File size: 4,518 Bytes
eafbe80 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 | # 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]
|