Field notes from AI filmmaking
Every article starts with something that actually went wrong or worked on a real production: the rejected prompt, the wasted credits, the shot that finally held together. What failed, why it failed, and what I do now.
A Movie Reference Got My Harmless Scene Rejected
A harmless scene failed because its prompt named a movie. Here is how I replaced the title with specific visual language.
Read the articleLatest articles
Nano Banana Pro Returned a Server Error Before Generating Anything
Nano Banana Pro returned an immediate HTTP 500 until I fixed the aspect ratio and reference image structure.
I Stopped the Image Batch but It Kept Spending Credits
Canceling a background batch on Windows left its child processes alive, generating PNGs and spending credits.
Your Vertical Flyer Fails Before Higgsfield Generates an Image
Higgsfield rejected a 4:5 flyer because direct GPT Image 2 generation supports different ratios than product photoshoots.
My Fictional Character Came Back With a Famous Person's Face
A fictional princess inherited a real famous face because her name overpowered the physical description in the image prompt.
Higgsfield Rejected My Portraits and Still Charged Me
Four harmless portraits were rejected as NSFW, but the attempts still reduced my Higgsfield credit balance.
My Image Timed Out but Higgsfield Had Already Finished It
A Higgsfield timeout can hide a completed, paid image. Here is how I recovered mine without spending another credit.
I Asked for a Clock Without Hands and Got Hands and Numbers
My prompt asked for a clock without hands. The fix was to describe the surface replacing them instead of naming them.
How to Use Four Variations to Diagnose an AI Shot
Treat four AI video variations as a diagnostic batch, inspect every frame, and salvage the usable seconds before regenerating.
Build Prop References Without Hands or Backgrounds
Build clean 3:4 prop references that preserve shape, material, scale, wear, and markings without hands or scenery.
Build Wardrobe References Around What Your Character Always Wears
Group permanent clothing into one 3:4 look, separate optional pieces, and create independent references for each story state.
Build Character Sheets That Give the Clothes as Much Attention as the Face
Build one character sheet per costume, document the headless front view, and approve clothing construction alongside facial identity.
Turn Existing AI Portraits Into a Complete Character Reference Set
Audit existing AI portraits, preserve useful references, and generate only the character views your set is missing.
Give AI Portraits a Distinct Face Before Adding Skin Detail
Review nine portrait decisions in order so realistic skin does not hide an interchangeable AI-generated face.
Use a Floor Plan to Keep Doors and Furniture in Place Across AI Shots
Build a top-down scene map with landmarks and distances, then use it as a spatial reference across changing AI camera angles.
Start Your Next AI Shot From a Frame You Already Have
Capture an approved frame, use it as the next shot’s starting image, and let prop references complete the surrounding environment.
An English Speech Model Invented Japanese Dialogue, and Its Confidence Gate Caught It
A Japanese audio transcription failed, but its confidence gate correctly flagged 23 of 27 lines as doubtful.
How One Duel Shot Splits References Across Characters, a Sword, and a Location
A 14-second duel prompt shows how two characters, one sword, and one location divide the reference work.
Why the Thigh Wound Was Replaced With an Arm Wound
Why an unsupported thigh wound became a referenced arm wound, keeping the same mark consistent across AI video shots.
Why Four Wounds Would Not Fit Into a 25-Second Fight
Four wounds became three marks when the timing showed that each event needed more room to read inside a 25-second AI fight.
What an 808-Word Duel Prompt Keeps After a 63 Percent Cut
How I cut a duel prompt from 2,164 words to 808 while keeping its three impacts, hard dawn light, sword technique, and unfinished exit.
Why a 14-Second Fight Prompt Needs Every Beat on the Clock
My duel prompt was losing its opening four seconds to setup. Exact action beats gave each clash and the unfinished exit a place.
Inside the Available Prompt for a 14-Second Handheld Sword Fight
A close reading of a handheld duel prompt, from first-frame blocking to blade weight, with the missing sections clearly identified.
The 12 Checks This Duel Prompt Was Marked as Passing Before Generation
The twelve checks I marked as passing for a duel prompt, plus the length, timing, and continuity questions they leave open.
Only One of My Recorded AI Film Projects Reached a Verifiable Final Cut
A file audit shows why concepts, image libraries, unfinished edits, and delivered AI films require separate counts.
The 234-Image Library Behind One AI Short Film
The counted image library behind Borrowed Wounds shows how references and storyboard frames serve different production jobs.
Why 820 Successful Log Entries Counted as 805 Unique Image Jobs
A production audit shows why 820 successful log entries became 805 unique image jobs, while only 713 current outputs remained.
Why 45 Video Prompts and About 200 Final Shots Are Both Correct for the Same Film
Why 45 written video prompts and about 200 final shots describe the same AI film without measuring the same work.
Thirty-Eight Video Files Turned Out to Be Only 19 Distinct Reels
Why 38 MP4 files from Borrowed Wounds represent 19 distinct reels, plus what file counts miss about campaign output.
What 196 Character Images and 139 Anchors Do Not Tell You
Why I count 196 character images and 139 anchors without calling either total a count of unique character sheets.
We Reviewed 405 Frames to Audit One Finished AI Film
A timed sample of 405 frames turned the final visual audit of BORROWED WOUNDS into a defined review workload.
Why a Default Image Cost 1.5 Credits While the 2K Version Cost About 1
One default image cost more than a 2K request because the requested model and the model that ran were not the same.
A Coffee Trend Got Nine Times the Views of My Best Film Post
A coffee trend reached 2,743 views while my best film post reached 292. The gap changed how I promote an AI film from a new account.
My Character Sheets Got More Engagement Than My Film's Ending
My character sheets earned 17.2% participation, beating the film's final-shot post and changing what I publish between releases.
My Highest-Engagement Film Post Wasn't a Video
My Instagram data showed carousels earning stronger engagement while reels reached farther, changing how I choose a format.
A One-Line Caption Beat My 80-to-130-Word Story Posts
A one-line caption drew 2,743 views while story captions of 80 to 130 words never passed 292.
“Link in Bio” Was Costing Me Clicks on Threads
Two Threads posts said “link in bio” even though the full film link could have been tapped directly.
My 40 Scheduled Instagram Stories Still Needed Link Stickers Added by Hand
Forty scheduled Instagram stories still needed manual link stickers because the automatic connection could not add them.
27 TikTok Posts Taught Me Not to Optimize Yet
After 27 TikTok posts produced a flat distribution, I decided to wait instead of treating noise as a useful pattern.
I Misread Who Was in the Shot. Three Faces Settled It
I mistook hands for sword handles and identified the wrong character. Comparing three enlarged faces corrected my reading.
No article matches that search yet.