(WARNING) These weights are identical to the base Z-Image model — not a fine-tune (SHA256 proof)
Heads up to anyone considering this model before you download 12GB:
The transformer here is byte-for-byte identical to the official Z-Image base
model (Tongyi-MAI/Z-Image). It has not been fine-tuned, it was a false information.
You can confirm this yourself in seconds without downloading — the HF tree API
exposes the SHA256 of every LFS file:
Official Tongyi-MAI/Z-Image (transformer/):
shard 1: ecd5df7768856215812af84d5cc785ba9d84c6d8c84e6939c7ccf30a2e0d2425
shard 2: e2d5fb75ca504b2d669a33af9380b68e0c1632ecddbfca85f82631958a4b81b1
This repo (horus_transformer/):
shard 1: ecd5df7768856215812af84d5cc785ba9d84c6d8c84e6939c7ccf30a2e0d2425
shard 2: e2d5fb75ca504b2d669a33af9380b68e0c1632ecddbfca85f82631958a4b81b1
Matching SHA256 means the files are identical down to the last byte. If any
training had occurred — even one step — every tensor would change and the
hashes would differ completely. I also confirmed all 521 tensors are identical
after accounting for naming differences (this repo is diffusers format, the
ComfyUI z_image_bf16 release uses fused QKV, but the underlying weights match).
On the same seed it produces a pixel-identical image to base Z-Image, which is
exactly what you'd expect from the same weights.
Verify links:
https://huggingface.co/api/models/tokenaii/Horus-Lens-1.0/tree/main/horus_transformer
https://huggingface.co/api/models/Tongyi-MAI/Z-Image/tree/main/transformer
Posting so nobody else wastes bandwidth/time. If this is actually a fine-tune,
please share the trained weights, because these aren't them.