Unverified50% confidenceFactExact time
Continuous diffusion models suffer from a temporal compression problem stemming from the misalignment between geometric trajectory and temporal progression, causing token identity decay to concentrate in a narrow window near the noise endpoints.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified主体漂移本质上是模型在时序一致性建模上的不足，源于扩散模型每帧去噪引入随机性叠加时序注意力对长程依赖捕获能力有限79% similarUnverifiedContinuous-space diffusion first maps discrete tokens to a continuous embedding space to perform Gaussian noise diffusion and denoising, then maps back to the discrete vocabulary, but the discretization step may introduce error accumulation.74% similarUnverifiedWhen all models approach a perfect score, the benchmark loses its discriminative power and rankings become noise.73% similarUnverifiedWhen diffusion models directly infer at higher resolutions, the receptive field of the denoising network cannot cover the global structure, leading to issues such as multiple heads, limb misalignment, or texture repetition.72% similarUnverifiedFew-shot prompts are themselves much longer, and the mere difference in length is sufficient to shift the model's representations, meaning that previously measured drift confounded length noise unrelated to the content of the examples.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098601API
curl https://kongchang.com/api/v1/knowledge/claims/1098601MCP
get_claim(id=1098601)