LHMPP-700M-PixelShuffle (mirror)

Unmodified mirror of Damo_XR_Lab/LHMPP-700M-PixelShuffle from ModelScope, at revision 5f1c4274068e11b93721219618d36594b6087cb6.

All credit to the original authors (Damo XR Lab / LHM-plusplus). Redistributed under Apache-2.0, the licence declared on the source repository.

Why this mirror exists

ModelScope throughput from cloud datacenters in Europe/US measures around 0.7 MiB/s, so pulling the 4.87 GB model.safetensors takes upwards of three hours — longer than a Hugging Face Space stage timeout allows, and the partial download is discarded on failure. The same file over the HF CDN completes in about a minute.

Contents

File	Size	Note
model.safetensors	4984.3 MiB	byte-identical to source (5226465708 bytes, 1523 tensors)
config.json	3103 B	byte-identical to source
configuration.json	44 B	byte-identical to source
README.md	—	rewritten for HF model-card metadata; source card was ModelScope-specific

Only the model card differs from upstream. The weights and configs are unchanged.

This is the PixelShuffle variant

Not interchangeable with 3DAIGC/LHMPP-700M, despite the similar name. This checkpoint carries 41 shape_head.* tensors and predict_shape_dim: 10 in its config, giving it an image-predicted shape head; 3DAIGC/LHMPP-700M has no shape_head and instead defines a neural_renderer block that this one lacks. Loading one where the other is expected fails on a state-dict mismatch.

Verifying against the source

import json, struct

with open("model.safetensors", "rb") as f:
    header = json.loads(f.read(struct.unpack("<Q", f.read(8))[0]))

print(len([k for k in header if k != "__metadata__"]))            # 1523
print(len([k for k in header if k.startswith("shape_head.")]))    # 41