An SDXL LoRA that captures the visual language of the historical frescoes at Casa del Suono, the contemporary music museum housed in a historical palazzo in Parma, Italy. Trained as the style backbone for the slow-interpolation pipeline's Choire v2 installation (April 2026, Casa del Suono, Parma) and now released as a v0.1 demo LoRA for the open-source slow-interpolation repo.
Reference render: a 60s slow loop in the Casa del Suono fresco aesthetic, lakeside subject at dawn
Above: still frame from a 60s loop rendered with this LoRA at scale 0.35. Full clip: sample_clip.mp4 in the Files tab.
The LoRA gives SDXL the visual vocabulary of Italian fresco painting: oil-on-plaster handling, warm chiaroscuro, terracotta and ochre palette, the kind of compositions a ceiling-fresco painter would arrange.
Quick use
python
1from diffusers import DiffusionPipeline, EulerDiscreteScheduler
2import torch
34pipe = DiffusionPipeline.from_pretrained(5"stabilityai/stable-diffusion-xl-base-1.0",6 torch_dtype=torch.float16,7 variant="fp16",8).to("cuda")910# SDXL Lightning 4-step (the production backbone for slow-interpolation).11pipe.load_lora_weights("ByteDance/SDXL-Lightning", weight_name="sdxl_lightning_4step_lora.safetensors")12pipe.fuse_lora()13pipe.unload_lora_weights()14pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")1516# Casa del Suono LoRA fused at the calibrated default scale.17pipe.load_lora_weights("VRuins/casa-del-suono-sdxl-lora", weight_name="casa-del-suono-sdxl-lora.safetensors")18pipe.fuse_lora(lora_scale=0.35)19pipe.unload_lora_weights()2021image = pipe(22 prompt="Italian fresco on aged plaster, warm chiaroscuro, visible brushwork, terracotta and ochre palette, a vast lake at dawn",23 negative_prompt="angel, wings, halo, saint, religious, crucifix, biblical, cherub, photograph, 3D render, anime",24 num_inference_steps=4,25 guidance_scale=1.5,26).images[0]
Recommended settings
No trigger word. The LoRA is activated by a descriptive prefix that names the style. Lead every prompt with: "Italian fresco on aged plaster, warm chiaroscuro, visible brushwork, terracotta and ochre palette, <subject>". Without the prefix the style does not engage.
lora_scale:0.35 is the calibrated default. The LoRA is strong; even small scale changes have large visual effects. Range: 0.25 to 0.5 typical; 0.5+ risks the fresco character overwhelming the subject.
Base model:stabilityai/stable-diffusion-xl-base-1.0.
Lightning backbone: stack with ByteDance/SDXL-Lightning 4-step LoRA for fast keyframe rendering; without it, use 30 steps at guidance 5.5 on plain SDXL base.
Resolution: 1344x768 (horizontal) or 768x1344 (vertical / portrait, the original Choire v2 orientation). Native SDXL training-bucket sizes.
Negative prompt (recommended): "angel, wings, halo, saint, religious, crucifix, biblical, cherub, photograph, 3D render, anime, cartoon, digital art, signature, watermark". The training data includes religious-iconographic content; the negatives suppress incidental religious bias unless the prompt explicitly asks for it.
Training data
Synthetic dataset. Not photographs. The author studied the historical frescoes at Casa del Suono through online reference images (the venue and its decoration are publicly documented), then generated a synthetic dataset that emulates the studied style. The LoRA was trained on this synthetic set, not on the reference images themselves.
The methodology in one line: study a real style from public reference material, generate a synthetic dataset that matches what you found, train the LoRA on the synthetic set. This is a reusable pattern for studying a target style without needing direct access to the originals; the slow-interpolation repo's dataset-mosaic protocol documents the workflow.
The dataset captures the venue's distinctive palette and brushwork conventions (terracotta, ochre, soft chiaroscuro, oil-on-plaster surface), not specific compositions or named scenes. The LoRA generalises the surface and palette across arbitrary subjects.
The underlying real frescoes at Casa del Suono are 16th to 18th century work in the public domain by age; this LoRA studies their style, it does not reproduce or claim specific works.
Training settings
Engine: Kohya / sd-scripts.
Trained in 2025 for the Choire v2 installation; checkpoint at epoch 4 is the canonical pick (the file in this repo).
The slow-interpolation pipeline uses this LoRA as the style backbone for ~60s looped videos that drift through subject and light variations. The technique was first shown publicly at Casa del Suono, Parma (Choire v2 installation, April 2026). The published LoRA lets anyone reproduce the technique's surface in their own work.
Reference clips from the Choire v2 production work, all rendered with this LoRA, are at examples/outputs/ on GitHub:
E14_lake_horizontal.mp4: lakeside subject (the sample clip in this repo).
E24_garden_horizontal.mp4: garden subject.
notturno_city_horizontal.mp4: nocturnal city.
siege_harbour_horizontal.mp4: harbour with ruins.
harbour_market_horizontal.mp4: harbour market.
Provenance and attribution
LoRA trained by Luca Martinelli (Vandalo Ruins) in 2025 as part of the slow-interpolation project. Released under MIT.
Training data: synthetic dataset generated by the author, derived from study of the historical frescoes at Casa del Suono through publicly available online reference material. No photographs of the venue are shipped with this LoRA; the reference material was study input, not training input. The training set itself is a derivative work of the author's study and is not distributed.
The historical frescoes themselves at Casa del Suono (Parma, Italy) are in the public domain by age. Casa del Suono is the contemporary cultural space (music museum) housed in the palazzo whose frescoes were the study subject; the LoRA's name references the location, not a claim of association or endorsement by the venue.
License
MIT for the LoRA weights. The underlying SDXL base model is licensed separately by Stability AI under the CreativeML OpenRAIL-M license; users must comply with both.
Cite
If you use this LoRA in a published work, a citation is appreciated but not required:
