RORem-mixed distilled to four steps, LCM LoRA, ONNX fp16, for greycard
The fast variant of the generative remove in greycard: a latent consistency LoRA (Luo et al., LCM-LoRA: A Universal Stable-Diffusion Acceleration Module, 2023) trained by greycard on top of greycard-rorem, the ONNX export of RORem-mixed (Li et al., RORem: Training a Robust Object Remover with Human-in-the-Loop, CVPR 2025). Three UNet calls replace twenty-four, at about a third of the time, with the base model's weights left as they are.
These files are only the adapters and two graphs that apply them.
They do nothing without unet.onnx_data from greycard-rorem, which
greycard places beside them; the VAE and the embeddings are
greycard-rorem's too.
How the files were made
- Training. The authors' LCM distillation script on RORem-mixed's UNet: DDIM solver, 50 skeleton steps, boundary scalings at timestep scaling 10, a rank 64 LoRA on their module list, the same UNet with its adapters off as the teacher at guidance 1.0 on the empty prompt. 4000 steps at batch 8 and learning rate 1e-4, then 1500 more at batch 6 and 2e-5 from step 3000 with an exponential moving average of the adapter weights (decay 0.997); the average at step 4000 is what ships. The data was 2082 of greycard's author's own photographs, 8000 random 1024 crops with LaMa-style masks.
- Export. The LoRA's 788 adapters (A and B, fp16) are kept
unfused in
unet-lora.onnx_data. The two graphs are greycard-rorem's UNet graphs at latent 128 and 64 with each adapted linear and conv rewritten as the base weight plus the scaled product of its adapters, reading the base weights from greycard-rorem'sunet.onnx_databy name and the adapters from this file. - The Resize fix and the
conv_inpad are greycard-rorem's, unchanged.
| File | Bytes | SHA-256 |
|---|---|---|
unet-lcm-1024.onnx (latent 128) |
3,419,476 | 0b5320148e08fc354f2b4e6b1f2f1927d8fb3c8bf2f522188db193d3ad90f4f2 |
unet-lcm-512.onnx (latent 64) |
3,419,472 | c12747ce96160ec6ad0c89a941fc70452ad8784d563594183a18a5cb905bfc62 |
unet-lora.onnx_data (the adapters, both graphs) |
393,502,720 | 2beaa20d700c59e191e51fd1d9a0698110a9354c6ec16b7c570ba96e7bff6d33 |
Inputs and outputs are greycard-rorem's UNet's. 400,341,668 bytes in all.
What it gives up
On 34 held-out windows the student sits about one level above the teacher on average with a spread of two, where the teacher's own spread against the truth is seven; against the truth both score the same. On large holes its fills are rougher (shards where the teacher draws shapes), which is why greycard offers it as Fast beside Best and lets each patch choose.
License
The adapters are a derivative of RORem-mixed, itself a fine-tune of
SDXL inpainting, and carry the CreativeML Open RAIL++-M License
with its use-based restrictions (Attachment A), in
LICENSE, as greycard-rorem does. The training data is
the author's own photographs; see greycard-rorem's card for the base
model's training data.
How greycard fetches it
greycard downloads these files when Fast is first chosen, after
showing this license, into the generative remove's store
(~/.cache/greycard/models/rorem-mixed-fp16/ on Linux, beside
greycard-rorem's files), checks each against the size and SHA-256
above, and runs them through ONNX Runtime on WebGPU: the LCM
scheduler over four steps from strength 0.99 (three UNet calls), in
about 3 s at 1024 on a 16 GB card.