Datasets:
Rank int64 1 10 | Model stringclasses 10
values | Provider / interface stringclasses 2
values | Test date stringdate 2026-09-01 00:00:00 2026-10-24 00:00:00 | Duration (seconds) int64 8 10 | Resolution stringclasses 4
values | Aspect ratio stringclasses 1
value | Native audio setting stringclasses 1
value | Recorded cost (USD) float64 1 5.6 | Listed or estimated cost (USD) float64 1.7 3.6 ⌀ | Generation time (seconds) float64 57.8 805 ⌀ | Strongest result stringclasses 10
values | Main failure stringclasses 10
values | Overall verdict stringclasses 10
values | Reroll policy stringclasses 1
value | Source URL stringclasses 1
value | Model ID stringclasses 10
values | Final score (out of 10) float64 3.5 8.4 | Final score before display rounding float64 3.48 8.37 | T1 score (out of 10) float64 5.3 9 | T2 score (out of 10) float64 1.9 8.3 | T3 score (out of 10) float64 3.3 7.9 | Generation time minimum (seconds) float64 76 303 ⌀ | Generation time maximum (seconds) float64 85 411 ⌀ | Generation time scope stringclasses 2
values | Cost note stringclasses 4
values | Delivered duration note stringclasses 3
values | Scoring version stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | Seedance 2.0 | WaveSpeed | 2026-10-24 | 10 | 720p | 16:9 | Enabled | 2.16 | 2.4 | null | Best weighted family and fight results; strong commercial timing. | Weapon continuity, a small fight-ending slip and weaker drink textures. | Main recommendation in this ten-model test. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | seedance-2-0 | 8.4 | 8.366667 | 9 | 8.3 | 7.8 | 303 | 315 | observed range across recorded runs; not an average | Paid after discount from $2.40. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
2 | Seedance 2.5 | WaveSpeed | 2026-09-01 | 10 | 720p | 16:9 | Enabled | 3.24 | 3.6 | 805.03 | Best picture quality and highest weighted advert score. | Oversized monitor, missing attacker, misspelled MANGOW label and a second can; $3.24 per clip. | Second; best picture, behind Seedance 2.0 on the weighted recommendation score. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | seedance-2-5 | 8.3 | 8.316667 | 8.9 | 8.2 | 7.9 | null | null | one recorded observation; not a three-run average | Paid after discount from $3.60. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
3 | Wan 3.0 | WaveSpeed | 2026-10-24 | 10 | 720p | 16:9 | Enabled | 1 | null | null | Strong family scene and fight, realistic advert people, $1 per clip. | Identity drift in the fight; early can opening and container swap in the advert. | Third; TGK editorial best-value pick. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | wan-3-0 | 7.9 | 7.916667 | 8.7 | 7.6 | 7.6 | 158 | 160 | observed range across recorded runs; not an average | Recorded session charge, not a current provider quote. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
4 | MiniMax H3 | WaveSpeed | 2026-09-01 | 10 | 2K | 16:9 | Enabled | 1.4 | null | 380.69 | Convincing father, active gameplay and a usable advert. | Weak fight choreography and continuity; fixed 2K output affects sharpness comparisons. | Fourth; strong calm scenes and advert, poor fight. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | minimax-h3 | 6.6 | 6.616667 | 8.8 | 3.4 | 7.8 | null | null | one recorded observation; not a three-run average | Recorded session charge, not a current provider quote. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
5 | FLUX 3 Video | Black Forest Labs | 2026-09-01 | 10 | HD | 16:9 | Enabled | 1.87 | 1.7 | 112.02 | Coherent family scene, readable MANGO COLA text and strong voiceover. | Fight resembles a gun handover; advert faces look generated. | Fifth; solid family scene and advert, weak action. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | flux-3-video | 5.8 | 5.8 | 7.5 | 2.7 | 7.3 | null | null | one recorded observation; not a three-run average | $1.87 charged; $1.70 was the estimate. | See clip-level measured duration; family clip approximately 15.04 seconds. | tgk-video-v2-2026-10 |
6 | Gemini Omni 1.1 Flash | WaveSpeed | 2026-09-01 | 10 | 720p | 16:9 | Enabled | 1 | null | 57.77 | Fastest recorded generation and energetic advert pacing at $1. | Wrong family-scene eyeline, merging people and moving elevator, scrambled branding. | Sixth; useful for inexpensive, fast drafts. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | gemini-omni-1-1-flash | 5.7 | 5.666667 | 7.8 | 3 | 6.3 | null | null | one recorded observation; not a three-run average | Recorded session charge, not a current provider quote. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
7 | HappyHorse 1.1 | WaveSpeed | 2026-09-01 | 10 | 720p | 16:9 | Enabled | 1.4 | null | 123.36 | Fourth-best weighted fight result; convincing running, impacts and sound. | Closed eyes in the family scene, stiff people, incorrect elevator ending and CONGA branding. | Seventh; the fight outperformed its other scenes. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | happyhorse-1-1 | 5.4 | 5.383333 | 6 | 5.8 | 4.4 | null | null | one recorded observation; not a three-run average | Recorded session charge, not a current provider quote. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
8 | Veo 3.1 | WaveSpeed | 2026-10-24 | 8 | 720p | 16:9 | Enabled | 3.2 | null | null | Realistic people, settings and polished product imagery. | Inactive game, broken fight logic and inconsistent advert editing; $3.20 for eight seconds. | Eighth; attractive frames without dependable scene logic. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | veo-3-1 | 4.5 | 4.516667 | 5.9 | 3.4 | 4.3 | 79 | 85 | observed range across recorded runs; not an average | Recorded session charge, not a current provider quote. | Approximately 8 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
9 | Sora 2 Pro | WaveSpeed | 2026-10-24 | 8 | 1080p | 16:9 | Enabled | 5.6 | null | null | Attractive people, interiors and individual shots. | Robotic dialogue, broken fights and advert continuity; $5.60 for eight seconds. | Ninth; worst editorial value in this tested group. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | sora-2-pro | 4.4 | 4.4 | 6.7 | 3.3 | 3.3 | 220 | 411 | observed range across recorded runs; not an average | Recorded session charge, not a current provider quote. | Approximately 8 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
10 | Grok Imagine Video 1.5 | WaveSpeed | 2026-10-24 | 10 | 720p | 16:9 | Enabled | 1.4 | null | null | Active gameplay and some recognisable advert beats. | Cartoon-like faces, merging action, weak advert and the only recorded main-test prompt refusal. | Tenth in this tested group. | First completed result kept; no quality rerolls | https://theseguysknow.io/best-ai-video-generators/ | grok-imagine-video-1-5 | 3.5 | 3.483333 | 5.3 | 1.9 | 3.3 | 76 | 113 | observed range across recorded runs; not an average | Recorded session charge, not a current provider quote. | Approximately 10 seconds; exact measured values in evidence_records.csv. | tgk-video-v2-2026-10 |
Ten AI Video Generators Tested on Dialogue, Action and an Advert
This dataset accompanies TGK's ten-model AI video comparison. It records 30 first completed outputs across three tasks: a father-and-son gaming conversation, a hotel fight and a Mango Cola advert. Five models were tested on 1 September 2026 and five more on 24 October 2026. There were no quality rerolls.
Version: v2, published 25 October 2026. Version-specific DOI: 10.5281/zenodo.23637815. All versions; original v1.
Results
| Rank | Model | Final /10 | Father & son | Hotel fight | Mango Cola |
|---|---|---|---|---|---|
| 1 | Seedance 2.0 | 8.4 | 9.0 | 8.3 | 7.8 |
| 2 | Seedance 2.5 | 8.3 | 8.9 | 8.2 | 7.9 |
| 3 | Wan 3.0 | 7.9 | 8.7 | 7.6 | 7.6 |
| 4 | MiniMax H3 | 6.6 | 8.8 | 3.4 | 7.8 |
| 5 | FLUX 3 Video | 5.8 | 7.5 | 2.7 | 7.3 |
| 6 | Gemini Omni 1.1 Flash | 5.7 | 7.8 | 3.0 | 6.3 |
| 7 | HappyHorse 1.1 | 5.4 | 6.0 | 5.8 | 4.4 |
| 8 | Veo 3.1 | 4.5 | 5.9 | 3.4 | 4.3 |
| 9 | Sora 2 Pro | 4.4 | 6.7 | 3.3 | 3.3 |
| 10 | Grok Imagine Video 1.5 | 3.5 | 5.3 | 1.9 | 3.3 |
Seedance 2.0 leads the weighted recommendation score. Seedance 2.5 is TGK's picture-quality pick and the advert winner; Wan 3.0 is the editorial value pick. These judgments describe the recorded clips and prices, not every output those models can produce.
