wcycn/touhou-ai-yolo

Touhou AI YOLO Detector

An experimental 21-class object detector used by Touhou AI, an unofficial visual-control project for Touhou Koumakyou: the Embodiment of Scarlet Devil.

This is a fan-made technical experiment. It is not affiliated with, endorsed by, or sponsored by Team Shanghai Alice, ZUN, or the Touhou Project.

Demo

The overlay is reconstructed from a recorded AI Control session and shows detections, tracked bullet trajectories, collision risk, planned movement, and the keys held by the controller.

Model details

• Task: object detection
• Framework: Ultralytics YOLO / PyTorch
• Checkpoint format: PyTorch .pt
• Training image size: 640
• Classes: 21
• SHA-256: 78eb395d277bb5f35f27025a7bada7725928d6e7f7b15681f659a43b5bf60ab2

Classes

ID	Class
0	boss
1	boss_occluded
2	bullet_enemy
3	bullet_enemy_big_blue
4	bullet_enemy_big_green
5	bullet_enemy_big_red
6	bullet_enemy_big_yellow
7	bullet_enemy_small_blue
8	bullet_enemy_small_green
9	bullet_enemy_small_red
10	bullet_enemy_small_yellow
11	bullet_enemy_unique_blue
12	bullet_enemy_unique_green
13	bullet_enemy_unique_red
14	bullet_enemy_unique_yellow
15	bullet_player
16	character
17	enemy_small_blue
18	enemy_small_red
19	powerup_blue
20	powerup_red

Training data

The repository owner trained this checkpoint on a locally collected and annotated screenshot dataset:

• Total images: 1,834
• Training split: 1,794 images and 1,794 YOLO label files
• Validation split: 40 images and 40 YOLO label files
• Annotation format: YOLO bounding boxes

The raw screenshots are not distributed with this model because they contain game imagery. The repository owner keeps a private archival backup for recovery and reproducibility; it is not a public dataset. A separate final_train folder found in the historical project was produced later through automatic annotation and was not used to train this checkpoint.

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict(
    source="frame.png",
    imgsz=640,
    conf=0.15,
)

The checkpoint is designed for the original project's capture pipeline and should not be assumed to generalize to other Touhou games, resolutions, visual mods, scaling settings, or capture methods.

Known limitations

• No independent, manually reviewed benchmark is currently published.
• Detection quality is uneven across the 21 classes.
• Lasers are not represented as a dedicated class.
• Dense boss patterns and partially occluded player sprites remain difficult.
• Some training samples were augmented; reported sample counts are not a substitute for real-world evaluation.

This model should not be presented as a reliable game-clear system or as a general-purpose Touhou detector.

License

The checkpoint is released under the GNU Affero General Public License v3.0, consistent with the Ultralytics YOLO open-source license.

The Touhou Project, game imagery, names, characters, and other third-party materials remain the property of their respective rights holders and are not licensed by this model repository.