt.