@ljwei-stak/dsh-mv-cli

LIMITATIONS

Known limitations

- About 2% of the 1232 reference frames differ from the Python renderer. All of them are in the 75–81 s legacy-mesh section, caused by float-ulp / z-buffer ties. - dsh PV: the timeline is fixed to 211.9 s; other edits need an audio sync offset and different lengths may drift. The AI-assisted dancer is a remake, not hidden original-film material. Windows fonts remain local and glyphs can differ between systems. Legacy 1.0.0 packs remain supported; new raster/font assets require 0.9.5. Only dsh-pv packs allow 64 files / 24 MiB; images stay capped at 1 MiB, JSON at 512 KiB and decoded atlases at 64 Mi pixels. Other packs retain 40 files / 8 MiB. - 用 AI 制作新 MV needs the agent session API of the Harness client (otherwise copy & paste the prompt). Scene scripts run in a Blob Web Worker; if a Harness build forbids blob workers, script packs play with the generic renderer. The agent tools need the Host `tools` service; without it the agent checks its work by reading AGENT.md. - Audio files over 1 GB are refused; the WAV cache is limited to 1.5 GB per file (about 2.5 hours). - The `79c4e5…` offset is inferred. - Automatic timing: recognition quality depends on the mix; fast rap, heavy effects and spoken parts produce yellow lines to check. LRCLIB only knows songs others have uploaded and needs lrclib.net to be reachable (it is skipped with a note otherwise). The CPU-only PyTorch profile was not installed on a test machine; GPU needs an NVIDIA driver for CUDA 12.6. Ja/zh alignment works per character and was only unit-tested. - 创意工坊: the audio fingerprint is coarse (energy envelope only) and can miss or mis-flag edits; lyric retiming works only for lines whose text matches the pack's hashes; GitHub's upload page needs the files dragged in by hand; raw.githubusercontent.com caches for about 5 minutes, so new packs appear with a delay; the CI's `node:vm` run is a check, not a security boundary (the panel's Worker sandbox is).