Sellestia — background removal that keeps the plate

Sellestia is a BiRefNet-based background-removal model, fine-tuned from Lucida for top-down food photography.

General-purpose models treat the plate as background: they cut the food out and drop the plate, or leave a half-transparent ghost of it. Sellestia keeps the dish and its plate as one subject, whatever the plate's colour, and behaves like Lucida on everything else.

All images above are held out: none of them was used for training.

Results

Held-out test set of top-down dish photos (plate kept = correct answer), the test split of the Sellestia dataset:

23× lower error than Lucida on food photos, and every test image above 0.95 IoU (worst case 0.965).

Outside food, Sellestia stays close to its base model: on held-out objects, animals and illustrations its masks agree with Lucida's at 0.987 IoU.

Usage

import torch
from PIL import Image
from torchvision import transforms
from transformers import AutoModelForImageSegmentation

model = AutoModelForImageSegmentation.from_pretrained(
    "franckdsf/sellestia", trust_remote_code=True, dtype=torch.float32)
model.eval()

t = transforms.Compose([
    transforms.Resize((1024, 1024)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
img = Image.open("dish.jpg").convert("RGB")
with torch.no_grad():
    preds = model(t(img).unsqueeze(0))[-1].sigmoid()
alpha = transforms.functional.resize(preds[0], img.size[::-1]).squeeze(0)
rgba = img.copy()
rgba.putalpha(Image.fromarray((alpha.numpy() * 255).astype("uint8")))
rgba.save("dish.png")

Same architecture, preprocessing and 1024×1024 input as Lucida: it is a drop-in replacement anywhere Lucida or BiRefNet is loaded through transformers.

Training

• Starts from Lucida's weights. The backbone is frozen and only the decoder is fine-tuned, which keeps the base model's general knowledge intact.
• Data: top-down dish photos with plate-inclusive masks (Sellestia dataset, CC-BY 4.0), mixed with a rehearsal set of varied non-food images (objects, vehicles, animals, glass, people, anime) labelled by Lucida itself, so the model learns "keep the plate" without forgetting the rest.
• Loss: BCE + IoU. Held-out validation and test splits per data source.

Base model & attribution

• Fine-tuned from egeorcun/lucida (MIT), itself a fine-tune of ZhengPeng7/BiRefNet_HR (MIT). The original copyright notices are preserved in LICENSE.
• The rehearsal set includes images from DIS5K, AIM-500, P3M-10K and anime-segmentation (CC0). Some of these are distributed for research purposes; evaluate suitability for your use case.

License

MIT (weights and code).