Round 2 · checks done, blind panel grading6 models · 5 people · 2 attempts eachBy Yicheng ZhouTested 31 Oct 2026 on Higgsfield, graded blind 1 Nov 2026
The short answer
Every model produced a usable LinkedIn photo from every casual snapshot, so the real question is whether it still looks like you. Measured with face recognition, GPT Image 2.5 kept the closest likeness (0.92 on average, where 1.00 is identical) and is also the cheapest, at 0.5 credits. Grok Image 2.0 (0.88) and FLUX 3 Image (0.86) came next. Nano Banana 2.1 (0.77) and Ideogram 4.5 (0.75) drifted most, smoothing faces into a more generic look.
Results
Two attempts per model per brief, no cherry-picking. A headshot passes when it is a usable LinkedIn photo (head and shoulders, the navy blazer, a plain background) and face recognition says it is still clearly the same person. Likeness is the average face-recognition similarity to the source photo: 1.00 would be identical.
Ideogram 4.5Ideogram · medium quality, magic prompt off, 1K
8 / 10reliability 90%
0.75
2
2
The blind panel is grading these outputs now. BiBi Scores, which add the panel's looks score and adjust for each grader's habits, appear here when grading finishes. How scoring works.
Every output
Unedited, both attempts, in the order the models were run. Click an image to see it full size.
Source photo 1
PromptTurn the person in this photo into a professional corporate headshot for LinkedIn: head and shoulders, facing the camera, wearing a navy blazer over a plain white top, soft even studio lighting, plain light grey background, natural skin texture, gentle confident expression. Keep the person's face exactly the same: same identity, face shape, eyes, nose, skin tone and hair. No text.
PromptTurn the person in this photo into a professional corporate headshot for LinkedIn: head and shoulders, facing the camera, wearing a navy blazer over a plain white top, soft even studio lighting, plain light grey background, natural skin texture, gentle confident expression. Keep the person's face exactly the same: same identity, face shape, eyes, nose, skin tone and hair. No text.
PromptTurn the person in this photo into a professional corporate headshot for LinkedIn: head and shoulders, facing the camera, wearing a navy blazer over a plain white top, soft even studio lighting, plain light grey background, natural skin texture, gentle confident expression. Keep the person's face exactly the same: same identity, face shape, eyes, nose, skin tone and hair. No text.
Source photoThe input every model receivedGPT Image 2.5Attempt 1PassGPT Image 2.5Attempt 2PassNano Banana 2.1Attempt 1PassNano Banana 2.1Attempt 2Small slipUsable headshot, but likeness only 0.65Seedream 5.0 ProAttempt 1PassSeedream 5.0 ProAttempt 2PassFLUX 3 ImageAttempt 1PassFLUX 3 ImageAttempt 2PassIdeogram 4.5Attempt 1Small slipUsable headshot, but likeness only 0.66Ideogram 4.5Attempt 2Small slipUsable headshot, but likeness only 0.62Grok Image 2.0Attempt 1PassGrok Image 2.0Attempt 2Pass
Source photo 4
PromptTurn the person in this photo into a professional corporate headshot for LinkedIn: head and shoulders, facing the camera, wearing a navy blazer over a plain white top, soft even studio lighting, plain light grey background, natural skin texture, gentle confident expression. Keep the person's face exactly the same: same identity, face shape, eyes, nose, skin tone and hair. No text.
PromptTurn the person in this photo into a professional corporate headshot for LinkedIn: head and shoulders, facing the camera, wearing a navy blazer over a plain white top, soft even studio lighting, plain light grey background, natural skin texture, gentle confident expression. Keep the person's face exactly the same: same identity, face shape, eyes, nose, skin tone and hair. No text.
Five synthetic people from our Reference Kit, each in an ordinary phone photo: a park selfie, a car selfie, a beach photo, a kitchen photo and an office selfie. No real person's photo was used. Every model that takes a reference image got the same photo and prompt, twice, at about 1K, on Higgsfield on 3 Nov 2026. Recraft V4.1 doesn't take a reference image there, so it isn't tested.
Likeness is measured by machine: ArcFace face embeddings (InsightFace) compare each headshot with its source photo. Our blind panel then grades whether the person would actually use the photo.
