Video-to-SFX Models Ranked (2026): Four Models on Eight Real Clips

A door can look shut and still sound wrong. We gave four video-to-SFX routes the same eight silent AI-drawn clips, then hid the model names while Syver-August rated the results. Sonilo v1.1 leads this small test at 72.50%. Every rated video plays below with its generated sound, not just a score or a waveform. This is one listener’s test, not a universal verdict.

By Syver-August, Developer, North Noir · October 27, 2026 · 8 min read

What we measured, and what the input actually was

The input was a silent AI-drawn video; the output was a generated sound track underneath that picture. Four routes ran on all eight cues: Sonilo v1.1 and Mirelo SFX on fal, HunyuanVideo-Foley on Replicate, and MMAudio V2 on fal. The videos were generated on 26 October 2026. The owner’s completed rating CSVs were supplied on 27 October; they do not timestamp individual listening sessions.

The cue names match our text-to-SFX benchmark: small harbor bell, wooden door, rain on tin, gravel footsteps, paper, sword, floorboard and distant thunder. The video inputs and rated muxes are ten seconds long, not four. Four seconds was the text benchmark’s target. We have retained the real ten-second files rather than cutting the evidence to fit that earlier target.

The first blind pack contained the door, gravel and bell; the second contained the other five cues. Each pack had its own sealed identity key and shuffled route order within a cue. Reusing an anonymous ID across packs does not mean reusing a model. The CSVs were de-blinded with their own keys after the owner had finished rating. Every rating and note is included above, including critical comments. Notes were lightly edited for spelling and grammar; meaning and scores are unchanged.

Each route/cue had one request. Mirelo returned a required pair of samples; sample 0 was fixed before generation, not selected because it sounded better. Generated audio was encoded to AAC and tail-padded or tail-trimmed only to fit the ten-second picture. There was no gain change, loudness matching, looping, time stretch or quality reroll. The gallery contains the exact MP4s the owner rated. The first pack’s picture was re-encoded during muxing; the second pack copied its video stream.

The door, the gravel, and the sounds nobody asked for

The door gives a useful starting point. Sonilo received 8/10, with the note “The sound is perfect.” Mirelo received 7/10, “The sound is solid and correct.” Hunyuan received 6/10; the listener heard an early closing sound before the real one. MMAudio received 2/10, with a note about extra sounds before and after the close. Those are observations about these four outputs, not promises about the next door each route will generate.

On gravel, Sonilo’s 8/10 came with “Contains only the correct gravel footstep sounds.” Hunyuan’s 7/10 identified correct gravel and timing but unwanted background music. Mirelo’s 6/10 note described the man’s grunting rather than gravel; MMAudio received 1/10, “The sound is wrong.” The difference here is not just a plausible sound texture. It is whether the track attends to the event the listener expected from the picture.

The remaining cues resist a neat winner-takes-everything story. Sonilo’s paper received 8/10, “The sound is perfect.” Mirelo’s rain scored 7/10 against Sonilo’s 6/10. For the floorboard, Sonilo’s 7/10 note explicitly says the video could have been clearer and that the sound was not creaking. On distant thunder, both Sonilo and Mirelo received 6/10 and the identical note “Sounds more like wind than thunder.” Input clarity and the event itself still matter.

Watch every rated clip with its sound

Choose a cue, then play the four routes in rank order. Native levels vary, so start quietly. These are the finished blind-test muxes with model names revealed afterwards, not a fresh demonstration generation. The silent source picture is the same for every model within a cue. The published score is the owner’s original rating, not an automated analysis of the file.

Cost per selected clip, not cost per second of finished film

The captured meter-and-tariff calculation is $0.09 per Sonilo clip, $0.21 per Mirelo selected clip, $0.014140273172025 on average for Hunyuan, and $0.01 for MMAudio. Mirelo’s figure includes both required samples: it would be misleading to halve the request cost when only sample 0 was rated. Hunyuan’s figure uses actual prediction time at the recorded L40S tariff, not ten seconds of GPU work simply because the video lasts ten seconds.

Cost/clip is the request cost to obtain the selected rated output, derived from returned provider meters and the captured 26 October 2026 tariffs, not invoice-audited charges. Mirelo includes its required two-sample batch, not half the request cost. Hunyuan uses actual predict_time, not ten seconds of GPU time. Input-video creation, storage and delivery are excluded; prices can change.

Verdict: start with Sonilo, then listen to your own cue

For this listener and these eight inputs, Sonilo is the first route to audition: the highest mean, strong door, gravel and paper results, and a lower request cost than Mirelo. That is the whole claim. Eight cues and one listener do not establish a universal ranking, repeatability, listener consensus or a statistically proven gap. The ±SEM describes variation across cues, not agreement between people. No model that was not run receives a score here.

In an AI-drawn video, sound remains a separate craft decision from the drawings. A benchmark can help you choose what to audition, not tell you to keep an unwanted voice or music track. For the text-only results, read Text-to-SFX Models Ranked. For all 16 route-task rows and the two carefully labelled family summaries, read Sound-Effects Models Ranked: Text and Video Together.

Key takeaways

• 32 owner ratings, four routes and eight ten-second AI-drawn input videos. The real rated MP4s, posters and every listening note are published; notes were lightly edited for spelling and grammar with meaning unchanged.
• Sonilo 72.50% ±3.13 pp; Mirelo 61.25% ±3.98 pp; HunyuanVideo-Foley 43.75% ±6.25 pp; MMAudio V2 30.00% ±5.00 pp. Each route has n=8.
• Quality % is mean rating × 10. One listener, not a jury; SEM is cue variation, not proof that a quality gap generalises.
• Costs are meter/tariff-derived, not invoice-audited, and Mirelo includes its required two-sample batch. Price and observed time do not change rank.

Frequently asked questions

Which video-to-SFX model ranked first?

Sonilo v1.1 led this four-route, eight-cue test at 72.50% with ±3.13 percentage points SEM. This is one listener’s blind rating of one selected output per cue, not a universal best-model claim.

Can I watch and hear all the benchmark outputs?

Yes. All 32 exact rated ten-second MP4s play with generated audio, grouped by cue and ordered by route quality rank. Each shows a poster, model name revealed after rating, score and listening note.

Were the videos four seconds long?

No. All eight input videos and 32 rated muxes are ten seconds long. The related text-to-SFX test targeted four seconds; those durations and different inputs must not be treated as interchangeable.

Is the price an audited provider charge?

No. Costs use returned provider meters and captured tariffs from 26 October 2026. They exclude source-video generation and storage. Mirelo’s selected clip cost includes the required two-sample request.

Explore AI-drawn video

Keep reading

• Sound-Effects Models Ranked (2026): Text-to-SFX and Video-to-SFX Together
• Text-to-SFX Models Ranked (2026): Blind Listening and Prices
• Background Music Models Ranked (2026): Vocals, Instrumental and Prices