ViRAP pre-trained models
参考图引导的摄影调色模型权重。代码仓库：https://github.com/edzee3000/ViRAP
Model index
| Model | Batch | Seed | Checkpoint | Validation L1 | Folder |
|---|---|---|---|---|---|
| MAE 末4层及末端LN解冻＋前8层和patch LoRA (r8, alpha16), K=4 | 16 | 17 | 15,328 | 0.0276138513 | 2026-10-30 release |
| MAE 全参数微调, K=4 | 16 | 17 | 15,328 | 0.0261653330 | 2026-10-30 release |
| 冻结MAE＋全12层和patch LoRA (r8, alpha16), K=4 | 16 | 17 | 15,328 | 0.0273889780 | 2026-10-30 release |
Directory convention
<model-family>/batch-<N>/seed-<S>/<release-date>/
不同架构、batch、种子与后续版本分别保存，新增模型使用新目录，不覆盖既有发布。每个发布自带完整权重、配置、源码、推理脚本与指标。当前权重均包含MAE骨干；LoRA方案发布完整模型，全参数方案不使用LoRA；均不是Transformers AutoModel。
Download and run
from huggingface_hub import snapshot_download
folder = "mae-hybrid-lora/batch-16/seed-17/2026-10-30"
snapshot_download("edzee3000/ViRAP-pre-train", allow_patterns=[folder + "/**"], local_dir="ViRAP-models")
cd ViRAP-models/mae-hybrid-lora/batch-16/seed-17/2026-10-30
pip install -r requirements.txt
python infer.py --content input.jpg --reference reference.jpg --output result.png --device cuda
验证集L1用于选取检查点；目标图由PETAL生成，该指标不等同于人类调色偏好。此仓库不包含训练图片、数据清单、优化器或训练恢复状态。
Attribution and license
This complete checkpoint includes the official MAE ViT-B/16 initialization, adapted for ViRAP. MAE is by Meta / Facebook Research, He et al., CVPR 2022: https://github.com/facebookresearch/mae . The upstream MAE license is CC BY-NC 4.0; its attribution and noncommercial terms apply to the included material. See the release's ATTRIBUTION.md and MAE_LICENSE.txt. The initial empty repository's MIT placeholder metadata has been corrected for this full MAE-derived weight release.