Unverified50% confidenceTradeoffExact time
CFG requires two forward computations during sampling and sacrifices a certain amount of diversity.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedClassifier-free guidance (CFG) runs both a conditional branch and an unconditional branch simultaneously, amplifying the conditional signal by subtracting the unconditional direction from the conditional direction, trading a doubling of inference cost for quality improvement.69% similarUnverifiedUsing excessively high CFG values during low-step inference amplifies the correction magnitude at each step, leading to oversaturated colors, distorted contrast, and even artifacts; when using an accelerated LoRA, the CFG value often needs to be reduced from 7.5 to 2–4.68% similarUnverifiedThe study proposes correcting CFG in the early flow steps: reducing or reweighting the CFG influence during the noise stage, then gradually restoring standard CFG; this method requires no retraining of the model and adds no inference cost, being completed entirely within the sampling stage.64% similarUnverifiedMany existing SMC (Sequential Monte Carlo) inference-time sampling schemes rely on equal-weight resampling, which aggressively prunes low-weight trajectories and causes a sharp decline in lineage diversity of the search space.63% similarUnverifiedCFG guidance strength typically seeks the best balance between 5 and 12; too low deviates from the prompt, while too high produces oversaturation or artifacts.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1046703API
curl https://kongchang.com/api/v1/knowledge/claims/1046703MCP
get_claim(id=1046703)