Unverified50% confidenceFactExact time
Codebook collapse is a common failure mode of VQ-VAE, meaning only a few codebook entries are used while the rest become dead codes; common mitigations include exponential moving average updates, codebook resets and diversity regularization.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedPosterior collapse refers to the decoder learning to completely ignore the latent encoding and outputting an average result regardless of the input, and is one of the most common failure modes in VAE training.70% similarUnverifiedVQ-VAE uses a straight-through estimator to pass gradients back from the decoder to the encoder, and the training objective typically includes three terms: reconstruction loss, codebook alignment loss, and commitment loss.69% similarUnverifiedRepresentation collapse is the most notorious failure mode of JEPA architectures, where the encoder learns to output constant embeddings independent of the input, reducing prediction loss while making the representations entirely uninformative.66% similarUnverifiedCross-model fallback can bypass provider-level failures but introduces potential issues such as response format inconsistency, context truncation, and billing anomalies.65% similarUnverifiedThe VAE used by Stable Diffusion has been specifically optimized with perceptual loss and adversarial training, and its reconstruction quality far surpasses that of early standard VAEs, but issues such as high-frequency detail loss and color shift still remain65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088797API
curl https://kongchang.com/api/v1/knowledge/claims/1088797MCP
get_claim(id=1088797)