{"observation":{"id":"21963d51-7f03-4ec0-8609-3298f72e007f","tool":"gemini","tool_name":"Gemini","criterion":"identity-preservation","criterion_name":"Identity preservation","criterion_definition":"How closely the generated face matches the reference image across scenes.","criterion_evidence_type":"transformation","criterion_rank_role":"decisive","criterion_rank_role_reason":"This is the core of the task: the tool must keep the same character looking like the reference image across scenes. (3 of 3 judges)","scenario":"near-profile-occluded-portrait-stress-test","scenario_name":"Near-profile occluded portrait stress test","group_tag":null,"scenario_description":"Near-profile portrait reference used as the maximum-difficulty stress test: face turned roughly 80–90 degrees, one eye partially occluded by fringe, short dark wavy hair, and natural outdoor lighting. Designed to probe whether tools can preserve identity under severe pose and occlusion.","modality":"image","input_text":null,"input_artifact_refs":[],"stresses":["Near-profile identity preservation","Occlusion handling","Extreme pose consistency","Lighting and hair-detail retention under stress"],"verdict":"failed","score":null,"score_total":null,"note":"Under the near-profile stress test, the tool produces a generic-looking face that no longer closely matches the reference.","evidence_state":"verified","source":null,"artifacts":[{"url":"https://d3epheqghktydj.cloudfront.net/gemini-image-c0417a5d4a92.png","role":"input","alt":null},{"url":"https://d3epheqghktydj.cloudfront.net/gemini-gemini-input3-rooftop-96359a1309c7.png","role":"output","alt":null}],"run_id":"1dfb8fa4-f7f0-47dc-911b-7db3af467e9a","study_title":"Generate Consistent AI Characters Across Different Scenes and Poses","study_kind":"generation","research_task":"86b96df11","tested_at":null,"completeness":"input-and-output","input":{"state":"not-captured","text":null,"files":[],"modality":"image","stresses":["Near-profile identity preservation","Occlusion handling","Extreme pose consistency","Lighting and hair-detail retention under stress"]},"tool_page_slug":"gemini","tool_url":"https://aidemos.com/tools/gemini","permalink":"https://aidemos.com/evidence/21963d51-7f03-4ec0-8609-3298f72e007f","api_url":"https://ai.aidemos.com/v1/observations/21963d51-7f03-4ec0-8609-3298f72e007f"},"peers":[{"id":"7813fc39-8282-479d-adb5-7b09e09a84cc","tool":"chatgpt","tool_name":"ChatGPT","verdict":"worked","score":null,"score_total":null,"note":"Holds the near-profile angle closely and keeps the character recognizable, though the facial features are slightly refined and smoothed compared with the reference.","artifact_count":1,"thumbnail":"https://d3epheqghktydj.cloudfront.net/chatgpt-chatgpt-image-jun-9-2026-11-44-28-pm-48266f3c5749.png","evidence_url":"https://aidemos.com/evidence/7813fc39-8282-479d-adb5-7b09e09a84cc"},{"id":"517d4338-7ba5-41dc-a17c-445ab66e188b","tool":"imagineart","tool_name":"ImagineArt","verdict":"struggled","score":null,"score_total":null,"note":"The rooftop output does not maintain the near-profile stress condition, because the face is rendered more frontally visible than specified.","artifact_count":2,"thumbnail":"https://d3epheqghktydj.cloudfront.net/imagineart-image-3fe56456bc45.png","evidence_url":"https://aidemos.com/evidence/517d4338-7ba5-41dc-a17c-445ab66e188b"},{"id":"27c857ae-8e2f-4012-a206-01a699650a0c","tool":"leonardo-ai","tool_name":"Leonardo AI","verdict":"failed","score":null,"score_total":null,"note":"When the face is rotated past near-profile, the tool over-rotates into a near-back view and identity verification becomes impossible.","artifact_count":2,"thumbnail":"https://d3epheqghktydj.cloudfront.net/leonardo-ai-image-2-2f13c9b713fe.png","evidence_url":"https://aidemos.com/evidence/27c857ae-8e2f-4012-a206-01a699650a0c"},{"id":"c4f4dfac-4f6c-4b1e-9055-19236fe89767","tool":"scenario","tool_name":"Scenario","verdict":"worked","score":null,"score_total":null,"note":"It handles a near-profile reference cleanly, preserving the side-angle jawline, nose profile, and lip shape instead of reverting to a frontal face.","artifact_count":1,"thumbnail":"https://d3epheqghktydj.cloudfront.net/scenario-scenario-input3-rooftop-be5ae814b119.png","evidence_url":"https://aidemos.com/evidence/c4f4dfac-4f6c-4b1e-9055-19236fe89767"}],"other_criteria":[{"id":"9202f072-5120-46d1-99de-222cf57efc23","criterion":"accessory-and-detail-retention","criterion_name":"Accessory & detail retention","rank_role":"decisive","verdict":"struggled","score":null,"score_total":null,"note":"Hair detail weakens under stress, with the output hair flatter and less wavy than the reference.","artifact_count":2,"evidence_url":"https://aidemos.com/evidence/9202f072-5120-46d1-99de-222cf57efc23"},{"id":"f0e88898-addc-4e27-aaf7-43d6c35286fc","criterion":"scene-compliance","criterion_name":"Scene compliance","rank_role":"decisive","verdict":"mixed","score":null,"score_total":null,"note":"The rooftop scene is only partially followed: the outfit, rooftop, skyline, and body pose are present, but the face is too frontal, the occlusion stress condition is missed, and the golden-hour lighting shifts cooler and more daytime-like.","artifact_count":2,"evidence_url":"https://aidemos.com/evidence/f0e88898-addc-4e27-aaf7-43d6c35286fc"}],"appears_in":[{"page_type":"ranking","slug":"consistent-ai-characters","title":"Best AI Tools for Consistent AI Characters Across Scenes and Poses","url":"https://aidemos.com/best/consistent-ai-characters","binding":"run"}],"same_scenario":[{"id":"7813fc39-8282-479d-adb5-7b09e09a84cc","tool":"chatgpt","tool_name":"ChatGPT","verdict":"worked","score":null,"score_total":null,"note":"Holds the near-profile angle closely and keeps the character recognizable, though the facial features are slightly refined and smoothed compared with the reference."},{"id":"517d4338-7ba5-41dc-a17c-445ab66e188b","tool":"imagineart","tool_name":"ImagineArt","verdict":"struggled","score":null,"score_total":null,"note":"The rooftop output does not maintain the near-profile stress condition, because the face is rendered more frontally visible than specified."},{"id":"27c857ae-8e2f-4012-a206-01a699650a0c","tool":"leonardo-ai","tool_name":"Leonardo AI","verdict":"failed","score":null,"score_total":null,"note":"When the face is rotated past near-profile, the tool over-rotates into a near-back view and identity verification becomes impossible."},{"id":"c4f4dfac-4f6c-4b1e-9055-19236fe89767","tool":"scenario","tool_name":"Scenario","verdict":"worked","score":null,"score_total":null,"note":"It handles a near-profile reference cleanly, preserving the side-angle jawline, nose profile, and lip shape instead of reverting to a frontal face."}]}