Unverified50% confidenceFactExact time
Research by Zhang (2019) demonstrated that standard pooling and downsampling break translation equivariance, and shifting the input by a single pixel can cause non-trivial changes in intermediate feature maps.
1
Sources
50%
Confidence
Long-term
Relevance
8/8/2026
First Seen
Sources
谱池化详解：从频域视角重新理解CNN下采样机制
Related Claims
UnverifiedStable Diffusion、Midjourney、DALL·E 3等主流图像生成模型都存在文字渲染效果差的问题69% similarUnverifiedPixel art generated by AI is often not true pixel art; when enlarged, the pixel grid is uneven, and individual pixel blocks contain gradients, blurred edges, and inconsistent colors.69% similarUnverifiedEarly diffusion models 'drew' character shapes in pixel space rather than truly understanding character structure, leading to problems of missing strokes and misplaced radicals.68% similarUnverifiedHunyuan tends to oversimplify and over-smooth in 3D generation, producing sharp imagery but with noticeable aliasing, indicating that resolution remains insufficient.67% similarUnverifiedDice Loss and IoU Loss are more sensitive to the relative proportion of foreground pixels, naturally alleviating class imbalance; Focal Loss forces the model to focus on difficult pixels by dynamically reducing the weight of easily classified samples.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/709915API
curl https://kongchang.com/api/v1/knowledge/claims/709915MCP
get_claim(id=709915)