| The weights in transformer/, octree_decoder/ and gaussian_decoder/ are converted from: |
| Idol, Sperid Labs, Apache-2.0 (see LICENSE). |
| https://huggingface.co/speridlabs/idol |
| Reference revision: 2731c22 |
| The conversion renames modules and splits fused attention projections; the values are unchanged. |
|
|
| image_encoder/ is DINOv3 ViT-H+/16 from: |
| facebook/dinov3-vith16plus-pretrain-lvd1689m, Meta, DINOv3 License (see LICENSE_DINOv3). |
| https://huggingface.co/facebook/dinov3-vith16plus-pretrain-lvd1689m |
| Reference revision: c807c9e |
| It is stored in bfloat16, cast from the original float32 weights; it is otherwise unchanged. |
|
|
| vae/ is the FLUX.2 VAE from: |
| FLUX.2 [klein] 4B, Black Forest Labs, Apache-2.0 (see LICENSE). |
| https://huggingface.co/black-forest-labs/FLUX.2-klein-4B |
| Reference revision: e7b7dc2 |
| Its files are copied unchanged. The original Idol loads the same VAE weights, in float32, from |
| black-forest-labs/FLUX.2-dev. |
|
|
