RORem pipeline with the LCM-distillation LoRA merged into the UNet (788 adapters, scale 1)
f8a5222 verified | license: apache-2.0 | |
| library_name: diffusers | |
| pipeline_tag: image-to-image | |
| base_model: LetsThink/RORem | |
| tags: | |
| - inpainting | |
| - object-removal | |
| - lcm | |
| - sd-xl | |
| # RORem (merged) | |
| SDXL - RORem (CVPR 2025) repackaged as a plain diffusers | |
| `StableDiffusionXLInpaintPipeline`: the [LetsThink/RORem](https://huggingface.co/LetsThink/RORem) | |
| weights with their LCM-distillation LoRA already merged into the UNet, so | |
| conversion pipelines install straight from this repository with no adapter | |
| step. Nothing was retrained — the merge is a plain B@A bake at scale 1 | |
| (788 adapters), numerically equivalent to diffusers' own `fuse_lora`. | |
| ## Inference contract | |
| Verified by instrumenting the reference `inference_RORem_4S.py` call | |
| (diffusers 0.41): | |
| ```python | |
| pipe(prompt="", image=image, mask_image=mask, guidance_scale=1.0, | |
| num_inference_steps=4, strength=0.99) # LCMScheduler | |
| ``` | |
| The scheduler places [999, 759, 499, 259] and the strength slice drops the | |
| head — three denoiser passes at **[759, 499, 259]**, with the photo noised | |
| to 759. Guidance is off and the prompt is empty: RORem is prompt-free. | |
| ## Provenance | |
| Derived from [LetsThink/RORem](https://huggingface.co/LetsThink/RORem) | |
| (Apache-2.0), the LCM LoRA from the RORem-LCM release, on the | |
| [CreativeML Open RAIL++-M](https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1) | |
| SDXL inpainting base. Project page: github.com/leeruibin/RORem. | |