Unverified50% confidenceFactExact time
Mainstream text-to-video models (including Seedance, Kling, Sora, etc.) are all based on the diffusion Transformer architecture; each generation is an independent random sampling, and two generations with the same prompt can produce significantly different results.
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDiffusion models are inherently stochastic processes while video requires determinism, and the mathematical tension between the two is the fundamental reason why even the most advanced models still exhibit subject distortion or scene jumps.71% similarUnverifiedThe inherently stochastic sampling of diffusion models means that each generation result differs even given the same prompt.71% similarUnverified主流图像/视频生成模型如Stable Diffusion、Midjourney、Sora缺乏持久化记忆机制，每次生成依赖当前提示词和随机噪声种子，不内置跨次调用的状态保持能力69% similarUnverifiedSeedance 2.0属于视频扩散模型（Video Diffusion Model），底层原理与Stable Diffusion、Sora等模型类似68% similarVerified扩散模型在推理阶段从随机高斯噪声出发，经数十至数百步迭代去噪得到输出，即便固定随机种子，提示词的细微差异也会导致人脸特征、服装细节的显著偏移67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/933517API
curl https://kongchang.com/api/v1/knowledge/claims/933517MCP
get_claim(id=933517)