Generative video artifacts are frame-to-frame failures: a face changes, a hand gains a finger, a label becomes nonsense, or an object ignores gravity. The skill is naming the defect precisely enough to choose the cheapest reliable repair.

By the end, you will have a repaired three-clip set and a log of bad frames, causes, repair choices, prompts, and acceptance decisions.

// concept

Diagnose the Artifact Before Spending Another Generation

Watch each clip at normal speed, frame-by-frame around the suspicious moment, and in the final phone crop. Record a timecode range, not “looks weird.” A defect outside a vertical crop differs from a face that melts in the centre.

Artifact family	What to look for	Likely pressure on the generation	Targeted first fix
Identity drift	Face, clothing, product colour, or logo changes	Appearance details compete across a long action	Hold one identity description constant; simplify action; use an approved reference workflow if supported
Motion morph	A cup becomes a bowl while turning; a bag merges with a hand	Occlusion or a complex rotation forces the model to invent hidden geometry	Shorten the action; keep the object visible; split “pick up” and “turn” into separate shots
Anatomy	Extra fingers, fused wrists, changing teeth, uneven eyes	Hands or faces are small, fast, or hidden	Reframe to forearms, profile, or wider; replace when anatomy is the focus
Text gibberish	Labels, signs, screens, or Urdu letters mutate	The model is being asked to render exact typography across frames	Generate a blank sign or clean surface; add exact text later in the editor
Physics glitch	Liquid rises, wheels slide, shadows move against the light	Multiple interacting objects or unclear cause-and-effect	State one physical action and one direction; lock the camera; separate the event into shorter clips
Style flicker	Lighting, grain, colour temperature, or illustration style pulses	Conflicting style words or a scene change inside one generation	Remove competing descriptors; reuse one style suffix; cut before the visible switch
Texture popping	Fabric weave, product pattern, or background detail crawls	Fine repeating detail changes faster than the model can preserve it	Use a simpler material/background, reduce camera motion, or crop away the unstable texture

Prompt only what the shot needs. Lifting, tying, placing, and orbiting together combine anatomy, occlusion, object permanence, and camera motion. Split them into single-action shots. Google’s Veo guide recommends explicit framing, motion, style, lighting, location, and action. Check current official pages for models, edit controls, regional access, and credits.

// concept

Choose Reroll, Isolate, Replace, or Edit Around

Use this decision tree instead of rerolling by habit:

Does it alter identity, meaning, consent, safety, or the main product? Replace or reroll from a simpler prompt; never hide a changed face or product.
Is it confined to a short range, with a suitable edit available in your current Flow model? Isolate that range and request one change. Flow documents version History, but feature and region support vary—confirm the current page. Otherwise regenerate the short shot.
Is it outside the important action? Crop it, cover it with relevant B-roll, or speed-ramp only when the faster motion still looks intentional.
Does it repeat across most frames? Replace the concept, reduce interactions, shorten the clip, or remove in-scene text.

Reject a changing hero product, mutating text, central facial/anatomical errors, reversed physics, or visible flicker. A background error may pass only outside the crop, beneath a motivated cutaway, or genuinely unseen at feed size. A small screen never excuses concealing a material alteration.

With limited quota, duplicate the prompt, change only one failure driver, and compare. Use an existing cutaway for a brief background wobble instead of regenerating the whole shot. Check Flow’s current per-generation credit cost before committing.

// concept

Keep a Targeted Repair Log

Use one row per defect, even when a clip has two failures.

Clip/version	Timecode	Taxonomy	Severity	Evidence	Decision	One change made	Result
`C02-v1.mp4`	`00:02.4–00:03.1`	Text	Kill	label characters mutate	Edit around	generate blank label; overlay type in edit	pending

Severity is kill, repair, or pass. Evidence must be observable—“right wrist fuses for 11 frames,” not “bad anatomy.” Mark accepted, rejected, or pending after phone review. Keep rejected versions until approval.

// worked_example

Worked Example

This hypothetical sample for an imaginary Lahore stationery seller claims no performance result. The deliverable is three vertical clips: parcel reveal, ribbon detail, and dispatch.

