{"id":1193175,"date":"2026-11-04T10:51:11","date_gmt":"2026-11-04T17:51:11","guid":{"rendered":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/publication\/logit-conditioned-diffusion-decoding-for-frozen-discrete-token-vlms\/"},"modified":"2026-11-04T12:46:10","modified_gmt":"2026-11-04T19:46:10","slug":"logit-conditioned-diffusion-decoding-for-frozen-discrete-token-vlms","status":"publish","type":"msr-research-item","link":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/publication\/logit-conditioned-diffusion-decoding-for-frozen-discrete-token-vlms\/","title":{"rendered":"Logit-Conditioned Diffusion Decoding for Frozen Discrete-Token VLMs"},"content":{"rendered":"\n\n\n<p class=\"wp-block-paragraph\">Pre-trained discrete-token vision-language models (VLMs) generate images by sampling sequences of codebook indices, but their visual fidelity is bounded by the VQ-VAE decoding stage. Prior work addresses this bottleneck by replacing the discrete tokenizer with continuous or hybrid visual token representations, or by integrating diffusion into the generation pipeline; both routes require foundation-scale retraining or realignment of VLM, which is undesirable when the VLM&#8217;s broader capabilities should be preserved or when retraining is infeasible. We show that such retraining is unnecessary: before each token is sampled, discrete-token VLMs already compute a full logit distribution over the codebook, but standard decoding discards it after selecting a single index. Under Otsu&#8217;s threshold, the pre-sampling logits of capable VLMs assign statistically meaningful probability to at least two codebook entries per token, and recovering this distribution can substantially close the fidelity gap. We propose DCDD, a post-hoc framework that conditions a diffusion decoder on this discarded distribution while leaving the VLM frozen. At inference, DCDD converts the VLM&#8217;s pre-sampling logits into distribution-weighted code vectors; during training, it uses VQ-VAE encoder-derived proxy logits calibrated to match the distributional statistics of the VLM&#8217;s inference-time logits. Our diffusion decoder then maps these representations to high-fidelity images as an optional alternative to the native VQ-VAE decoder. Trained on ImageNet-1K for 50K steps, DCDD reduces reconstruction FID by 74% and Janus-Pro 7B&#8217;s generation FID on MJHQ-30K by 34%, while semantic-alignment across GenEval, DPG-Bench, and WISE remain consistent. Comparison against a diffusion model from the same backbone family applied as an image-to-image refiner confirms that the gain comes not from the diffusion model alone, but from conditioning on the VLM&#8217;s pre-sampling logit distribution.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Pre-trained discrete-token vision-language models (VLMs) generate images by sampling sequences of codebook indices, but their visual fidelity is bounded by the VQ-VAE decoding stage. Prior work addresses this bottleneck by replacing the discrete tokenizer with continuous or hybrid visual token representations, or by integrating diffusion into the generation pipeline; both routes require foundation-scale retraining or [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-author-ordering":[{"type":"text","value":"Ji Woo Hong","user_id":0},{"type":"text","value":"Hee Suk Yoon","user_id":0},{"type":"text","value":"Gwanhyeong Koo","user_id":0},{"type":"text","value":"Eunseop Yoon","user_id":0},{"type":"text","value":"SooHwan Eom","user_id":0},{"type":"user_nicename","value":"Qi Dai","user_id":"36689"},{"type":"user_nicename","value":"Chong Luo","user_id":"31450"},{"type":"text","value":"Chang Yoo","user_id":0}],"msr_publishername":"","msr_publisher_other":"","msr_booktitle":"","msr_chapter":"","msr_edition":"","msr_editors":"","msr_how_published":"","msr_isbn":"","msr_issue":"","msr_journal":"","msr_number":"","msr_organization":"","msr_pages_string":"","msr_page_range_start":"","msr_page_range_end":"","msr_series":"","msr_volume":"","msr_copyright":"","msr_conference_name":"NeurIPS 2026","msr_doi":"","msr_arxiv_id":"","msr_mag_id":"","msr_other_authors":"","msr_other_contributors":"","msr_speaker":"","msr_award":"","msr_affiliation":"","msr_institution":"","msr_host":"","msr_version":"","msr_duration":"","msr_release_tracker_id":"","msr_highlight_type":"","msr_date_display_format":"","msr_main_download_label":"","msr_external_link_label":"","msr_doi_label":"","msr_published_date":"2026-10","msr_startdate":"","msr_presentation_date":"","msr_highlight_text":"","msr_notes":"","msr_longbiography":"","msr_publicationurl":"","msr_external_url":"","msr_secondary_video_url":"","msr_conference_url":"","msr_journal_url":"","msr_year":2026,"msr_month":10,"msr_day":0,"msr_microsoftintellectualproperty":false,"msr_pub_id":"","msr_publication_uploader":[{"type":"url","title":"https:\/\/neurips.cc\/virtual\/2026\/poster\/153860","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_original_fields_of_study":[],"msr_s2_paper_id":"","msr_s2_pdf_url":"","msr_citation_count_updated":"","msr_citation_count":0,"msr_influential_citations":0,"msr_reference_count":0,"msr_s2_open_access":false,"msr_s2_author_ids":[],"msr_pub_ids":[],"msr_hide_image_in_river":0,"footnotes":""},"msr-research-highlight":[],"research-area":[13556,13562],"msr-publication-type":[193716],"msr-publisher":[],"msr-publication-cta":[],"msr-focus-area":[],"msr-locale":[268875],"msr-post-option":[],"msr-field-of-study":[263185],"msr-conference":[259048],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1193175","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-research-area-computer-vision","msr-locale-en_us","msr-field-of-study-computer-vision-and-pattern-recognition"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2026-10","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":0,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"url","title":"https:\/\/neurips.cc\/virtual\/2026\/poster\/153860","label_id":243109,"id":false,"viewUrl":false}],"msr_related_uploader":[],"msr_citation_count":0,"msr_citation_count_updated":"","msr_s2_paper_id":"","msr_influential_citations":0,"msr_reference_count":0,"msr_arxiv_id":"","msr_s2_author_ids":[],"msr_s2_open_access":false,"msr_s2_pdf_url":null,"msr_attachments":[],"msr-author-ordering":[{"type":"text","value":"Ji Woo Hong","user_id":0,"rest_url":false},{"type":"text","value":"Hee Suk Yoon","user_id":0,"rest_url":false},{"type":"text","value":"Gwanhyeong Koo","user_id":0,"rest_url":false},{"type":"text","value":"Eunseop Yoon","user_id":0,"rest_url":false},{"type":"text","value":"SooHwan Eom","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Qi Dai","user_id":36689,"rest_url":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Qi Dai"},{"type":"user_nicename","value":"Chong Luo","user_id":31450,"rest_url":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Chong Luo"},{"type":"text","value":"Chang Yoo","user_id":0,"rest_url":false}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[],"msr_group":[],"msr_project":[],"publication":[],"video":[],"msr-tool":[],"msr_publication_type":"inproceedings","related_content":[],"_links":{"self":[{"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1193175","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":3,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1193175\/revisions"}],"predecessor-version":[{"id":1193272,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1193175\/revisions\/1193272"}],"wp:attachment":[{"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1193175"}],"wp:term":[{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=1193175"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1193175"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=1193175"},{"taxonomy":"msr-publisher","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-publisher?post=1193175"},{"taxonomy":"msr-publication-cta","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-cta?post=1193175"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=1193175"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1193175"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1193175"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=1193175"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=1193175"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=1193175"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1193175"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1193175"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}