Unverified50% confidenceFactExact time
iREPA is a class of representation alignment techniques that align the intermediate-layer features of a diffusion model with the features of a pretrained visual encoder (such as DINOv2) to accelerate convergence.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedDINOv2 is a visual foundation model released by Meta AI in 2023, pre-trained on large-scale unlabeled image data using a self-supervised learning paradigm.69% similarUnverifiedDINOv2的训练框架以iBOT为主干，结合DINO的自蒸馏目标和SwAV的聚类一致性损失，并加入masked image modeling任务67% similarUnverified在图像生成中LoRA主要作用于U-Net或DiT架构中自注意力与交叉注意力层的Q、K、V投影矩阵，使RTX 3090（24GB显存）等消费级显卡也能完成模型定制训练66% similarUnverifiedInstructPix2Pix uses GPT-3 to generate image editing instruction pairs, then uses the Prompt-to-Prompt technique to generate before-and-after image pairs in Stable Diffusion, constructing a triplet training set.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026760API
curl https://kongchang.com/api/v1/knowledge/claims/1026760MCP
get_claim(id=1026760)