Marigold V2
Copyright 2026 Huawei Technologies Co., Ltd.

Inference graph and preprocessing adapted from:
https://github.com/huawei-bayerlab/marigold-v2
Revision: cc6a7031abcd59fd9e1ceff7fdd0d9687d389bc5
License: Apache-2.0 (LICENSE alongside this notice).
Sources: experiments/20260316_qwen_depth/component_loader.py and network_graph.py,
dataset/dataloading/util.py and transform.py, validation/folder_steps.py and
validate_steps.py. Paths above are relative to upstream marigoldv2/.
LibreYOLO changes: registry-free graph, strict adapter loading, tensor-only
prompt deserialization, checkpoint metadata, isolated inference RNG state,
original-canvas results and explicit depth/normal/albedo conventions.

The original upstream NOTICE credits Marigold (https://github.com/prs-eth/Marigold),
Copyright 2023-2025 Marigold Team, ETH Zurich, Apache-2.0, for reused evaluation,
training and utility code. Those evaluation/training implementations are not
bundled by this family.

Architecture implementations load from separately installed dependencies:
huggingface/diffusers v0.38.0 (Apache-2.0), huggingface/peft v0.18.1
(Apache-2.0), bitsandbytes-foundation/bitsandbytes 0.49.2 (MIT).
No DINOv3 teacher implementation is bundled or used in inference.

Pretrained adapter and prompt tensors are sourced separately from:
https://huggingface.co/huawei-bayerlab/marigold-v2-0
Revision: 6fd6d1ca246c9d2d99a4d8ac375a4eccc87178ad
License: Apache-2.0, as declared in that model repository's LICENSE.
Copyright 2026 Huawei Technologies Co., Ltd.
The converted checkpoints retain inference and prompt tensors unchanged and
omit two named training-only iREPA projection tensors.

The frozen base is downloaded separately from:
https://huggingface.co/Qwen/Qwen-Image-Edit-2509
Revision: d3968ef930e841f4c73640fb8afa3b306a78167e
License: Apache-2.0, as declared by the base model card.
