Comparing three AI video models in the same 95 second scene

written with Chatgpt + EAW review · October 9, 2026

A beautiful five-second demo does not tell me whether an AI model can carry a scene. I want to know whether the characters keep their faces, listen when they should be listening, and finish a conversation without the editor having to rescue every shot.

For this experiment, we recreated the same one minute and 35 seconds from the Chapter Five audiobook drama of Bridesmaids & Bourbon, by Sydney Marsh. Taylor, stranded and exhausted, weighs Beau’s offer of a room. It is a quiet scene, which makes small changes in expression, mouth movement and timing especially noticeable.

The new versions use Hailuo 3 Max, MiniMax H3 and FLUX.3 Video. The earlier Kling 3.0 Pro and HeyGen edit is included as a reference. It combines two tools, so it is a comparison with our existing workflow rather than a pure Kling result.

The full videos are published as Cinematic Read Alongs, and the vertical dramas are also published on YouTube at Drama is Mine (@dramaismine). You can read the Bridesmaids & Bourbon ebook for free at FutureFictionPress.com.

Second pass: correcting the edit before judging the models

EAW asked us to revisit the voice-to-picture timing, including Kling/HeyGen. The first edits remain in the player: switch between First edit and Alignment pass without losing your place.

Astra does the editing, with EAW directing and reviewing the results. Getting these cuts into shape takes about two to three editing passes. The work here breaks down into an initial assembly, a focused dialogue-alignment pass, and targeted cleanup of trims and framing. EAW spots what feels wrong in the scene; Astra inspects the source footage, adjusts the cuts, renders both formats and checks the finished files. These are editorial passes, which can include several small fixes within one saved version. They do not each represent another paid generation of the whole scene.

That editing effort belongs in the comparison. A low provider bill does not mean a finished scene arrived ready to publish, and an export number is not a reliable count of how many times the edit was revised.

The audit found an editing error. The three new model cuts had inherited source offsets from the HeyGen edit, even though their native performances began at different times. Some short answers occurred two or three seconds later in the generated take. The first edits sometimes cut away before that mouth movement even happened. That mismatch should not be counted solely against the video model.

We transcribed all 15 native dialogue takes, checked the results against the original audiobook word timings and sampled mouth frames, then selected the correct source ranges. Picture speed changes follow phrase boundaries in Beau’s longer explanation. A slightly longer Taylor reaction covers part of his reassurance. The closing “Okay” and “Good” now use their actual speaking moments. H3 also needed a portrait-framing repair: Beau moves across the source during his longer speech, so the phone crop now follows his face instead of losing him offscreen. The original voices, narration, rain, music, captions and 1:35 runtime stay fixed; only the picture edit changes.

Kling/HeyGen is a different case: its generated audio already matches the supplied recording. It received small two- or three-frame picture delays based on sampled articulation, a tighter entrance to “Okay,” and the revised reaction cut. The earlier listening-shot repairs and music pass did not establish three dedicated dialogue-alignment passes. This is its first documented focused alignment revision in the comparison.

This is an editorial correction, not a new lip-sync model. Fine consonants and mouth shapes can still disagree with the audiobook, and picture retiming can affect how natural a gesture feels. Transcription timings and automated video critique are aids, not proof of perfect synchronization. Compare the exchange at 0:46, the two-clause reply at 0:52, the explanation at 0:59, and both answers at 1:33.

No new video generation was needed for this pass. The added provider charges below cover speech-timing analysis and full audiovisual review of both formats; local editing time and compute remain unpriced.

The finished package contains eight revised videos: four workflows in widescreen and phone formats. Every file runs exactly 95 seconds / 2,280 frames at 24 fps, decodes without errors and has a decoded soundtrack identical to the scored reference. The original cuts are preserved. The new analysis and review charges total $0.18471067—about 18.5 cents; that figure measures provider charges, not Astra’s editing effort or the full production cost.

Automated review was useful but unreliable. It correctly flagged H3’s cropped-out speaker, which direct frame checks confirmed and the moving portrait crop repaired. It also claimed that mouths never moved in places where dense frame strips clearly showed changing mouth shapes. We kept the useful findings, checked the others against the exported footage, and left fine phonetic synchronization as an open limitation. A technical check or an AI critique does not replace EAW’s creative viewing.

Workflow	First pass / historical	Added alignment pass	Cumulative provider cost	Cumulative per finished minute
Hailuo 3 Max	$10.2365	$0.0484	$10.2849	$6.50
MiniMax H3	$15.5038	$0.0409	$15.5447	$9.82
FLUX.3 Video	$24.1866	$0.0484	$24.2351	$15.31
Kling + HeyGen	$14.3405	$0.0470	$14.3875	$9.09

H3’s earlier unconfirmed charge still has $2.12 reserved, separate from the confirmed amounts above. The per-model $39.58 ceiling is cumulative across both edits; a revision does not reset it. Timing method and anchors and updated cost record are available for inspection. EAW’s creative review and final selection remain open.

What stays the same

The first comparison held the 95-second timeline and all 21 picture-cut boundaries fixed. The alignment pass keeps the same runtime, 21-shot structure, fictional character references and original scored recording, while allowing model-specific trims and small changes to reaction-cut boundaries. The narration, cast voices, rain and music are held fixed. “Room to Exhale,” supplied by EAW / elizabethannwest, enters at the same point in every version. Captions also follow the same words and timing.

We adapt the prompts to each model. We keep the native generated files, then make common 1280 × 720 review copies and separately framed 720 × 1280 phone copies. The returned source sizes differ: Hailuo Max supplied 1344 × 768, H3 supplied 2560 × 1440, and FLUX supplied 1280 × 704 in the initial samples. Resolution is therefore one of the differences being disclosed, not a controlled variable.

This is a single-scene production comparison, with selected takes and paid repairs counted. It is not a blind ranking or evidence that one model wins on every story. The earlier reference cut also had its own prompts, reference choices and lip-sync treatment.

Watch the listening as well as the talking

Hailuo 3 Max gives Taylor visible blinks, breathing and small listening changes in the inspected takes. Some breaths leave her lips parted despite the closed-lip instruction. Its decision shot contains a quick down-up nod. It finished this test without a paid video retake and has the lowest recorded cost of the three new scene versions.

MiniMax H3 keeps the sampled faces and wardrobe consistent, with downward glances and restrained reactions. Its native files have the largest frame dimensions in this test. One requested reaction failed at the provider, so the cut repeats an earlier successful H3 listening performance in those inserts. That is an editing compromise to look for, not an invisible success. The decision nod is small and brief.

FLUX.3 Video needed two paid replacements: the first listening take was too still, and the first decision take lacked the intended nod. Clearer motion directions produced a blink, downward glance and return in the listening replacement. The nod replacement includes the action later in the source, so the edit starts one second into that take. Both replacements, and the discarded first attempts, are included in its cost.

Kling + HeyGen remains our existing-workflow reference, now with its own second alignment edit. Kling supplied the visual coverage and HeyGen supplied audio-driven dialogue performances. Compare its mouth timing with the new native-video performances at 0:52 and 0:59, as well as the silent reactions at 0:03. It is a combined production method, so we show both vendors’ costs.

The six first-pass exports and all eight revised exports passed complete decode checks and run exactly 95 seconds. Their decoded soundtracks match the scored reference exactly. Automated review also made mistakes: it repeatedly mistook a lit lantern beside the dark fireplace for a fire, and missed brief nods visible in the source frames. Those critiques were checked against the footage before making editing decisions. These are production observations; EAW’s full creative review is still pending, and no overall winner has been selected.

The original audiobook voices are the soundtrack in all three new cuts. Native generated dialogue helps direct the visible performance, but those generated voices are replaced in the edit. Asking a model to finish a line at a particular second does not guarantee that its lips will match an existing recording. Exact mouth timing is something to judge in the player, alongside faces, hands and acting.

How we prompted each model

Hailuo 3 Max: Start with the established image, describe a concrete action, state what should stay consistent, and give sound its own direction. Put spoken lines in quotes and describe their delivery. Keep the requested action within the clip’s actual duration. These follow the principles in fal’s H3 Max guide. Our first probe used the structured H3 wording too; its natural breathing did not exactly follow the request for continuously closed lips.

MiniMax H3: Use its documented first-frame alignment and separate picture, soundscape and music instructions. The picture description states the person’s movement and eyeline; the dialogue includes the exact English line and a delivery cue. See the official H3 prompt guide.

FLUX.3 Video: Let the image establish the look, and use the text to direct what happens next. We describe a short chronological sequence of eye, head and shoulder movements. The first listening probe was too still, prompting a bounded retake with clearer movement beats. FLUX’s video documentation describes its image conditioning, keyframes and native audio.

For example, a listening prompt asks Taylor to blink, glance down, inhale gently, look back toward Beau and settle. The dialogue prompt asks Beau to say, “You’re not asking. I’m offering,” with calm, practical reassurance. These are performance directions tied to this scene, rather than a list of cinematic adjectives.

These runs use the controls available through our OpenRouter account. Vendor guides also describe endpoints and options that this route does not expose. The saved settings record what we actually sent; they are not a promise of an identical result on another service.

Download the exact prompts and settings used in this comparison.

See the FLUX listening take before and after clearer movement directions

These ten-second source takes use the same starting image. The first held too still. The replacement adds a timed blink, downward glance, breath and return to Beau. Each request cost $1.70; both are counted.

First take

Replacement selected for the cut

First-pass generation and historical costs

Original Kling + HeyGen cost: $14.34. That includes Kling 3.0 Pro: $12.992 across 16 video requests, HeyGen Avatar IV: $1.25 across five dialogue requests, and automated reviews: $0.098458, including the later scored-mix check. The exact historical total is $14.340458, or $9.06 per finished minute. Reusing its footage adds $0 in new generation charges; the alignment-pass review charge is shown separately above. Kling and HeyGen contributed to one mixed-workflow film; these are not separate 95-second model results.

FLUX’s recorded total includes $2.55 for replacement listening and decision-nod takes. Both first attempts remain in the production evidence. One H3 reaction request failed with no price returned; $2.12 remains reserved and is excluded from its confirmed total. Its insert reuses an earlier successful H3 reaction. The 127-second first-pass allowance per model includes five seconds of dedicated portrait coverage; these source takes are edited into the same 95-second scene. Automated review is included in the totals and itemized in the dated cost summary.

The new-model rows show incremental provider charges for this comparison: generated picture, discarded paid attempts, diagnostics and metered automated review. The Kling + HeyGen row shows its historical production cost separately. We count a widescreen and phone pair as one 95-second scene, not twice the finished runtime. We also show requested generation seconds, because handles, minimum clip lengths and discarded attempts mean buying more than 95 seconds of footage.

The figures exclude the earlier audiobook and character-image production, supplied Suno songs, subscriptions, editing time, software, electricity and storage. Reusing the earlier Kling and HeyGen footage creates no new generation charge; its historical production cost and the second-pass review cost are shown separately. A cheap successful request and a cheap finished scene are different measures.

EAW set a ceiling of $25 per finished minute for this test: $39.58 per 95-second model version, including its paid retries and review. That is a limit, not a target to spend.

Why Seedance has a diagnostic instead of a scene

On this account, Seedance 2.5 rejected the realistic fictional character image with a possible-real-person input error before returning a video. The rejected portrait request recorded no charge. A separate empty-armchair request succeeded, showing that the route could generate a video without that character reference.

Five-second empty-room diagnostic. Recorded provider cost: $1.16523.

The empty-room test cannot establish Seedance’s ability to perform this two-character scene. It is evidence about what was available through our provider route on the test date, not a general claim that Seedance cannot animate people.

What this tells me about the Mac arriving in February 2027

The forthcoming Mac is intended to move more experiments onto our own machine. Its 256GB of unified memory offers substantial capacity for model weights, and the 2TB internal drive will need to accommodate both models and production media. Capacity, generation speed and software compatibility are separate questions; this hosted test does not measure that Mac’s speed.

H3 has published base-model weights. Its hosted 2K service also uses stages that are not part of the released base pipeline, so a downloaded version should not be assumed to reproduce this hosted output exactly. Hailuo Max is a post-trained variant; access to H3’s base weights does not establish that Max’s weights are available.

Black Forest Labs has announced FLUX 3 Dev as an open-weight multimodal backbone, and its video launch roadmap includes that release. EAW is planning around a possible release by February 2027 or soon afterward. We will check the actual download, license and Apple Silicon support when it lands. The current hosted FLUX video result is useful preparation, but it is not a benchmark of an unreleased local package.

The eventual local comparison should repeat this scene and record generation time, peak memory, output quality and repair effort. That will answer the practical question: how much of this production can we comfortably run ourselves?

Story: Sydney Marsh. Music: EAW / elizabethannwest. Human direction and editorial choices: EAW. AI-generated character visuals. Draft comparison cuts await EAW’s creative review.

Watch the full Cinematic Read Alongs and vertical dramas on YouTube at Drama is Mine (@dramaismine), and read the Bridesmaids & Bourbon ebook for free at FutureFictionPress.com.