Takotak/darkslide-moebius-coreml

Moebius for Core ML, as used by Darkslide

The object-removal model of Darkslide, a RAW editor for photographers: the photographer marks what to remove, this model fills the area from the picture around it, on their Mac. No text input, no other use.

moebius-ft_places2-512-coreml.zip holds three Core ML ML Programs (fp16 weights, 512 × 512 pixels, latents 64 × 64):

Model	Inputs	Output
vae_encoder.mlpackage	image 1×3×512×512 in −1…1	latent 1×4×64×64, mean × 0.13025
unet.mlpackage	latent, mask 1×1×64×64, masked_latent, timestep int32, guidance	noise, classifier-free guided
vae_decoder.mlpackage	latent 1×4×64×64	image 1×3×512×512 in 0…1

The U-Net has Moebius's twenty conditioning ids baked in (unconditional 10–19, conditional 0–9). Sampling: DDIM, scaled-linear betas 0.00085–0.012, 10 steps with the first skipped, guidance 2.0, noise offset 0.0357.

Sources and licences

• Moebius by hustvl (ECCV 2026), the ft_places2 weights: code and weights under the Apache License 2.0, https://github.com/hustvl/Moebius. The CelebA-HQ and FFHQ fine-tunes are not used.
• SDXL VAE by Stability AI, MIT licence, https://huggingface.co/stabilityai/sdxl-vae (as shipped with PixelHacker).
• Converted to Core ML by Darkslide with coremltools 9.

The ft_places2 weights were fine-tuned on Places2 (Places365), whose image-download terms are limited to non-commercial research and education; those terms say nothing of models trained from the images.

SHA-256 of the archive: a4a8e1bbad685883a96682f08e7e9ef0754a3318b5d0c1deb932370c1a2e53d6.