DreamSim (HuggingFace format) — unofficial port
Copyright (c) 2026 bigshanedogg. MIT License (see LICENSE).

This is a DERIVATIVE, re-packaged distribution and is NOT an official release by the
DreamSim authors.

Attribution and third-party notices
------------------------------------

1. DreamSim (MIT)
   Copyright (c) 2023 Shobhita Sundaram, Netanel Tamir, Stephanie Fu, Richard Zhang
   https://github.com/ssundaram21/dreamsim
   The architecture, ensemble configuration, and released checkpoints originate here.
   The weights in `model.safetensors` are the DreamSim "ensemble" LoRA adapters merged
   into their base backbones.

2. DINO — Apache License 2.0 (see LICENSE.apache-2.0.txt)
   Copyright (c) Facebook, Inc. and its affiliates.
   https://github.com/facebookresearch/dino
   The ViT backbone code in `modeling_dreamsim.py` is adapted from DINO's
   `vision_transformer.py`. MODIFICATIONS: vendored into a single module, trimmed to
   the inference path, restructured as a `transformers.PreTrainedModel`, and wired to
   load merged safetensors weights. Per Apache-2.0 §4(b), these files carry notices
   that they were changed.

3. Backbone pretraining weights (merged into `model.safetensors`)
   * DINO ViT-B/16       — Apache-2.0 (Meta AI)
   * CLIP ViT-B/16       — MIT (OpenAI, https://github.com/openai/CLIP)
   * OpenCLIP ViT-B/16   — MIT code (https://github.com/mlfoundations/open_clip);
                           pretraining data is LAION-based.
   Review these upstream terms — and, for OpenCLIP, the LAION data provenance —
   before any commercial use.