A　fixed　SVM　model　setting　is　not　suitable　for　the　evolvement　of　the　pattern　of　user＇s　interest.　In　this　paper　a　relevance　feedback　algorithm　based　on　SVM　model＇s　dynamic　adjusting　for　image　retrieval　is　presented.　In　this　algorithm,　there　is　no　need　to　fix　the　model＇s　parameters　beforehand,　and　the　parameters　of　SVM　model　will　be　automatically　adjusted　corresponding　to　the　changing　of　the　training　samples.　Experimental　results　show　the　proposed　algorithm　outperformed　other　algorithms　with　fixed　model＇s　parameter.　©　2007　IEEE.