# 可视化评测配置:以肉眼查看为主,不依赖数值 metric。 # 用途:固定 prompt 与首 chunk/首帧,便于同一条件下对比不同模型或不同设置。 # 使用方式见 VISUAL_EVAL_DESIGN.md;运行示例见 run_visual_eval.py。 # ----------------------------------------------------------------------------- # 1. Prompt 设计(便于观察记忆/一致性/身份/逻辑) # ----------------------------------------------------------------------------- # 原则:英文、与首帧语义匹配时生成更稳定;按“考察维度”分组便于选看。 prompts: # 身份/角色保持(看长时是否同一角色、人脸是否漂移) - id: identity_single_person text: "A single person in a room, clear face and body, consistent lighting." category: identity note: "适合首帧为单人时,观察多 chunk 后是否仍是同一人、五官是否保持。" - id: identity_face_clear text: "Close-up of a person's face, neutral expression, indoor lighting." category: identity note: "强调人脸,便于观察身份保持力。" # 场景/长时一致性(看场景是否塌缩、风格是否一致) - id: scene_indoor_room text: "An indoor room with furniture, walls and floor visible, consistent perspective." category: long_horizon note: "室内场景,观察旋转多 chunk 后房间是否还一致、有无穿模或崩坏。" - id: scene_outdoor_street text: "An outdoor street or building, clear geometry, natural lighting." category: long_horizon note: "室外,观察长时几何与光照一致性。" # 物体/状态(看物体是否合理存在、位置是否连贯) - id: object_furniture text: "A room with a table and chairs, objects stay in place." category: state note: "固定物体,观察状态是否稳定。" - id: object_simple text: "A simple scene with one or two objects, clean background." category: state note: "简单物体,便于观察状态追踪。" # 回环/重访(看回到起点时视角是否一致) - id: loop_revisit text: "A 360-degree view of a space, same view when returning to start." category: loop_closure note: "强调回到起点,便于肉眼看 4chunk 回环末帧与首帧是否像同一视角。" # 通用/兜底 - id: generic_scene text: "A scene." category: generic note: "与现有默认一致,用于对比。" # ----------------------------------------------------------------------------- # 2. 首 chunk / 首帧预设 # ----------------------------------------------------------------------------- # 同一首帧 + 同一 prompt 下对比不同模型,便于控制变量。 # type: fixed_image 表示使用本地图片路径;dataset_frame 表示从数据集取 (video_name, start_frame)。 first_chunk_presets: # 固定图片:需事先准备好,路径可覆盖 - id: fixed_default type: fixed_image path: "" # 留空则用 run_generalization_fixed_first_frame 的默认(如 train/ctx_5_20_per_frame_vae/image.png) note: "默认首帧图,与现有 evals 一致。" - id: fixed_indoor type: fixed_image path: "" # 例如 /path/to/indoor_sample.png note: "建议放一张室内代表图,便于与 scene_indoor_room 等 prompt 搭配。" - id: fixed_outdoor type: fixed_image path: "" note: "建议放一张室外代表图。" - id: fixed_face type: fixed_image path: "" note: "建议放一张人脸/单人图,与 identity_* prompt 搭配。" # 从数据集取首帧:用于与训练分布一致、可复现 - id: dataset_mid type: dataset_frame video_name: "" # 从 metadata 或 frames 目录选一个 start_frame: 0 note: "从数据集指定 (video_name, start_frame) 作为首帧;run_visual_eval 会先导出该帧再跑。" - id: dataset_another type: dataset_frame video_name: "" start_frame: 40 note: "另一段,可改为实际存在的视频名与帧号。" # ----------------------------------------------------------------------------- # 3. 推荐组合(用于快速跑一小组“可视化对比”) # ----------------------------------------------------------------------------- # 只跑这些 (prompt_id, first_chunk_id) 时,可写在这里;留空表示用全部组合或由命令行指定。 recommended_pairs: [] # - [identity_single_person, fixed_face] # - [scene_indoor_room, fixed_indoor] # - [loop_revisit, fixed_default]