The draft prompts were overloaded. Clip A requested a readable label and produced changing letters in this sample. Clip B combined fingers, ribbon tying, and an orbit; the sample defect is fused fingers. Clip C moved a motorcycle, traffic, parcel, and tracking camera together; the sample defects are wheel slide and flicker.

Bad repair prompt:

// prompt — copy me2 lines

Make the same video perfect. No bad hands, no weird text, no flicker, no morphing,
perfect physics, perfect product, cinematic and flawless.

That request identifies neither the failing frames nor the cause. Replace one risky instruction per clip:

// prompt — copy me3 lines

CLIP A — static close-up, a plain kraft parcel on a clean packing table.
The parcel has one empty cream label with no writing or symbols. Slow camera push-in.
Soft daylight from the left, warm realistic commercial B-roll, stable brown-and-cream palette.

// prompt — copy me3 lines

CLIP B — macro side view of a green ribbon already tied around a kraft parcel.
One adult forearm gently straightens the loose ribbon tail; fingertips remain outside frame.
Locked camera, shallow depth of field, soft daylight, one slow movement, no text.

// prompt — copy me3 lines

CLIP C — wide rear three-quarter view of a stationary delivery motorcycle beside a closed gate.
A plain kraft parcel is secured flat on the rear rack. The rider remains still; only a tree leaf moves
slightly in the breeze. Locked camera, overcast daylight, realistic documentary B-roll, no logos or text.

Clip A reserves its blank label for an editor overlay. Clip B shoots around anatomy by keeping fingertips outside frame. Clip C replaces complex driving physics with a stable dispatch image. If Clip A flickers only in its final half-second, trim it or cover it with Clip B; do not speed-ramp Clip C to imply motion it never contained.

The log records text → edit around, anatomy → reframe and reroll, and physics/flicker → replace concept, plus prompt versions and phone review. “Fixed” alone says nothing.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

Rerolling the unchanged prompt
the same complexity remains. Change one cause—action count, framing, text, occlusion, or camera motion—and version the prompt.
Adding a long negative list
state the single subject, action, framing, and stable details positively.
Repairing text inside the generation
use a blank surface and add approved English or Urdu type in the editor.
Cropping away evidence that changes the offer
hiding a product variation or a changed face is misleading. Reject the clip when the defect affects meaning or identity.
Using a speed ramp as camouflage
use it only when the action reads correctly at final speed.
Checking only the first frame
inspect start, midpoint, end, and suspicious timecodes.

// pakistan_angle

Pakistan Angle

For a creator paying in PKR, repeated USD-priced generations compound exchange-rate and card charges. Check the current credit cost and your bank’s converted amount, set a generation ceiling, and pay for the tool that removes your actual bottleneck. With tight quota, crop, trim, or use B-roll instead of rerolling a usable clip.

With unstable electricity or mobile data, generate short drafts and download only accepted candidates at full quality. Save prompts and logs locally through load-shedding. For WhatsApp review, send a compressed, watermarked copy with timecodes; keep the master for final transfer. Never prompt with a real phone number, CNIC, address, or Daraz customer label.

// hands_on

Hands-On Exercise

6 steps

Select three authorized generations. Name them C01-v1, C02-v1, and C03-v1.
Review each at normal speed, frame-by-frame, and in the phone crop. Log a precise timecode per defect.
Classify every defect as identity, motion, anatomy, text, physics, flicker, or texture; assign kill, repair, or pass.
Apply three different decisions across the set: one targeted reroll, one isolated/cropped repair, and one replacement or motivated B-roll cutaway.
Save the revised prompt and output as v2. In Flow, use History where available; otherwise store the prompt beside the clip.
Review the three-clip sequence at normal speed on a phone and mark each log row accepted or rejected.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

4 official sources — check every claim yourself

// check_yourself

Check yourself

4 questions · answers and options are taken word-for-word from this course

0/4

1 / 4 · diagnose
Your work shows this failure mode: “Rerolling the unchanged prompt.” The lesson describes it like this: “The same complexity remains.” What does the lesson tell you to do about it?