What Our AI Critic Rejected This Week: 274 Frame Reviews and the Checklist They Make

Two things stood out:

• Nothing scored under 7 was kept. All 114 frames scored 1 to 6 were rejected.
• A 7 was a coin flip. 46 frames scored 7 were kept and 25 were rejected. Even 8 wasn't safe: 65 kept, 8 rejected. Every 9 was kept.

So the score is a guide, but the written reason decides. A pretty frame that breaks the story still goes back.

What it rejected, in its own words

We counted how many of the 147 rejection notes mention each kind of fault (by keyword, so a note can count more than once and the numbers are approximate):

A few real notes from this week, lightly trimmed:

• The film-still look. "The dark vignette and cinematic grade look AI-made, and they do not match the flat bright morning light of frame 1." Real phone photos are flat and a little dull. Image models love drama.
• Continuity. "The loose hair at the edges is dark brown, not her honey blonde, so it reads as a different woman's neck cut in." One wrong strand of hair in a close-up turns the character into someone else.
• A spoiled reveal. "She is already wearing the lace slip and leather biker, so the reveal is spent before the cut and the before/after has nothing to turn on." Before-and-after posts fail when the model draws the "after" in the "before".
• Sound off. "This is the frame that carries the answer, and nothing visible happens in it: the reply is only heard, so muted it is a still-life of a disc." Most people scroll with the sound off. If the point is only spoken, it isn't there.
• The product. "The disc on the collar reads as a plain black coat button." Small products drawn at phone size lose the details that make them recognisable.
• Food that cooks itself. "The food is fully roasted and browned, but one cut earlier it was raw." Each frame can be fine alone and still break the order of time.
• Not the character. "The visible lower face shows a grey beard and no handlebar moustache, so this is a different man."

Seven checks for your own AI influencer posts

You don't need our pipeline to use this. Look at each picture at phone size, with the caption hidden, and ask:

1. Does it look like a phone took it? Flat light, no dark corners, no orange-and-teal grade, no mirror-wet streets. If it looks like a film still, redraw it with plain daylight.
2. Is it the same person as in the last post? Hair colour, hair style, signature jewellery, face. Check the edges of the frame too, where stray hair gives it away.
3. Do the frames agree with each other? Same outfit, same room, same product, same time of day, unless the story changes them on purpose.
4. Is the point visible with the sound off? If the joke or the answer is only spoken, add a picture that shows it.
5. Is the payoff saved for the end? In a before-and-after or a reveal, the first frame must not already show the after.
6. Is the product the real one? Draw it from its real photos, big enough that its shape and marks read at phone size. A product that looks like a button is not in the ad.
7. Does each frame move the story on? Two frames with the same framing read as one picture shown twice.

What this means if you are starting an AI influencer

Expect to throw away about half of what you generate. In the earlier cut we published, 18 to 27 October, in our 672-frame review, the keep rate was 48%; this week it was 46%. Budget for that: at our logged cost of about two cents per picture (see Influgen in numbers), the redraws are cheap. Posting the rejects is what costs you, in followers who notice.

The most useful habit is the boring one: look at every frame next to the last one before you post. If you want a team that does this for you, that is what Influgen is: a writer, a creative director and a critic behind each influencer, and nothing goes out that the critic hasn't passed. For the label side of posting, our free AI disclosure checker tells you what to say on each platform.

Method

• Source: our own review logs, every frame reviewed between 26 October and 1 November 2026 (UTC), 274 reviews of 54 posts by 8 AI influencers. A redrawn frame counts as a new review.
• The critic is a vision model that sees the frame, the character's reference images and the post's plan. Frames are drawn with GPT Image 2.5.
• Fault shares are keyword counts over the 147 written rejection notes, rounded; a note can mention several faults. They are a rough guide, not a study.
• Quotes are from real notes, shortened; product and character names are removed.
• This is one product's pipeline over one week, not a measure of AI images in general.

Most weak product ads aren't badly filmed. They're made by someone who never learned how the product works. Here is the research good creators and agencies do before writing a single line, and what we learned building it into our own pipeline.