VRuins/thomas-cole-sdxl-lora

Thomas Cole SDXL LoRA

An SDXL LoRA trained on paintings by Thomas Cole and the Hudson River School (mid-1800s American romantic landscape painting). Trained as the style backbone for the slow-interpolation pipeline, which generates slow, painterly looped video at the pace of weather, not the pace of cinema.

Above: still frame from a 60s loop rendered with this LoRA. Full clip: sample_clip.mp4 in the Files tab.

Public release of this LoRA is the v0.1 demo asset for the slow-interpolation repo. Train your own LoRA in any style by following the dataset-curation protocol and train-on-Modal protocol in that repo.

Quick use

from diffusers import DiffusionPipeline, EulerDiscreteScheduler
import torch

pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda")

# SDXL Lightning 4-step (the production backbone for slow-interpolation).
pipe.load_lora_weights("ByteDance/SDXL-Lightning", weight_name="sdxl_lightning_4step_lora.safetensors")
pipe.fuse_lora()
pipe.unload_lora_weights()
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")

# Thomas Cole LoRA fused at scale 0.75 (the calibrated default).
pipe.load_lora_weights("VRuins/thomas-cole-sdxl-lora", weight_name="thomas-cole-sdxl-lora.safetensors")
pipe.fuse_lora(lora_scale=0.75)
pipe.unload_lora_weights()

image = pipe(
    prompt="tcole, romantic landscape oil painting, Hudson River School style, a wide valley at dawn",
    num_inference_steps=4,
    guidance_scale=1.5,
).images[0]

Recommended settings

• Trigger word: tcole. Lead every prompt with it.
• Prompt prefix template: tcole, romantic landscape oil painting, Hudson River School style, <subject>.
• lora_scale: 0.75 (calibrated default). The slow-interpolation reference config examples/configs/tcole_valley.yaml uses this value.
• 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 (SDXL's 16:9 training bucket; gives the best Hudson-River-School horizontal aspect).
• Negative prompt: photograph, 3D render, anime, cartoon, digital art, smooth blending, sharp edges, frame, border, signature, watermark.

Training data

100 to 110 public-domain paintings sourced from Wikimedia Commons + WikiArt, primarily Thomas Cole plus closely-adjacent Hudson River School painters (Asher B. Durand, Frederic Edwin Church, Albert Bierstadt, Thomas Moran). All works pre-1929 and in the public domain in the United States.

The dataset was curated via the dataset-mosaic 5-phase protocol: Wikimedia source, Gemini-assisted multi-pass crop + non-target rejection, gallery review by the author, rule-based captioning with tcole trigger.

Training settings

• Engine: Kohya / sd-scripts (the canonical training engine for the slow-interpolation project; CivitAI / AI Toolkit produces different but viable LoRAs).
• Network rank: 16, alpha: 8.
• UNet LR: 3e-5, AdamW8bit, 6 repeats, 10 epochs.
• Resolution: 1024 (square training bucket).
• Cost on Modal L40S: ~$1.50, ~45 min wall.
• See the full hyperparameter rationale in docs/findings/lora-training.md.

The published checkpoint is epoch 10, the canonical strong-style pick. Epoch 1 (the atmospheric / lighter alternative) is available on request; future versions may publish both.

What this LoRA is for

The slow-interpolation pipeline uses this LoRA as the style backbone for ~60s looped videos. The LoRA generates SDXL Lightning keyframes through an img2img chain with a slowly-evolving noise tensor; RIFE v4.25 interpolates 64x at a linear timestep so motion is glacial and the loop closes without a cut.

Second sample: a cloudship over a wooded coastline at sunset. Same LoRA, different subject + light.

Full reference clips of the slow-interpolation repo are at examples/outputs/cole_valley_horizontal.mp4 and examples/outputs/tcole_cloudship_horizontal.mp4 on GitHub.

Provenance and attribution

LoRA trained by Luca Martinelli (Vandalo Ruins) in 2026 as part of the slow-interpolation project. Released under MIT.

The training images are public-domain works by deceased artists (Cole 1801-1848, Durand 1796-1886, Church 1826-1900, Bierstadt 1830-1902, Moran 1837-1926). Wikimedia Commons and WikiArt list each work's attribution; specific image-level provenance was not preserved in the dataset metadata because the works are all in the public domain.

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:

Luca Martinelli (Vandalo Ruins), Thomas Cole SDXL LoRA, slow-interpolation v0.1, 2026. https://huggingface.co/VRuins/thomas-cole-sdxl-lora

Related work in the slow-interpolation project

• v0.1 demo LoRA (this one): Thomas Cole / Hudson River School.
• v0.2 (alongside the objkt labs Renoir art release):
  • Casa del Suono fresco LoRA (vertical/portrait fresco aesthetic).
  • Renoir Flowers LoRA.
  • Soutine Figures LoRA.

Repository: https://github.com/VandaloRuins/slow-interpolation.