Instructions to use teawhite/ActionRoPE with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Wan2.2
How to use teawhite/ActionRoPE with Wan2.2:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
ActionRoPE(动作走 RoPE 的 Wan2.2-TI2V-5B 世界模型)
《恶意不息》2.5D 俯视角漫游数据(teawhite/EYBX-processed)上训练:动作不走文本,写进每帧 token 的 RoPE 世界坐标;
known/new 支持掩码;文本只含场景。零新增参数(mask 通道 conv 15,360 参数可关)。
| 路径 | 内容 |
|---|---|
DISTILL_HANDOFF.md |
从零搭环境、拉数据、加载权重、前向/训练/评测接口(先读这个) |
code/ |
训练 / 推理 / 评测 / baseline 代码(actionrope/、baseline/、configs/、scripts/、tests/、项目 README、requirements.txt) |
checkpoints/arope_10k_ga2/step-10000.safetensors |
当前最好的权重(10 GB,只含 DiT + arope_mask_embedding.*)+ config.json + 训练/验证日志 |
data/arope_sidecar.pt |
22,782 个 clip 的逐 cell 相机偏移(训练必需,32 MB) |
data/text_table_actionrope.pt |
只含场景的 UMT5 文本表(361 键,1.5 GB) |
data/text_table_eybx_mirror.pt |
动作词翻回真实方向的 UMT5 文本表(plain/prompt 对照臂用,7.5 GB) |
data/gain_fit_y112.json |
相机增益校准结果(重建 sidecar 用) |
data/eval/eybx_bg_50/ |
50 个未见背景(bg50 四科评测) |
训练配方:8×H200、DeepSpeed ZeRO-2、bf16、每卡 batch 1 × 梯度累积 2、lr 1e-5、10k 步(5k + 续训 5k)。 bg50 四科(1×)@5k:增益斜率 0.94、折返 PSNR 19.1 dB、倒退帧 9/50 景、方向误差中位 0.76°、失控 0。
- Downloads last month
- 2,167
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for teawhite/ActionRoPE
Base model
Wan-AI/Wan2.2-TI2V-5B