[shoyofight]

• 新增统一评测编排器 eval_all.py：按任务路由 T2I/edit/video 指标， 模型只加载一次，从 config.resolved.yaml 权威读真实 prompt，质量分回写
• 新增 edit/video 专属指标：directional CLIP、相邻帧时序一致性
• 新增全指标出表 make_matrix_tables.py（DiT 纯采样独立列 + 端到端标口径）
• cal_ir.py 内联 transformers>=5 自足兼容层（干净安装可跑）
• src 组件级 phase marker（ED_BENCHMARK_MARKERS gate，默认零开销）
  • model_runtime offload 无 max-vram 时自动降级修复
• 补 benchmark/requirements.txt 与 README 评测章节

Summary

Describe the change and the affected model/backend/API surface.

Validation

• CPU build or tests
• CUDA build or tests, if relevant
• Python tests, if relevant
• Correctness or image/video quality check, if generation behavior changed
• Benchmark numbers, if performance changed

Notes

Include hardware, model paths in generic form, and any known limitations. Do not
include secrets, private paths, model weights, or generated large artifacts.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.