Short answer: this week our critic kept 69% of the frames it reviewed, up from 46% the week before, and the reasons it rejected the rest changed. The glossy "film still" look, last week's top fault, came up in only 2 of 62 rejection notes. The new top fault was a picture that gives the punchline away before the end (22 of 62 notes). Next came a face that drifted from the character's reference (14) and a look that didn't match the outfit signed off for the post (13).
This is first-hand data from our own pipeline, 2 to 8 November 2026 (UTC): 198 frame reviews of 42 posts by 5 AI influencers. Like last week's digest, it has no pictures of the influencers, only what the critic wrote.
Faces get all the attention, but clothes drift too: a coat that shortens, a fan that opens, a ponytail that comes down. Our AI critic rejected 15 of 59 frames for this in one week. Here are the 7 rules that came out of those misses.
Nothing under 7 was kept, again: all 45 frames scored 2 to 6 went back.
Every 7 was kept (10 of 10). Last week a 7 was a coin flip (46 kept, 25 rejected).
Seventeen frames scored 8 or 9 and were still rejected. Fourteen of those failed a separate face check, and three were a grid of several pictures instead of one photo.
That last line is the interesting one. A frame can look great to the critic and still be the wrong person.
What changed between the two weeks
We changed several things at once, so we can't say which change did what. Between 2 and 8 November the team behind each influencer got:
a face check: every frame with the character in it is compared with her reference photos by a separate face-matching model, not by the critic's eye;
a signed-off look: the stylist's outfit, hair and accessories for the post are written down, and every frame is checked against that list;
a start check: a frame that will become a video clip must show the moment before the action, not after it;
a rule that the payoff is held for the last frame, and the critic checks the plan for it before anything is drawn.
So the faults below are partly new because the checks are new. Last week nobody was counting a coat hem that ended at mid-calf instead of on the boot.
What it rejected, in its own words
We counted how many of the 62 rejection notes mention each fault (by keyword, so a note can count twice and the numbers are rough):
Fault
Notes that mention it (of 62)
Payoff shown too early, or not shown at all
22
Face doesn't match the reference (face check)
14
Not the signed-off look (outfit, hair, accessory)
13
Same framing as another frame
5
Someone or something else in the frame
5
Action already finished (start check)
4
Several pictures in one, a grid
3
Wrong shot size
3
Glossy, rendered look
2
Product wrong
1
A few real notes, lightly trimmed, with names removed:
The spoiled reveal. "The row of boots, which is the whole payoff, is already in frame at the bottom right." The post was a slide along a clothes rail that ends on a row of identical boots. The plan had the boots along the bottom of nearly every frame, so the drop revealed nothing. The critic sent the plan back: no floor in view until the drop.
The payoff that never came. "His wrist is not bound to the bottle neck; he is holding the bottle neck with his bare fingers beneath the bow, leaving his wrist completely free and losing the visual gag." After three rounds the critic held the post with a plain reason: the image model can't draw that knot. That's an honest stop, not a failure to hide.
The face check overruling a 9. "Not her: the face doesn't match her reference (face match 0.36, hers score 0.45 and up)." The critic liked the frame. The face model measured it against her own photos and said no. All 14 face rejections measured between 0.34 and 0.45, at or just under the line set by the character's own photos. How the face scores work, and what they look like for real and AI faces, is in our face consistency guide.
The signed-off look. "Signed closed lace fan across lap, drawn open." A fan that opens in frame one has nothing left to do in frame four.
The start check. "The frame shows him already seated on the chair rather than halfway off the tabletop; the dismount is already finished, so the clip starts with nothing left to animate." A still that becomes a video has to be the moment before.
The wrong person. "A person who is not her is in a frame that should be an object close-up with nobody in it." Image models like to add a model to an empty product shot.
The new top fault: the reveal is open before the end.
The part the critic can't fix
Passing the critic isn't the same as being a good post. Of the 42 posts this week, the owner approved 4. Three were rejected with a reason: one "Looks fake, AI-made" and two "Doesn't make sense at a glance". Two were rejected without one. Nineteen were cleared on purpose to start a clean test, so they were never judged. Four were held by the critic after repeated redraws, seven are still waiting and three expired.
Cover the last frame and look at the rest. If you can guess the punchline, it's already showing. Move the reveal object out of the earlier frames, usually by keeping the floor or the edge of the frame out of view.
Check the face with something other than your eyes. After a few hours you stop seeing drift. Put each frame next to the same reference photo at the same size, or use a face-similarity tool, and redraw anything that looks like a cousin.
Write the look down before you draw. Coat length, hair up or down, which hand holds what, open or closed. Then tick it off frame by frame. Most of our look faults were one item: a fan open, a cape parted, a ponytail let down.
If a still will become a video, draw the moment before. The model animates what happens next. A frame where the action is finished gives it nothing to do.
Know when to stop. If the image model can't draw the gag after three tries, change the gag. Our critic held four posts this week instead of passing a picture that missed the point.
What this means if you are starting an AI influencer
Expect your keep rate to climb as you write things down: the look, the reveal, the reference face. Ours went from 46% to 69% in a week of adding checks. But don't confuse clean frames with a good post. Most of what the owner turned down this week was turned down for the idea, not the picture.
Quick answers
What share of AI influencer frames did the critic keep? 136 of 198 (69%) from 2 to 8 November 2026, up from 127 of 274 (46%) the week before.
What was the most common reason for a rejection? A frame that showed the payoff too early or not at all: 22 of 62 rejection notes.
Why were frames scored 9 still rejected? Fourteen high-scoring frames failed a separate face check against the character's reference photos, and three were grids of several pictures instead of one photo.
Did the glossy "AI look" go away? Almost. It came up in 2 of 62 notes this week against about 3 in 10 the week before, but we changed several things at once, so we can't credit one change.
Does a higher keep rate mean better posts? No. The owner approved 4 of 42 posts this week; the rejections with a reason were about the idea or a fake look, not small faults.
Make yours
If you want a team that writes the look down, checks every face and holds a post it can't get right, that's Influgen: a writer, a creative director, a stylist and a critic behind each influencer. Create your own influencer free: 7-day trial with 50 credits, no card needed. For the label side of posting, the free AI disclosure checker tells you what to say on each platform, and Influgen in numbers has the costs.
Method
Source: our own review logs, every frame reviewed for posts created between 2 and 8 November 2026 (UTC): 198 reviews of 42 posts by 5 AI influencers. A redrawn frame counts as a new review. The same query reproduces last week's 274 reviews and 127 kept.
The critic is a vision model that sees the frame, the character's reference images, the signed-off look and the post's plan. The face check is a separate face-matching model. Frames are drawn with GPT Image 2.5.
Fault counts are keyword counts over the 62 written rejection notes; a note can mention several faults. A rough guide, not a study.
Post outcomes are read from each post's status and the owner's stated reason.
Quotes are from real notes, shortened; character and product names are removed.
One product's pipeline over one week, not a measure of AI images in general.
Influgen gives your influencer a team that plans, creates and posts on its own. Create your own influencer free. The first 1,000 accounts get 100 extra credits.
We made AI influencer dance videos by giving a video model a real dance clip to copy. Every beat matched. Here is the 5-step method, what it cost per second on Kling 2.6 and 3, what broke, and which clips you are allowed to copy.
Before an Influgen influencer posts, a critic reviews every picture. From 26 October to 1 November 2026 it reviewed 274 frames and kept 46%. Here's what it rejected, in its own words, and the seven checks you can run on your own AI influencer content.
