# Mirelo SFX 2.0: AI Sound That Finally Matches the Picture (Keep Dialogue, Drop the Music)

Mirelo SFX 2.0 rebuilds video soundtracks while preserving dialogue, removing unprompted music, and syncing effects across 10-minute edits. Full analysis of features, benchmarks, and workflows.

Canonical URL: https://aiidelist.com/blog/mirelo-sfx-2-0-review

Language: en

Published: 2026-10-15

Updated: 2026-10-15

## Key Takeaways

- **Mirelo SFX 2.0** is a video-to-sound model that analyzes finished footage and generates synchronized foley, effects, and ambience while preserving original dialogue and removing unprompted music.
- It processes full edits up to **10 minutes** in a single pass, maintaining acoustic consistency across cuts — a major improvement over per-clip generation.
- Blind listening tests show strong preference: 81% over Sony Woosh and 63% over Kling Foley for video-to-sound competitors, plus advantages over native audio from models like Veo 3.1 Fast.
- Dialogue preservation is **on by default**; effects automatically duck under speech. Optional prompts and negative prompts allow steering without regenerating the video.
- Trade-offs versus the prior version (sfx-1.6) include a single mixed output (no separate stems by default) and the removal of extend/inpaint operations, in exchange for dramatically longer speech-preserving runs.

## What Is Mirelo SFX 2.0?

Mirelo SFX 2.0, released by Mirelo AI GmbH on October 14, 2026, is a context-aware video-to-sound model designed for post-production of finished video. It takes existing footage — whether AI-generated, camera-captured, or a stitched edit — and rebuilds the non-dialogue soundtrack so that effects, foley, and ambience align with on-screen action, timing, and environment.

The core value proposition is straightforward: **keep the picture and the performance, replace everything else in the audio**. Spoken dialogue remains as performed. Music that video models often insert without being prompted is stripped. New sound is generated around the preserved speech, with automatic level adjustment so effects sit lower while people talk.

This approach differs from text-to-sound tools and traditional sound libraries. Those systems never see the footage, so every effect must be located, trimmed, and manually synced. Mirelo SFX 2.0 starts from the picture itself.

## Core Capabilities and How It Works

Analysis of the model’s behavior shows several interlocking mechanisms:

- **Video-driven generation**: The model reads action, timing, and setting. Footsteps land on footfalls, impacts match collisions, and room tone or outdoor ambience fits the visible space (a hall produces different acoustics than a forest).
- **Dialogue preservation with ducking**: Speech is kept intact by default. Surrounding effects and ambience are built around it and attenuated during dialogue. This can be disabled when only effects are desired (for example, before recording a separate voiceover).
- **Unprompted music removal**: Music that video generation models frequently add even when not requested is removed, leaving a clean bed for custom scoring.
- **Cross-cut consistency**: Running a stitched edit once produces continuous ambience and coherent object sounds across shots of the same location. Per-clip generation previously caused the same car or room to sound different from cut to cut.
- **Long-form single pass**: Studio and API support up to 10 minutes per run. Longer material is available through the Enterprise tier.
- **Steering**: An optional text prompt (up to 2,000 characters) can request specific sounds; a negative prompt (up to 1,000 characters) excludes unwanted elements. Up to four variations can be generated per run for selection.

Variants include the primary video-to-sound mode (highest quality, dialogue-aware), a text-to-sound mode for off-screen or non-synced elements, and a stem mode (announced as coming soon) intended for layering individual synced sounds.

Audio output is 48 kHz mono, available as WAV, FLAC, or MP3. The upgrade from 44.1 kHz in the prior generation aligns better with common video production standards.

## Key Improvements Over the Previous Version

Official API documentation and side-by-side comparisons highlight deliberate trade-offs:

| Aspect | sfx-1.6 | SFX 2.0 |
|--------|---------|---------|
| Speech-preserving length limit | 60 seconds | 10 minutes |
| Speech preservation default | Off (must enable) | On (must disable) |
| Output when preserving speech | Separate mix / speech / SFX stems | Single mixed track |
| Extend & inpaint | Supported | Not available |
| Prompt length limit | 5,000 characters | 2,000 characters |
| Negative prompt | Not accepted | Up to 1,000 characters |
| Sample rate | 44.1 kHz mono | 48 kHz mono |

The longer speech-preserving window and default-on behavior make 2.0 practical for finished short films, ads, and stitched AI sequences. The loss of separate stems and local editing tools means workflows that require granular track control must still rely on traditional DAWs or the previous model for those operations. Negative prompts are a new control surface that helps suppress residual music or unwanted ambience.

## Blind Listening Test Results and Performance Benchmarks

Mirelo published results from blind pairwise listening tests. Methodology involved 25 raters evaluating 30 videos each, with options to prefer one track or declare a tie. Preference reflects overall quality and match to on-screen events.

