DaVinci Resolve 21 AI Features: I Tested Them on a Real Project [2026 Review]

DaVinci Resolve 21 AI Features: I Tested Them on a Real Project [2026 Review]

Last month I had to cut a 12-minute internal demo video for a product launch. