Original + template
$0.00 generation
Original MP4KINGY / PRODUCTION BENCHMARKS · SEPTEMBER 27, 2026
Gen-4.5 vs Veo 3.1 Fast: 9 Product Ads Tested (Results + Cost)
Watch both AI workflows beside an original-image template. Compare the same brief, the finished files and generation costs.
The short answer
All nine 15-second ads passed the file checks. No clear winner yet: human reviewers haven't rated these ads. The original-image template keeps the source image still; both AI workflows add generated shots that need a close product review. Gen-4.5 and Veo 3.1 Fast each used $3.60 in generation credit value across three ads; the template used $0 in new generation. Watch the videos side by side and judge for yourself.
October 19 update: We rechecked all nine ads and found a headline/cap overlap, with product fidelity uncertain in two other edits. Inspect the sampled-frame evidence.
This small test does not tell us the full cost of an ad a human editor would accept.
RESULTS / SAME PRODUCT, SAME COPY
Results at a glance
| Workflow | File checks | Generation credit value across three ads | What to weigh |
|---|---|---|---|
| Original image + template | All three passed | $0 in new generation | Keeps the product tied to the source image. The image stays still; only the copy changes. |
| Gen-4.5 + template | All three passed | $3.60 | Adds generated shots. Review the product closely; file checks do not show whether an ad is usable. |
| Veo 3.1 Fast + template | All three passed | $3.60 | Adds generated shots. Each request pays for six seconds; the edit uses the first five. |
The twelve AI requests succeeded without retries or replacements based on quality. They used 720 existing credits, valued at US$7.20. We made no new cash purchase. These costs exclude human editing time.
WATCH THE SAME BRIEF
Watch the same brief side by side.
Three 15-second exports from the same source. The AI workflows combine two generated shots with template copy and an original-image end card. We added the copy and end card after generation.
Inspect source imageNew product introduction
A restrained product reveal, complete bottle always visible. Soft studio lighting. A subtle slow camera move without cropping the cap or base.
Approved copy: Meet the Field Bottle → KINGY • 500 mL → Explore the collection
All clips are silent. Moving the timeline pauses all three. Use the arrow keys for 0.1-second steps. Playback can vary slightly by browser.
Select a brief to compare. Without JavaScript, the Launch players and original download links below remain available.
Prices show generation credit value only. Editing time and the cost of an accepted ad are still unknown. Both AI models ran through Runway.
Separate checks found a cropped cap and two free downloads
In a separate five-second Gen-4.5 test, the cap leaves the frame at 4.5 seconds. The file passed its checks and used $0.60 in existing credit value. That single clip does not show how often the model misses a brief. Watch the original and read the test details.
| Tool | What we downloaded | What to watch for |
|---|---|---|
| PixVerse | A free 720 × 1280 MP4, 5.04 seconds | The watermark stays on the video. Later frames crop the bottle. Removing the watermark required a subscription. |
| Canva | A free 940 × 1672 MP4, five seconds | The export size does not prove the resolution of the generated video. We have not verified its native resolution. |
No clear watermark-free, native-resolution winner emerged. See every tested account, skipped route and limit.
NEW ASSESSMENT / OCTOBER 19, 2026
Nine valid exports still need a product and copy check.
All nine ads pass the file checks. One finished edit has a visible headline problem. In the Gen-4.5 launch ad, the cap overlaps the headline at 4.000 and 4.9583 seconds. Two other Gen-4.5 edits leave product fidelity uncertain: a turning label hides letters, and strong lighting changes the apparent finish and dot color.
This is a new assessment of the retained September 2026 files. It is not a new model run. Two separate AI review contexts inspected sampled stills; a coordinator checked the disputed evidence. Human acceptance and correction time remain unmeasured.
Machine checked
9/9 finals pass all nine export gates. All 3,240 final frames and 12 raw clips decode without reported errors.
AI assessed
135 sampled stills, two separate reviews, full-resolution checks where needed. No full-playback or all-frame visual assessment.
Human reviewed
0 acceptance decisions. The separate human-operated study remains 0/27. No cost per human-accepted ad.
Inspect the evidence
The images below are delivery previews and may be recompressed by the CDN. For exact original image bytes and checksum verification, use the assessment ZIP.
Gen-4.5 launch · 4.9583s. The phrase is present, but dark letter portions merge with the cap. This is a finished-edit layout issue; the overlay was added after generation.
Gen-4.5 retarget · 4.9583s. The ends of KINGY and FIELD BOTTLE turn out of view. That does not prove misspelling.
Gen-4.5 detail · 7.000s. Strong glare changes the apparent finish and dot color. The stills cannot establish a changed material.
Compare with the original product image and all three workflows in the player. These examples come from one fictional product and one final per brief/workflow; they do not establish a model failure rate.
Every final, including the uncertain results
| Workflow / brief | AI sampled outcome | What we observed | Inspect |
|---|---|---|---|
| Original image + template Launch | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
| Original image + template Detail | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
| Original image + template Retarget | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
| Gen-4.5 + template Launch | Observed copy overlap | At 4.000s and 4.9583s the cap overlaps the headline. Dark letter portions lose contrast against the dark cap. This is an edit/layout issue, not invented wording or a cropped bottle. | Original MP4 All 15 samples |
| Gen-4.5 + template Detail | Uncertain | Strong illumination at 7.000s changes the apparent cap/body finish and dot color. Exact color fidelity remains uncertain; no wrong word is demonstrated. | Original MP4 All 15 samples |
| Gen-4.5 + template Retarget | Uncertain | At 4.9583s the label turns around the bottle edge, hiding the ends of KINGY and FIELD BOTTLE. Hidden letters remain unverified, not proven corrupt. | Original MP4 All 15 samples |
| Veo 3.1 Fast + template Launch | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
| Veo 3.1 Fast + template Detail | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
| Veo 3.1 Fast + template Retarget | No violation seen in sample | No violation identified in the inspected samples. This does not certify every frame or human usability. | Original MP4 All 15 samples |
“No violation seen” is not a production-ready certification. Tiny label text was expanded only at the times listed in each review. A correct original-image end card does not prove the earlier generated shots preserved the product.
All five criteria and the reviewers' disagreements
Criteria: product framing; identity and geometry; label; added copy; consistency across examined stills. Pass means the cited inspected evidence supports that criterion. Uncertainty is retained where the evidence or interpretation does not resolve it.
Original image + template — Launch (A05)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Original image + template — Detail (A06)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Original image + template — Retarget (A09)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Gen-4.5 + template — Launch (A08)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: fail. Reviewer A: uncertain; reviewer B: fail. Coordinator inspected full-resolution frames 96 and119. Both reviews identified the same overlap. Dark headline letter portions merge with the dark cap, supporting an AI-assessed copy-legibility failure. The phrase is present; this does not establish wrong spelling or a model-only failure.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Gen-4.5 + template — Detail (A02)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: uncertain. Reviewer A: pass; reviewer B: uncertain. Coordinator inspected frames144 and168. Silhouette remains recognizable; lighting versus actual finish/color change cannot be resolved from these stills. Preserve uncertainty.
- Label: uncertain. Reviewer A: uncertain; reviewer B: uncertain. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: uncertain. Reviewer A: uncertain; reviewer B: fail. Large illumination and shadow changes are visible at6s versus7s. Reviewers differ on whether these breach the gentle-light/stable-shadow direction. No calibrated tolerance was frozen; interpretation remains uncertain.
Gen-4.5 + template — Retarget (A04)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: uncertain. Reviewer A: uncertain; reviewer B: uncertain. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: uncertain. Reviewer A: uncertain; reviewer B: fail. Warm rim lighting and a harder shadow at9.9583s are visible versus0s. Reviewers differ on the stable-shadow interpretation; retain uncertainty rather than infer material corruption.
Veo 3.1 Fast + template — Launch (A07)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Veo 3.1 Fast + template — Detail (A03)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Veo 3.1 Fast + template — Retarget (A01)
- Framing: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Identity / geometry: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Label: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Added copy: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
- Sampled consistency: pass. Reviewer A: pass; reviewer B: pass. Both reviewers agree within their stated sample coverage.
Both reviewers identified the cap/headline collision. One marked copy uncertain and one failed it; the coordinator inspected frames 96 and 119 and retained an AI-assessed legibility failure. Disagreement about the severity of lighting/shadow changes remains uncertain. No wrong spelling is asserted.
Review A, evidence and prompt · Review B, evidence and prompt · Final decisions and reasons
Cost, effort and what to do with the result
Before approving an ad, check the product and the copy together as the framing changes. Inspect generated shots before the static end card. The template retains the original image; motion adds review work that these file checks cannot measure.
This assessment used $0 in new external API consumption, made 0 paid calls and performed 0 repairs. The original twelve requests used $7.20 in historical generation credit value, with no retries or quality-based replacements. That is not an invoice or a current price quote. Machine checking and sample extraction took 10.7261 seconds; review, writing and publication are excluded. Human time, local operating cost, agent-service allocation and total production cost remain unmeasured.
Method, dates and limitations
The protocol was frozen before this assessment, after the original generation. The cohort report is dated September 27, 2026 and includes three retained September 24 templates. Evidence was rechecked October 19 Pacific / October 20 UTC. This separately dated assessment preserves the historical files, attempts, costs and original study ledger.
For every final, we inspected frames at 0, 1, 2, 3, 4, 4.9583, 5, 6, 7, 8, 9, 9.9583, 10, 12 and 14.9583 seconds: 135 of 3,240 final frames. All frames were decoded, but not all were visually assessed. Twelve samples cover the first ten seconds; three cover the original-image end card. All 61 recorded hash comparisons and ten checker-control expectations passed.
Two separate agent contexts received randomized aliases with workflow and cost labels hidden, reviewing in opposite orders. Visual content can reveal a workflow. Exact served model versions were not exposed. They viewed all contact sheets and expanded 32 and 28 frame images respectively from the same fixed sample. The coordinator inspected disputed frames. These are AI judgments, not independent human acceptance. There is no repeatability estimate or general workflow winner.
Frozen assessment protocol · Full machine results · Full written report
Recheck it locally
The 4.1 MB supplement contains all nine contact sheets, seven cited frames, both AI reviews and prompts, JSON/CSV results, checksums and an offline recheck adapter. It reproduces the same 135 sample frames from the existing original-video bundle.
Download assessment + recheck ↓
Reproduction instructions · ZIP checksum · Results CSV · Original videos, every attempt and recipe
What we did
We used one fictional bottle image and the same approved copy for three briefs: launch, detail and retargeting. We made six AI-assisted ads and reused three original-image template ads. Each AI ad combines the first five seconds of two generated shots with a five-second end card showing the original image. All nine finals use the same renderer.
Both AI models ran through the Runway API. We checked the files' format, duration and every video frame for decoding errors. We also recorded credit use. Human reviewers haven't rated the ads.
Separate checks covered eleven older AI clips, one paid Gen-4.5 clip and free-export routes. Those use different tasks or conditions. Their results stay separate below.
Try it yourself
Download the videos, compare them with the source image, or run the editing recipe locally.
Rebuild the nine-ad comparison
Get all nine finals, twelve raw clips, the source image, exact requests and the local editing recipe.
Download videos + recipe ↓Try the image + template workflow
Get the source image, briefs, renderer and three videos. Runs locally with Python, Pillow and FFmpeg. No provider account required.
Download template workflow ↓Check your own export
The local checker checks duration, portrait dimensions and full decoding. It also records the file hash. You need Python 3 and FFmpeg/ffprobe.
Download export test kit ↓python3 verify_export.py your-video.mp4 > report.json
A passing file still needs checks for free access, native resolution, watermarks, generated motion, product accuracy and permitted use. Read the checker details.
Want a browser metadata check? Try Output Preflight. To plan editing costs, use Video Production Budget or the planning calculator below. Estimates stay separate from measured results.
All other originals, reports, results and checksums are in the test records below.
How we tested and what this can't tell you
A larger study with human editors is planned; those sessions haven't run yet. The original count remains 0 of 27 sessions complete, with 0 human acceptance decisions. We cannot name a workflow winner or give a cost per accepted ad.
Open a record for its full methods, limits, cost notes and downloads. The earlier reports keep their original wording.
Original opening and finished-edit scope
KINGY / PRODUCTION BENCHMARKS
Compare the completed job: quality, retries, generation spending and the time someone spends fixing it.
Technical checks published.
Human acceptance remains unmeasured.
Inspect nine matched-task exports and their generation costs, or download the local export test kit. This release includes no independent human acceptance scores. Read the final free-export findings and limits.
Three 15-second exports from the same source. The AI workflows combine two generated shots with template copy and an original-image end card. These are finished edits, not raw model outputs.
Matched-video player: exact product constraints, prompts, settings and hashes
Exact product constraints
One upright matte dark forest-green KINGY FIELD BOTTLE, 500 mL, ochre dot above KINGY, flat black cap, rounded shoulders, silver base band. Preserve every letter, color, proportion and the entire bottle. Cream background and stable contact shadow. No additional objects, people, claims, watermarks or invented text.
Source SHA-256: e2b72719dff6354d808dd4032de56c07a731006aeae704f8466ed7c440c7f6aa
Technical checks passed; human quality unmeasured. Prices show observed generation credit value for these outputs, not total production cost or cost per accepted ad. The retained original-image baseline used $0 in new generation. Both AI models ran through Runway.
Exact prompts, settings and file hashes
AI edits use seconds 0–5 of shot 1, seconds 0–5 of shot 2, then a five-second original-image end card. The baseline holds the source image while the same copy changes. No quality-based retries or replacements.
All nine mappings and exact prompts (JSON) · Raw clips, finals and reproduction recipe (ZIP)
Nine matched ads: full results, credit costs and reproduction steps
REPORT 004 / SEPTEMBER 27, 2026
Nine exports. One fixed editing template.
We completed six new AI-assisted product ads and reused three conventional baselines across the same three briefs. All nine final files pass the shared format, duration and full-decode checks. Independent human acceptance was not measured in this release.
Technical results are ready to inspect. These checks cannot establish product fidelity, readable copy or a usable ad. No independent scores were collected before this release closed. Prepared review packets remain private.
| Workflow | Launch | Detail | Retarget | New credit value |
|---|---|---|---|---|
| Original image + template | Inspect MP4 ↓ | Inspect MP4 ↓ | Inspect MP4 ↓ | $0 |
| Gen-4.5 + template | Inspect MP4 ↓ | Inspect MP4 ↓ | Inspect MP4 ↓ | $3.60 |
| Veo 3.1 Fast + template | Inspect MP4 ↓ | Inspect MP4 ↓ | Inspect MP4 ↓ | $3.60 |
Twelve new API requests succeeded, with no retries or quality-based replacements. They consumed 720 existing credits, valued at US$7.20; no new cash purchase was made. Each 60-credit balance decrease matched the planned request price. These observations are not invoices. Human time and cost per accepted ad remain unknown.
The same source image, approved copy and renderer were used throughout. Each ad uses the first five seconds of two raw clips, followed by a five-second original-packshot end card. The three conventional files are retained from Report 002. This separate agent-operated cohort has one final per brief/workflow; the original human-operated study remains 0/27. There is no provider ranking or reliability estimate.
Download all nine finals + twelve raw clips ↓
Full report and reproduction steps · Results JSON · ZIP checksum
Includes source image, exact requests, seeds, pre-run hashes, raw outputs, finals, cost evidence and the local editing recipe. Generated with Runway. Both models used the Runway API; this does not test free web plans or direct Google service reliability.
Earlier baseline and eleven AI clips: checks, sampled frames and limits
01 / MEASURED SEPTEMBER 24, 2026
Inspect the outputs and their limits.
What the evidence supports
A conventional template produced three valid exports without new API consumption. Eleven retained AI clips also passed their own technical checks, while sampled frames show problems that need creative review. No workflow winner or cost per accepted ad is established.
Cohort A · new local exportsThree briefs.
Three briefs.
A working conventional baseline.
Each file is a 15-second, silent portrait ad made from the supplied fictional bottle image and copy. All ten fixed technical gates passed, including a full decode and frame count. This template holds a still packshot and changes the text; it adds no generated motion.
| Brief / MP4 | Technical checks | Machine render | New API consumption |
|---|---|---|---|
| Launch ↓ | 10 / 10 passed | 0.987 seconds | $0 |
| Detail ↓ | 10 / 10 passed | 0.828 seconds | $0 |
| Retarget ↓ | 10 / 10 passed | 0.854 seconds | $0 |
720 × 1280 · H.264 · 24 fps · 360 frames · 15 seconds · no audio. Machine render time excludes authoring, review and human editing. Labor cost and human acceptance remain unknown.
Get the reusable workflow ↓Includes the source image, briefs, exact renderer, all three videos and a checksum verifier. Runs locally with Python, Pillow and FFmpeg; no provider account required.
New local study · fictional product · silent.
Technical pass; human review pending.
Text description
A stationary complete green bottle on a cream background. Seconds 0–5: “Meet the Field Bottle.” Seconds 5–10: “KINGY • 500 mL.” Seconds 10–15: “Explore the collection.” This is the unreviewed study output. Exact file bytes are preserved in the workflow ZIP.
Cohort B · historical raw files recheckedEleven AI clips decode.
Eleven AI clips decode.
That does not finish the job.
All eleven retained clips from Kingy's incomplete Creative Work Lab study passed their original duration, landscape aspect ratio, audio and decode checks; their hashes match the recorded originals. They still need the prescribed final copy, and the filter brief needs fixed narration. These approximately eight-second clips use different products and gates from the bottle baseline.
Historical request records total $8.54384 in provider consumption. No new paid generation was performed for this recheck. The provider account was not rechecked; these recorded observations are not provider-issued invoices. Taxes, credit commitments and human editing time remain unknown.
Inspect all eleven attempts and historical charges
| Attempt | Route / brief | Dimensions / duration | Historical USD | Result |
|---|---|---|---|---|
| attempt-001 | kling-v3-pro air-filter-explainer | 1920 × 1080 8.042 seconds | $0.896 | Technical pass Final edit incomplete |
| attempt-002 | kling-o3-image kettle-exact-offer | 1280 × 720 8.042 seconds | $0.672 | Technical pass Final edit incomplete |
| attempt-003 | kling-o3-image air-filter-explainer | 1280 × 720 8.042 seconds | $0.672 | Technical pass Final edit incomplete |
| attempt-004 | kling-o3-image invoice-three-shot | 1280 × 720 8.042 seconds | $0.672 | Technical pass Final edit incomplete |
| attempt-005 | kling-v3-pro invoice-three-shot | 1920 × 1080 8.042 seconds | $0.896 | Technical pass Final edit incomplete |
| attempt-006 | kling-v3-pro kettle-exact-offer | 1920 × 1080 8.042 seconds | $0.896 | Technical pass Final edit incomplete |
| attempt-007 | kling-o3-image air-filter-explainer | 1280 × 720 8.042 seconds | $0.672 | Technical pass Final edit incomplete |
| attempt-008 | veo-fast kettle-exact-offer | 1280 × 720 8 seconds | $0.79992 | Technical pass Final edit incomplete |
| attempt-009 | kling-v3-pro invoice-three-shot | 1920 × 1080 8.042 seconds | $0.896 | Technical pass Final edit incomplete |
| attempt-010 | kling-o3-image invoice-three-shot | 1280 × 720 8.042 seconds | $0.672 | Technical pass Final edit incomplete |
| attempt-011 | veo-fast invoice-three-shot | 1280 × 720 8 seconds | $0.79992 | Technical pass Final edit incomplete |
A prompt constraint can fail
Hands appear in the sampled frames of attempt 001, despite a no-people instruction. File checks cannot establish prompt compliance.
A story can stay unclear
Attempt 008's sampled indicator is on, off, then on. The intended shutoff action needs a full playback review.
Copy needs its own check
Several invoice samples contain distorted pseudo-text. Final overlays and editorial checks still matter.
These are agent observations of frames sampled at 1, 4 and 7 seconds, not independent human judgments or a full playback review. Samples were inspected for every retained attempt. No creative pass rate is calculated.
Keep the evidence boundaries visible.
The automated protocol was frozen before these local runs. The original human study is unchanged: 0 of 27 sessions complete. Human acceptance decisions: 0. Cost per accepted ad: unknown. The two groups above use different tasks and technical gates; do not combine their pass rates.
Earlier feasibility records · Report 001
The September 5 exploratory requests and $4.11 in recorded consumption remain separate. Their raw videos were not retained, and they are excluded from this study. The original pending human-cohort records and earlier demonstration are preserved.
One paid Gen-4.5 clip: framing miss, cost and test conditions
REPORT 003 / SEPTEMBER 27, 2026
A valid video file can still miss the product brief.
We generated one five-second Gen-4.5 clip with the instruction to keep the whole bottle visible. The downloaded file passes our duration, portrait dimensions and full-decode checks. At 4.5 seconds, the top of the cap is outside the frame.
That is a visible framing miss in this sample. It is an agent observation of the retained frame, not a human acceptance score or a model failure-rate estimate. Watch the original and compare it with the source image in the download.
Runway returned an estimate of 60 credits. The account balance fell by the same amount, equivalent to $0.60 at the checked credit rate. This used an existing balance; no new cash purchase was made. The balance observation is not an invoice or a terminal cost field. Human labor and cost per accepted ad remain unknown.
A planned silent Veo 3.1 Fast request was not submitted because the playground’s displayed pricing disagreed with the typed Fast-model settings. We stopped at the frozen cost limit. This is a configuration limitation, not a failed Veo generation.
Get the original + reproduction evidence ↓
Full report · ZIP checksum · Inspect the frame at 4.5 seconds
Separate agent-operated smoke test: one request, one retained raw output, no retries or editing. Gen-4.5 via Runway Dev playground, seed 111, 720 × 1280, H.264, 5.041667 seconds, no audio. No fixed model-weight version exposed. The fictional source, prompt, pre-run protocol hash, request IDs, file hashes and technical results are included. Original human study: still 0/27 complete. This paid request does not test a free tier.
Generated with Runway. One sample, no provider ranking.
Inspect the cap near the end.
Free exports: every account check, skipped route and limitation
FREE-EXPORT AUDIT / CLOSED SEPTEMBER 27, 2026
What we could actually download for free.
Final release: two free downloads, with limits.
This audit closed on September 27 with the evidence below. PixVerse produced a watermarked 720p file; Canva produced a free MP4 whose native generation resolution remains unverified. Further Kling, Luma and Adobe testing was excluded at close. No watermark-free native-resolution winner is established.
| Provider | Outcome | Evidence and limits |
|---|---|---|
| PixVerse | Free download obtained; partial result | 720 × 1280, 5.04 seconds. Watermark retained; later frames crop the bottle. Watermark-free option required subscription. |
| Canva | Free download obtained; native resolution unverified | 940 × 1672, five seconds. Canvas export dimensions do not establish native generation dimensions. |
| Pika | No export on tested account | Current Create Free account had zero credits; configured 720p request could not run. Legacy/promotional routes untested. |
| Hailuo | No export on tested account | Zero credits; configured request opened subscription selection. One-time trial eligibility unknown. |
| Kling | Excluded from final export comparison | Generation completed; original download was not verified before testing closed. Retained checkpoint is historical evidence. |
| Luma | Skipped at close | Phone verification unresolved in the tested session; no confirmed generation. |
| Adobe Firefly | Skipped at close | Account completion unresolved in the tested session; no generation. |
| Runway web | Not tested to export | Web login remained unresolved. Paid Runway API results are a separate test. |
Final report and budget · Final results JSON
Pika: the tested Free account had zero credits.
On Pika’s current Create site, we confirmed the Free plan, uploaded the fixed bottle image and set Pika 2.5 to 720p, five seconds and one generation. The account showed 0 credits, the request estimate was about 20 credits, and Generate was disabled after setup. We obtained no free export.
Pika’s official pricing page currently lists zero monthly credits for Free, with credit packs sold separately. This agrees with our account observation. We bought no credits, trial or subscription for this check.
One account and one current route, checked from Canada. Legacy and promotional accounts may differ; the linked old Pika site was not tested. No output means no verified watermark, native dimensions or quality result. This release does not cover every planned provider; see the final scope above.
Read the checkpoint and its limits · Exact prompt, settings and observations · Get the reusable export test kit
Hailuo: the configured request opened plan selection.
We uploaded the same source image and unchanged prompt to Hailuo 2.3 in Start/End Frame mode. The displayed settings were 768p, six seconds and one generation. The account had 0 credits against a 25-credit estimate. Pressing Create opened the subscription page; no queued generation or free export was confirmed. No purchase was made.
The plan page marked Free as Current and advertised one-time trial credits for Hailuo models alongside zero monthly credits. This account’s trial history and eligibility are unknown. The result does not establish that new or promotional accounts receive no allowance. The selected resolution is not a measured output; watermark and quality remain unverified.
Hailuo checkpoint and limits · Exact configuration and observations
Canva: a free MP4, with native resolution still unverified.
On a verified Canva Free account, we used Edit → Image to Video → Custom with the fixed image and unchanged prompt. One generation completed; we downloaded an MP4 without a purchase, payment details or paid trial.
The saved file is 940 × 1672, five seconds, silent H.264 and passes all four file checks. The editor’s playback resource reported 704 × 1216; it may be a preview copy. Neither that resource nor the larger canvas export proves the underlying generation dimensions, so our native 720 × 1280 requirement remains unverified. This is a free export, not a full audit pass.
Download the original and evidence · Method and limits · Observations JSON · ZIP checksum
PixVerse: free 720p download, with a watermark and a framing miss.
On a Basic account with 60 free credits, one PixVerse V6 request used the fixed source and unchanged prompt: 720P, five seconds, one output, audio and Multi-Shot off. It consumed 40 free credits. We downloaded the Watermarked version without payment details, a purchase or paid trial.
The original is 720 × 1280, 5.04 seconds, silent H.264 and passes all four file checks. The completed Details panel confirms 720P; no upscale or resize was used. The downloaded frames show PixVerse branding, and Watermark-free opens a subscription-required offer. This is a partial audit result. Later samples also crop the bottle despite the instruction to keep it fully visible.
Download the original and evidence · Method and limits · Observations JSON · ZIP checksum
Kling: generation completed; download access remains unresolved.
One VIDEO 3.0 request completed using 30 of 66 promotional credits, with the fixed image and prompt, 720p, five seconds, one output, Native Audio and Multi-Shot off. No purchase was made. The download action reached Kling’s download endpoint, where Chrome reported ERR_BLOCKED_BY_CLIENT. No original file was obtained.
This is a browser-access limit in our test, not evidence of a Kling export paywall or failure. Downloaded dimensions, watermark and file checks remain unverified. The completed generation is retained as historical evidence; further Kling testing was excluded at close.
Kling checkpoint and limits · Exact settings and observations
Export checker: requirements, open questions and validation
EXPORT CHECK / SEPTEMBER 27, 2026
Check the file before you count it as a free export.
A preview can look convincing while the downloaded file is too short, too small, or damaged. This kit checks a local video against our audit target: at least five seconds, portrait 9:16, and file dimensions of at least 720 × 1280. It also decodes every video frame and records the file’s SHA-256 hash.
A technical pass leaves six questions open.
Was generation and download free, without a card or paid trial? Is this the native output, without later upscaling? Is there a watermark? Does it contain generated motion? Does the product match the source? Do the applicable terms permit your intended use? Check those separately and retain the evidence.
Download the export test kit ↓ ZIP checksum
The kit contains a fictional product image and provenance, a fixed test prompt, a blank audit record, a Python checker, validation results and an offline example. The example is an existing conventional still-image video, so it demonstrates why a technical pass does not establish generated motion. Requires Python 3 and FFmpeg/ffprobe; the checker runs locally without uploading your file.
python3 verify_export.py your-video.mp4 > report.json
Checker validation: six tests passed, covering three retained baseline exports plus short, landscape, low-resolution, corrupt and missing-file cases. This is a completed verification tool, not a completed eight-provider free-export comparison. Kit validation uses retained files and adds no human acceptance decision or free-plan success claim; the paid smoke test above is separate. Inspect validation records.
For a quick browser metadata check, use Output Preflight. To estimate a production budget, use Video Production Budget. Keep estimated costs separate from observed bills and independently accepted outputs.
Workflow shortlist: price estimates and exact settings
02 / THE BUYING DECISION
One brief. Three complete workflows.
15 seconds · 9:16 · sound-off
Approved bottle packshot + approved copy
Current recommendation
Keep your current production workflow until a matched test clears your acceptance gates. We have no measured winner yet. Use this shortlist to scope a test for your own product.
01Practical baseline
Original asset
+ template edit
Use the supplied packshot, approved copy and conventional editing.
- Best question to test
- Does controlled simplicity meet the brief?
- Main tradeoff
- Product appearance stays tied to the source; movement is limited.
- Generation estimate
- $0 observed in the local baseline
- Accepted / tested
- Not measured
- Total finished cost
- Not measured
02Candidate
Gen-4.5
+ template edit
Generate two visual sections, then add the same copy and end card.
- Best question to test
- Is added motion worth the rework?
- Main tradeoff
- Product fidelity needs review on every attempt.
- Generation estimate
- $1.20–$2.40 per attempted ad*
- Accepted / tested
- Not measured
- Total finished cost
- Not measured
Exact route and settings
Runway Dev API · gen4.5 · version 2024-11-06 · 720:1280 · two five-second shots · at most one retry per shot. The provider's model family ID is not a frozen weight snapshot.
03Candidate
Veo 3.1 Fast
+ template edit
Generate on the same billing surface, then use the same final edit.
- Best question to test
- Which route preserves the product for this job?
- Main tradeoff
- Pay for six seconds; use the first five in each section.
- Generation estimate
- $1.20–$2.40 per attempted ad*
- Accepted / tested
- Not measured
- Total finished cost
- Not measured
Exact route and settings
Runway Dev API · veo3.1_fast · audio off · 720:1280 · two six-second requests · at most one retry per shot. This is not a direct Google API price comparison.
*Rate-card estimates, not observed costs or prices per accepted ad. Excludes editing, tax and software allocation. Rates rechecked September 24, 2026: Runway API pricing. This is a bounded shortlist, not a claim about the newest or best models.
Cost calculator: a planning example you can change
03 / PUT EDITING IN THE BUDGET
A planning scenario
you can change.
Enter totals for the whole batch, including rejected work. Nothing here is a measured Kingy result.
How the calculation works
(Generation consumption + allocated software/assets + all active operator minutes × hourly rate ÷ 60) ÷ accepted final ads. A purchased credit pack is not added again when its consumption is already counted. Research and benchmark review overhead are separate.
Planned human study: protocol, scoring, disclosures and version history
04 / ACCEPTANCE COMES FIRST
The planned human comparison.
Freeze the job
Three briefs, three sessions each, identical source assets and a 30-minute operator limit. Every workflow produces the same 15-second deliverable.
Keep the misses
Log every request and retry. Preserve raw files, receipts, exact prompts, settings and edit time. Uncertain requests are reconciled before more paid work.
Review without the price
Two independent people review anonymized exports against the brief. Product, copy and technical gates must pass. Disagreements get a third review.
Cost all the work
Divide the full batch cost, including failures, by accepted ads. Show generation spending separately from the assumed dollar value of human time.
Acceptance criteria and scoring
Exact product identity and approved copy; no blocking geometry, label, framing or claim error; readable copy at phone size. H.264, 720 × 1280, 24 fps, 15 ± 0.1 seconds, silent and decodable. Human fidelity, legibility, motion and brief-fit scores must each be at least 4/5. Both blind reviewers must accept, or a third adjudicates.
Sample limits and freshness
The first cohort uses one fictional bottle across introduction, detail and retargeting briefs. Nine sessions per workflow cannot establish a universal reliability ranking. It measures production acceptance, not ad performance. The protocol and source prices were checked September 23, 2026; new controlled trials have no test date yet. New model versions, materially different assets or changed workflow settings need a new cohort.
Independence and commercial disclosure
The proposed agency fee buys testing and a readout, never a favorable score. Providers cannot suppress failures. Kingy's wider media business accepts sponsorships; any supplied credits, affiliate relationships or sponsorship relevant to a test will be disclosed with that cohort. No provider funded this launch, and no paid pilot customer has validated the offer.
Kingy editorial and sponsorship standardsWhy begin with product ads?
It is a repeatable job with concrete brand-approval gates and visible editing costs. IAB's 2025 research reported that 86% of surveyed buyers used or planned to use GenAI for video creative. That is a category signal, not proof that agencies will pay for this product. Kingy's existing asset and pilot make it a practical first test.
We considered captioned shorts, support answers and coding-agent work. Product ads offered the best combination of existing Kingy material and a narrowly testable buyer decision. The conclusion remains a hypothesis.
IAB source · Commercial research and alternativesCorrections and version history
September 27, 2026 closeout: nine matched-task exports, two retained free-account downloads, dated route findings and a local export checker. Kling, Luma and Adobe further testing excluded at close. No independent human scores or provider ranking. Final report and budget.
v1.2 · September 24, 2026: separate preregistered automated study, three conventional exports, eleven retained raw-file rechecks, agent sample observations and two reusable downloads. No human acceptance claims.
v1.1 · September 23, 2026: agency pilot capped at US$500 total for one brief and nine sessions.
v1.0 · September 23, 2026: protocol, system demonstration and historical evidence audit published. No new paid benchmark results.
Corrections retain the old record and identify affected totals. The ledger regenerates JSON, CSV and report together. Price updates become dated scenarios; historical consumption stays fixed. Send a correction through the inquiry form and identify the record.
Agency pilot: scope, price, limits and inquiry
05 / TEST YOUR ACTUAL WORK
A private pilot
for your agency.
US$500 maximum total
For a production lead deciding how to deliver recurring ecommerce creative. Price and demand are hypotheses; scope comes before payment.
- One product brief, three workflows and three runs each: nine production sessions.
- Retained attempts, accepted exports where available, a cost and edit-time ledger, and your reusable workflow recipe.
- A private recommendation and 30-minute readout. One factual-correction round.
Target delivery: five business days after scope, assets, access and reviewers are confirmed. The $500 maximum includes up to $25 generation consumption and any applicable taxes or fees. No add-on charges for the agreed pilot. No guarantee of usable ads, sales lift or platform approval. No purchase or newsletter signup occurs here.
FAQ
Which workflow won?
No clear winner yet. All nine matched ads passed file checks, but independent reviewers haven't rated them. The template keeps the source image still. Both AI routes add generated shots that need a close product review.
How much did Gen-4.5 and Veo 3.1 Fast cost?
Each AI workflow used $3.60 in generation credit value across three ads. Together they used 720 existing Runway credits, valued at US$7.20, without a new cash purchase. Human editing time and total cost per accepted ad remain unknown.
Did any tool give us a free video download?
Yes. PixVerse gave us a watermarked 720 × 1280 file. Canva gave us a 940 × 1672 file, but its native generation resolution remains unverified. Neither result establishes a watermark-free, native-resolution winner.
Can I reproduce the edits without a provider account?
You can run the original-image template locally with Python, Pillow and FFmpeg. The nine-ad kit includes the generated clips and editing recipe. Making new AI clips requires provider access.
Does passing the file checks mean an ad is ready to use?
No. You still need to check the product, copy and fit for the brief. The export checker also leaves free access, native resolution, watermarks, generated motion and permitted use for you to check.
The Kingy Brief
Get The Kingy Brief.
AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Signup help and newsletter schedule
After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.