**Versus other video-to-sound models:**
- Sony Woosh: 81% preference for Mirelo, 9% for competitor, 10% tie
- Kling Foley: 63% Mirelo, 18% competitor, 19% tie
- Sonilo: 46% Mirelo, 33% competitor, 21% tie

**Versus native audio from video generation models (cinematic AI clips):**
- Veo 3.1 Fast: 57% Mirelo
- MiniMax H3 Max: 49% Mirelo
- Flux 3: 48% Mirelo
- Seedance 2.5: 47% Mirelo
- Kling native: 45% Mirelo

These figures indicate consistent preference for the rebuilt soundtrack on the tested material. Independent verification on a given project remains advisable, as results vary by clip content and listener criteria. Community feedback and early platform integrations (Runway, Fal) have focused on the practical benefit of cleaning unprompted music while retaining usable dialogue takes.

## Practical Workflows and Use Cases

**AI short films and stitched sequences**: Generate individual shots with any video model, lock the cut, then run the full edit once. Continuity of ambience and object sounds improves perceived production value without regenerating visuals.

**Ads and product videos**: Preserve voiceover or dialogue, strip residual music, and add timed effects. Custom licensed music can be layered afterward.

**Social and platform content**: Timed swishes, impacts, and environmental detail land on the correct frames without library search and manual alignment.

**Game trailers and cinematics**: Frame-accurate impacts, mechanical sounds, and environmental layers can be generated from the picture rather than assembled from stock libraries.

**Agent and platform integration**: Video agents can call the model at multiple stages — while generating variants (to judge pacing with sound), after native audio is present (to clean music while keeping dialogue), after stitching (for consistency), or post-delivery (to refresh sound without re-rendering picture). API and MCP access support this.

**Text-only generation**: For off-screen elements or pure ambience beds, the text-to-sound path produces clips up to 10 minutes without requiring video input.

## Availability, Pricing, and Access

Mirelo SFX 2.0 is available in the Mirelo Studio and Playground, via the official API and MCP, and on partner platforms including Runway and Fal. Plugins for Adobe Premiere Pro and DaVinci Resolve exist; a Roblox plugin has been noted as forthcoming.

Pricing is credit-based. During the introductory period through December 12, 2026, SFX 2.0 is billed at 10 credits per second per variant (approximately 33% off the standard 15-credit rate). Example monthly capacity estimates when using the discounted rate for SFX 2.0:

- Free (5,000 credits): ~8 minutes
- Creator (€20 / 24,000 credits): ~40 minutes
- Studio (€99 / 120,000 credits): ~200 minutes
- Higher tiers scale further

After the discount period the rate returns to 15 credits per second. Free-tier limitations include non-commercial use, project deletion after one week, and watermarks on exports. Paid plans allow opting out of training-data usage. Enterprise arrangements cover unlimited length and custom terms.

## Advanced Tips, Edge Cases, and Common Pitfalls

- **Prompt discipline**: Keep prompts specific about materials, space, and intensity. Overly broad requests can produce generic results. Negative prompts are useful for excluding music, wind, or voices that the model might otherwise introduce.
- **Variation selection**: Generate 2–4 takes on important sections. Small differences in timing or texture often appear; the best match is rarely the first.
- **Dialogue edge cases**: Very quiet or heavily processed speech may be handled less reliably. Test critical dialogue sections. Silent or near-silent source clips are processed normally (unlike some earlier behaviors that rejected empty audio tracks).
- **Length and stitching**: Stay within the 10-minute standard limit for predictable results. For longer pieces, process logical sections and verify ambience continuity at the joins, or inquire about Enterprise.
- **Stem and mixing expectations**: Expect a single mixed track. If separate dialogue, effects, and ambience stems are required for a client deliverable, plan for additional processing or retain the prior model for those needs.
- **Acoustic fit**: The model adapts to visible environment, but extreme mismatches (for example, a close-up that should feel intimate inside a large space) may need prompt guidance.
- **Rights and final mix**: Generated sound is intended for the supplied video. When adding third-party music, verify licensing separately. Always listen in context on the final timeline; platform playback can mask issues that appear in a full mix.

Common failure modes reported in early use include residual low-level ambience that still feels slightly synthetic on highly detailed foley, and occasional timing drift on very fast action. These are typically addressable by selecting a better variation or refining the prompt rather than regenerating the entire video.

## Conclusion

Mirelo SFX 2.0 addresses a persistent gap in AI video production: visuals that look finished but sound incomplete or inconsistent. By reading the picture to place effects, preserving performed dialogue by default, stripping unprompted music, and maintaining continuity across multi-minute edits, it enables a cleaner “picture first, sound second” workflow.

The documented preference margins in blind tests, the jump from 60-second to 10-minute speech-preserving runs, and the availability across Studio, API, and partner platforms make it a practical tool for creators and teams producing short-form AI content, ads, and stitched sequences. The trade-offs around stem output and certain editing operations are real and should be factored into pipelines that require granular audio control.

Creators evaluating the model can test it directly in the Mirelo Playground or Studio with their own footage. For production integration, the API documentation and credit pricing (noting the December 12, 2026 discount window) provide the necessary operational details. Listening to results on representative project material remains the most reliable validation step.