BACKGROUND: The use of the Self-Compassion Scale (SCS) as a single measure has been pointed out as problematic by many authors and its originally proposed structure has repeatedly been called into question. The negative facets of this construct are more strongly related to psychopathology than the positive indicators. The aim of this study was to evaluate and compare the different structures proposed for the SCS, including a new measure based only on the negative factors, and to assess the psychometric features of the more plausible solution. METHOD: The study employed a cross-sectional and cross-cultural design. A sample of Brazilian (n = 406) and Spanish (n = 416) primary care professionals completed the SCS, and other questionnaires to measure psychological health-related variables. The SCS factor structure was estimated using confirmatory factor analysis by the maximum likelihood method. Internal consistency was assessed by squaring the correlation between the latent true variable and the observed variables. The relationships between the SCS and other constructs were analyzed using Spearman's r s . RESULTS: The structure with the best fit was comprised of the three negative first-order factors of "self-judgment", "isolation" and "over-identification", and one negative second-order factor, which has been named "self-criticism" [CFI = 0.92; RMSEA = 0.06 (90% CI = 0.05-0.07); SRMR = 0.05]. This solution was supported by both samples, presented partial metric invariance [CFI = 0.91; RMSEA = 0.06 (90% CI = 0.05-0.06); SRMR = 0.06], and showed significant correlations with other health-related psychological constructs. Reliability was adequate for all the dimensions (R ≥ 0.70). CONCLUSIONS: The original structure proposed for the SCS was not supported by the data. Self-criticism, comprising only the negative SCS factors, might be a measure of uncompassionate behaviors toward the self, with good psychometric properties and practical implications from a clinical point of view, reaching a stable structure and overcoming possible methodological artifacts.