Files
| File | Contents |
|---|---|
| model_results.csv | Ten model summaries, final and task scores, providers, dates, recorded costs and scoped generation-time observations. |
| evidence_records.csv | Thirty main results, with category scores, unrounded and displayed clip scores, observations, measured technical properties and hosted video links. |
| test_prompts.csv | Eight prompt/provenance records: the original three, the separate structured retest, three October standard-task records and the partially recorded Grok advert revision. |
| action_prompt_retest.csv | Four historical retest rows from 3 September: three completed outputs and Gemini Omni's three rejected attempts. |
| SCORING.md | Criteria, weights, formula, rounding and value interpretation. |
| DATA_DICTIONARY.md | Field definitions, identifiers, units and missing-value conventions. |
| CHANGELOG.md | Scope changes and corrections from v1. |
| CITATION.bib | Current v2 and historical v1 dataset citations. |
| LICENSE | TGK-authored data and text reuse terms, with video and third-party exclusions. |
The package contains metadata and annotations, not MP4 files. Hosted video links are provided in the evidence table and article; those external files are not preserved inside this archive.
Method and settings
Each model started from text without a source image or supplied audio track. FLUX ran through Black Forest Labs' interface; the other nine ran through WaveSpeed. Native generated audio was retained. Standard task text requested ten-second, 16:9 scenes, but the endpoints did not deliver identical settings.
- Most clips are approximately ten seconds at 1280×720. MiniMax returned 2560×1440. Sora was set to 1080p and delivered 1888×1088. Sora and Veo clips last eight seconds. FLUX delivered 1280×704, including an approximately 15.04-second family clip.
- Costs are recorded session charges, not current price quotations. Seedance 2.0 cost $2.16 after a $2.40 list price; Seedance 2.5 cost $3.24 after $3.60. FLUX was charged at $1.87 rather than its $1.70 estimate. Sora's $5.60 is an eight-second charge.
- A single first-round timing observation per model is kept at model level. October timing ranges are retained as ranges; neither is a three-run average or a separate measured time for every evidence row.
- Grok's advert prompt was refused before a revised setting produced the retained clip. The revised setting is recorded, but a complete verbatim export of that revised submission is unavailable.
Every main clip receives four editorial ratings out of five: accuracy 25%, scene logic 25%, audiovisual quality 30% and value 20%. The weighted sum is doubled to give a clip score out of ten. The three unrounded clip scores are averaged equally and the final result is displayed to one decimal using round-half-up. See SCORING.md.
Prompt provenance and retained retest
The September main prompts were reconstructed from TGK's testing records, rather than recovered from generation-interface histories. The legacy column “Exact published prompt” means the exact text published in the guide; it does not make a reconstructed September prompt an original request export. TGK reports the October standard prompts as exact, and their published wording is included with that attribution. The Grok revision has a partial record and a blank full-text field instead of an invented reconstruction.
The separate 3 September action-prompt experiment used named characters, fixed positions, timestamps and a locked camera. Its three completed results and one three-attempt rejection row are preserved outside the main 30 results. Its Seedance 2.5 win remains a result of that separate experiment.
Limitations
This is a small editorial comparison with one completed result per model and task, without repeated runs or statistical uncertainty estimates. Scores reflect TGK's audiovisual review and judgment. The archive preparation verified score arithmetic, file properties and media mappings; it did not supply a new independent audiovisual score. Scene logic describes visible continuity, not internal model reasoning. Different providers, versions, output settings and prompt exceptions limit controlled comparisons. Small score gaps are not proof of a general model advantage.
Blank CSV fields mean unknown, unavailable or not applicable as defined in DATA_DICTIONARY.md; they are not zero. Video durations and dimensions come from the original file records, with the replacement Seedance 2.0 advert checked again during this release preparation. Native video ratios can differ slightly from the requested 16:9 aspect ratio.
Citation and reuse
These Guys Know. (2026). TGK AI Video Generators 2026: Ten-Model Text-to-Video Test Dataset (v2). Zenodo. https://doi.org/10.5281/zenodo.23637815
The v2 identifier is for this release; the v1 DOI identifies its unchanged historical files. Use the all-versions DOI to find the evolving archive. See CITATION.bib and LICENSE. TGK-authored metadata, prompts, annotations and methodology are licensed under CC BY 4.0; generated videos, logos, interfaces and other third-party material are excluded.
Canonical article: https://theseguysknow.io/best-ai-video-generators/
Working repository: https://huggingface.co/datasets/These-Guys-Know/tgk-ai-video-generators
- Downloads last month
- 110