Unverified50% confidenceFactExact time
The actor's input is strictly identical to that of the real game client: about 1.4k-dimensional features, two 24x24 local grids, and a minimap, processed by a small CNN, with set encoding for up to 12 visible enemies, then fed into a 256-dimensional GRU for temporal memory.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedTechnologies such as IP-Adapter and ControlNet can be used to encode a character's facial features as reference vectors and inject them into subsequent generation processes to achieve cross-shot character locking.66% similarUnverified1280×720分辨率在模型训练数据集中出现频率极高，因为大量网页截图和UI录屏素材天然就是这个尺寸62% similarUnverifiedInteractive content relies on a real-time computation loop (Game Loop/Render Loop) that ideally reads user input, updates state, and renders output 60 times per second—a characteristic that pre-encoded video does not possess.60% similarVerifiedSolutions for enhancing character consistency in AI video include reference image injection (IP-Adapter), LoRA fine-tuning for specific characters, and controlling random seeds and attention mechanisms.60% similarUnverified在图像生成中LoRA主要作用于U-Net或DiT架构中自注意力与交叉注意力层的Q、K、V投影矩阵，使RTX 3090（24GB显存）等消费级显卡也能完成模型定制训练60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1021419API
curl https://kongchang.com/api/v1/knowledge/claims/1021419MCP
get_claim(id=1021419)