[roadmap] Q4 2026 Workflows and examples: reference video to film by November / 工作流与样例 #263

Parent roadmap: #107. First reference-video-to-film delivery is targeted for November 30, 2026, advancing the parent's early-December target. Owners and capacity remain to be confirmed.

Goal and scope / 目标与范围

Deliver reproducible, editable workflows for a reference MV, a product short and a no-key teaching film. Measure successful creative delivery and partial revision rather than tool-call counts.

交付参考MV、商品短片和免Key教学短片三条可复现、可编辑路径；首条参考视频到成片提前至2026年11月30日，以真实创作和局部修改验收。

Work and milestones / 工作与里程碑

• Mid-November · P0: select rights-cleared reference footage and assets; agree creative requirements, duration, quality and budget. Establish editable shot analysis and identify essential Agent, editing and provider gaps.

• By November 30 · P0: finish reference analysis, storyboard approval, per-shot generation, candidate inspection/adoption, timeline assembly, playable MP4 export and project reopening. Keep measured facts, model judgments and uncertainty distinct. Fix one local shot problem and export again without regenerating unaffected shots.

• First half of December · P0: strengthen partial recovery, shot replacement and reuse; complete product-stills-to-commerce-film and a small no-key teaching scenario.

• By December 31 · P1: extract named recipes from verified workflows and distribute editable samples, provenance/license manifests and bilingual instructions. Record output quality, elapsed time, human interventions, repeated work and resource consumption.

• Align one stable, version-pinned vLLM-Omni generation recipe with the November example. Add ASR-derived timed transcripts to shot/lyric/subtitle editing when a validated endpoint is available; declare an alternative if it is not ready.

• ComfyUI integration follows demonstrated user workflows and is not a dependency of the first film.

• **11月中旬 · P0：**确定授权素材、目标时长、质量和预算，形成可编辑分镜及Agent、编辑、provider必要缺口清单。

• **11月底 · P0：**完成分析、分镜确认、逐镜生成、候选采用、时间线、可播放MP4和重开；区分测量、模型判断和不确定项；修改单镜后再导出，不重做其他镜头。

• **12月上半月 · P0：**加强局部恢复、镜头替换与复用，完成商品图转电商短片及小型免Key教学。

• **12月底 · P1：**由实测样例提炼命名配方，提供可编辑工程、来源/许可清单及双语说明，记录质量、耗时、人工介入、重复工作和资源。

• 11月样例固定稳定的vLLM-Omni版本与配置；可用时用ASR时间信息辅助分镜、歌词与字幕，接口未就绪明确替代路径。

• ComfyUI按真实用户需求推进，不阻塞首条成片。

Implementation and dependencies / 实现影响与依赖

Use normal project objects, commands, jobs and receipts, not a separate workflow ledger. #171 covers reference analysis, not the whole generation/assembly chain. Related: #171, #170, #185, #187, #190, #191, #225. Reconcile overlapping reference-video proposals before new implementation tasks. Cross-project alignment: vllm-project/vllm-omni#8264 and #224.

复用正常项目对象、命令、任务与回执；#171的分析不替代后续生成组装。先协调重复提案，沿用已有实现Issue。

Acceptance / 验收标准

• November reference-film workflow includes approval, real generation, local revision, export and editable reopening. / 11月参考成片包含确认、真实生成、局部修改、导出和重开。
• Three workflows ship editable projects, meaningful assets, licenses and playable outputs. / 三条路径交付可编辑工程、有效素材、许可和成片。
• Costly generation follows authorization; one-shot edits do not unnecessarily regenerate other shots. / 高成本操作遵守授权，局部修改不无故重做其他镜头。
• No-key mode uses honest precomputed/authored media without hidden paid calls or fabricated inference receipts. / 免Key模式如实标注预置素材，无隐藏付费或伪造生成记录。
• Comparable end-to-end quality and efficiency evidence accompanies the examples. / 样例附可比较的端到端质量与效率记录。