InstantX/Qwen-Image-ControlNet-Inpainting
Full fine-tune (fewer layers) vs Qwen/Qwen-Image. InstantX/Qwen-Image-ControlNet-Inpainting matches part of Qwen/Qwen-Image (201 of its 1933 tensors; the rest of Qwen/Qwen-Image has no counterpart). On the shared part: Fine-tune: 93% of the weights changed, across 0 of 0 layers.
Tensors changed: other 27/28, Embeddings 4/4, Attention 92/96, Norms 24/25, MLP 44/48
3 other repos hold byte-for-byte identical weight files.