Unverified70% confidenceOpinionExact time
Diffusion-LM is widely recognized as the pioneering work in the direction of diffusion language models.
2
Sources
70%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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）77% similarUnverifiedThe 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.74% similarUnverifiedDiffusion-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.72% similarUnverifiedText in images generated by early diffusion models such as Stable Diffusion and DALL-E 2 often appeared garbled or blurry.70% similarUnverified生成式模型从早期的GAN发展到扩散模型（如Stable Diffusion）和自回归大语言模型（如ChatGPT）69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/845572API
curl https://kongchang.com/api/v1/knowledge/claims/845572MCP
get_claim(id=845572)