multishot / CHANGELOG.md
PencilHu's picture
Upload folder using huggingface_hub
85752bc verified

Changelog

以下记录了项目中所有值得关注的变更内容,其格式基于Keep a Changelog

本项目版本遵守Semantic VersioningPEP-440

Unreleased

Added

  • 暂无

Changed

  • 暂无

1.12 - 2026-01-12

Added

  • multi-shot/MULTI_SHOT_CORE_SUMMARY.md:多镜头核心流程总结文档
  • multi-shot/MODIFICATION_LOG.md:修改清单与前后对比记录
  • multi-shot/dry_run.py:生成 dummy 数据 + 数据集读取 dry-run
  • multi-shot/dry_run_train.py:构造 dummy batch 并跑 WanTrainingModule.forward_preprocesstraining_loss
  • multi-shot/tmp_dryrun/dummy_dataset.json:配套 dry-run 的 dummy 数据
  • multi-shot/OVERFIT_TROUBLESHOOTING.md:overfit 排查指南与常见问题
  • multi-shot/overfit_infer_debug.py:overfit 推理与调试脚本
  • multi-shot/check_dataset.py:导出训练输入样本(视频、ID 图、caption)

Changed

  • multi-shot/process_data.py:输出 shot_captions;补齐 id_dir/id_names 与拼接视频的产出
  • multi-shot/multi_view/datasets/videodataset.py:训练使用 shot_captions/text 生成 pre_shot_caption;固定取前 5 秒片段;支持 id_dir/id_names 作为 ref 图
  • multi-shot/overfit_infer_debug.py:默认 50 步、默认纯 prompt;支持 split=all、可选保存输入视频;自动转换 final_model/model.safetensors
  • multi-shot/multi_view/conf/multi-view.yamlsave_epoches 调整为 50;debug_infer 默认不使用 input_video
  • multi-shot/multi_view/train.py:读取 YAML 中 model_id_with_origin_paths/trainable_models/learning_rate 与 debug_infer 配置
  • multi-shot/multi_view/DiffSynth-Studio-main/diffsynth/trainers/utils.py:debug_infer 运行与恢复;loss 输出路径修正;新增按 epoch 保存 checkpoint
  • multi-shot/multi_view/DiffSynth-Studio-main/diffsynth/pipelines/wan_video_new.py:prompt 处理与 shot mask 逻辑补齐,空文本时回退
  • multi-shot/multi_view/DiffSynth-Studio-main/diffsynth/models/wan_video_dit.py:shot attention 与 mask 处理更稳健
  • multi-shot/multi_view/DiffSynth-Studio-main/diffsynth/models/wan_video_text_encoder.py:新增 state_dict_converter
  • multi-shot/multi_view/DiffSynth-Studio-main/diffsynth/trainers/utils.py:新增 --shot_rope CLI 参数

0.1.0 - 2025-12-05

Added

  • 创建项目