a243f1d9cf
fix: ktv_adapter model names — remove ky- prefix to match actual Sage llm records
...
- All 8 GPU service models now match tokentest llm names
- Add supporting app modules (quality_gate, model_selector, global_func, eval_music, eval_video)
2026-07-08 18:05:13 +08:00
212a0c1f85
fix: ktv_adapter pipeline_service导入改为可选
2026-07-07 22:14:10 +08:00
216f2f9f59
fix: demucs固定separate_full模式
2026-07-06 17:26:40 +08:00
bee01f6c3e
fix: wan2.6→wan2.7 作为推荐默认模型
2026-07-06 17:23:57 +08:00
225c654811
fix: 模型选择读取客户选项+定价改为llmage模型(wan2.6/happyhorse/kling)
2026-07-06 17:20:30 +08:00
3493444307
refactor: KTV产线全部改为通过llmage API调用，环境变量配置供应商/端点/APIKey
2026-07-06 17:11:51 +08:00
ffdba55fd7
feat: add full 4-stem demucs mode with MIDI instrument replacement
...
- Support mode='full' in handle_demucs_separating config
- Full mode: 4-stem separation → basic-pitch MIDI → fluidsynth render → mix
- Each instrument gets correct GM instrument (drums=channel10, bass=34, other=1)
- 2-stem mode preserved as default for backward compatibility
- Uses run_full.py standalone script on GPU server via SSH
2026-07-06 11:30:43 +08:00
8408570ed7
feat: add multi-model cost estimation and selection after storyboard
...
- Add cost_estimator.py: supports Wan 2.2, Wan 2.7, Vidu 2.0 pricing
- Add model_selector.py: creates human_task for customer model selection
- Update ktv_adapter.py: add model_selecting step, modify scene_video_generating to use selected model
- Add integration guide: docs/model-selection-integration.md
2026-06-25 18:51:29 +08:00
9d3f0ff38c
feat: 多维度质量评估体系 - 视频5维度/音乐7维度
...
新增模块:
- app/eval_video.py: 场景视频多维度评估 (视觉质量25%, 运动流畅度20%, 语义一致性30%, 角色一致性15%, 时间连贯性10%)
- app/eval_music.py: 音乐多维度评估 (节奏15%, 和弦10%, 听感20%, 情绪-歌词匹配20%, 指令匹配15%, 伴奏10%, 人声10%)
更新模块:
- app/quality_gate.py: 集成多维度评估函数，支持降级
- app/ktv_adapter.py: 注册6个质量门控 handlers (步骤9-10, 11, 12, 13, 14-15, 17)
质量门控流程:
1. 执行 handler
2. 多维度评估
3. 不达标自动重试(最多3次)
4. 仍不达标创建 human_task 暂停任务等待人工决策
2026-06-25 18:34:26 +08:00
f767f22717
feat: add KTV adapter with 17 step handlers
...
- Restore handlers_ktv.py as ktv_adapter.py (adapter pattern)
- Wire load_ktv_adapter() into pipeline_app.py init
- GPU ports: Wan2.2:8080, RealESRGAN:9082, fastwhisper:9925
- Handlers: audio/video preparing, demucs, lyrics, subtitles,
music gen, character design, storyboard, scene video, KTV synthesis
2026-06-25 14:26:19 +08:00
