heypatkay/solume-objectclear-coreml

ObjectClear, Core ML, for Solume

A Core ML conversion of ObjectClear (Jixin Zhao, Shangchen Zhou, Zhouxia Wang, Peiqing Yang, Chen Change Loy, NTU S-Lab: "ObjectClear: Complete Object Removal via Object-Effect Attention", 2025), SDXL 1.0 Inpainting 0.1's UNet fine-tuned to remove an object together with its shadow or reflection. Made for Generative Remove in Solume, a macOS photo app, and converted to Core ML by Solume. Solume downloads the files in objectclear/ from a pinned commit of this repository and installs them only if every file matches the size and SHA-256 pinned in the app.

This repository holds only the converted model, its licences and this card. It holds no photos or user data.

Non-commercial (please read)

ObjectClear is licensed under the S-Lab License 1.0 (LICENSE, unchanged from the authors' code repository): non-commercial use only. Redistribution and use for non-commercial purposes are permitted under its conditions; any commercial use or redistribution needs the authors' permission (contacts at the end of LICENSE).

Its base, SDXL 1.0 Inpainting 0.1 (Stability AI's Stable Diffusion XL 1.0, inpainting fine-tune by Hugging Face's diffusers team), is licensed under the CreativeML Open RAIL++-M License (LICENSE-SDXL.md). Its use-based restrictions (paragraph 5 and Attachment A) apply to every use of these files and of anything made from them. By downloading or using the model you agree to both licences, including not to use the model or its derivatives for any use listed in Attachment A. If you distribute the model or a derivative, you must pass on both licences and these restrictions.

Provenance

• Upstream weights: jixin0101/ObjectClear at revision c73af80888dbd519819d0a521b5c8d0f3cda6859, the fp16 safetensors: unet/diffusion_pytorch_model.fp16.safetensors (SHA-256 fb0c35b0be8bba4ecd61dfca4e716a0a9878639c5fc271a05a780b4dc2366c31), vae/diffusion_pytorch_model.fp16.safetensors (4ad62825e5c8b31eefb77355ba6693785619bf7d669d9b1a6fd9f19dec6d65b3), image_prompt_encoder/model.fp16.safetensors (0ebf6aa2b0f8db475ca7bd1148fcc2dc131f989cb452a45103ff23e3ece825c8), postfuse_module/model.fp16.safetensors (3e3f1dca78cfe25f36ff8b9a8b16ab52cfd0982c81afd04864cace1e4c2afaaf), text_encoder/model.fp16.safetensors (fc83cf401d930147807e7c44021c164bcc5508c9d4cc0ff35f4e354685ca9cd0), text_encoder_2/model.fp16.safetensors (a8622bd41f8d359df484fdf9de091edb9337a2fd747f0a5c9c320a62e24d1fd3), with their configs, tokenizers and scheduler.
• Method: re-implemented from the paper and the authors' code, zjx0101/ObjectClear at commit c052d91ecd8772744a5ab97527441c813ab83009 (read, not run).
• Licence texts: LICENSE is that commit's LICENSE, unchanged (1,924 bytes, SHA-256 73c84d2c2ebcdde03cb7fc853aaec0d4b3b280ee8b47f98a3d6c1931430d5aae). LICENSE-SDXL.md is LICENSE.md from stabilityai/stable-diffusion-xl-base-1.0 at revision 462165984030d82259a11f4367a4eed129e94a7b, unchanged (14,109 bytes, SHA-256 19b6998b569b53ac1fc2158a8a3202c8699a9a4605b47075715d9c96be7fb6d0).
• Conversion: Solume's scripts/remove-models/convert_objectclear.py (Solume commit 8eb5ee29, unchanged through dfa057c8), with diffusers and coremltools: the UNet as an fp16 ML Program with s512 and s1024 functions sharing one weight file and a second output (the object-effect attention map); the VAE encoder (image → the latent distribution's mean) and decoder as fp32 ML Programs (SDXL's VAE overflows in fp16); the object encoder (CLIP ViT-L/14 vision tower, projection and ObjectClear's post-fuse MLP, fp16); and prompts.bin, the two SDXL text encoders' output for the fixed prompt "remove the instance of object" (prompts.json), so no tokenizer or text encoder is shipped. The output is deterministic for the pinned toolchain (fixed Manifest identifiers and conversion date); converted on 2026-11-05.
• Toolchain: Python 3.12, torch 2.15.0, coremltools 9.0, diffusers 0.42.1, transformers 5.22.0, safetensors 0.8.0, numpy 2.5.4, on macOS 27 (Apple silicon).

Notice of changes: every file in objectclear/ was changed from the upstream files: the networks were converted from PyTorch safetensors to Core ML (UNet and object encoder fp16, VAE fp32), and the text encoders were replaced by precomputed embeddings of one fixed prompt. The weights were not fine-tuned. This conversion is not made, affiliated with or endorsed by S-Lab, NTU, Stability AI or Hugging Face.

Files

objectclear/README.md is the card the conversion writes (Solume checks its licence line); it also carries the S-Lab License's copyright notice, conditions and disclaimer.

File	Bytes	SHA-256
objectclear/README.md	2,661	1f40272215dfc5959be4c9fcef3dd820d44b16dc260114dc45b07f8c12249e95
objectclear/prompts.json	143	509857880b42163a1a0942f8673ded9beb840c6a42dea1d6b41efd9667183a84
objectclear/prompts.bin	635,904	63948bbe5384b0e3da687eb785e38d848d94f497369409a20c67df3856608d8e
objectclear/Unet.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
objectclear/Unet.mlpackage/Data/com.apple.CoreML/model.mlmodel	3,562,330	3de15178924f93539275587ef66d5cc2ba846ef91a8c0ff8782d2dac99558890
objectclear/Unet.mlpackage/Data/com.apple.CoreML/weights/weight.bin	5,135,095,872	45cf05d73c89a099465477f413d18383f34bc6bbb2417e19cc9ac28491748f73
objectclear/VAEEncoder.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
objectclear/VAEEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel	230,865	709097f786ad1417a6730f9a3f2ba775562c31070ec9ca5e99cce725c11654f5
objectclear/VAEEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin	136,668,992	790f273ef4dd2cd88c8b130f58cbd84a5ca88cebde0f2f8fc81219c9cf8753c7
objectclear/VAEDecoder.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
objectclear/VAEDecoder.mlpackage/Data/com.apple.CoreML/model.mlmodel	300,811	9589a770ee2ca0a145a57f78e86b12a8f817d842966abbb4a928b1dbde89f620
objectclear/VAEDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin	197,977,216	58010a20d9da788884ae924441b9aa931e1baa61303a033043d284f7ad8fa089
objectclear/ObjectEncoder.mlpackage/Manifest.json	617	a2576a37183e5a04265456092e18271e165564bd2e6578f6091e404236f3069b
objectclear/ObjectEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel	284,363	18fb2689afcfa7bd7cf3c490e766af808e20036fd06975c1fc564a16998f3236
objectclear/ObjectEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin	636,307,456	4cedfa866312444110cf6cae09bfde253fca8192f713375878466e8d5582f5b4

Licence

ObjectClear: S-Lab License 1.0, Copyright 2025 S-Lab, non-commercial use only; full text in LICENSE. Its base: CreativeML Open RAIL++-M License dated July 26, 2023 (Copyright 2023 Stability AI); full text with Attachment A in LICENSE-SDXL.md. Outputs: you are accountable for the images you generate and their use, which must not contravene either licence.