Unverified50% confidenceFactExact time
Diffusion-LM uses learnable embeddings rather than fixed embeddings because the geometric structure of the embedding space needs to mutually adapt to the denoising dynamics of the diffusion process.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe generation pipeline of Diffusion-LM consists of three steps: mapping words to continuous vectors using learnable word embeddings, iteratively denoising via the diffusion model to obtain a sequence of continuous latent variables, and restoring continuous vectors to discrete text through a Rounding operation.73% similarUnverifiedStable Diffusion 基于去噪扩散概率模型（DDPM）理论，将扩散过程从像素空间转移到低维潜在空间，由VAE负责编解码，CLIP模型将文本提示编码为语义向量72% similarUnverified离散扩散语言模型主流方案包括：将token映射到连续嵌入空间进行扩散（如Diffusion-LM），或设计离散扩散过程用掩码模拟加噪（如D3PM、MDLM）71% similarVerifiedStable Diffusion is based on the Latent Diffusion Model and generates high-quality images by performing a denoising process in a compressed latent space69% similarVerifiedDiffusion models circumvent the instability of adversarial training through a Markov chain mechanism that progressively adds Gaussian noise and then trains the model to learn the reverse denoising process.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843368API
curl https://kongchang.com/api/v1/knowledge/claims/843368MCP
get_claim(id=843368)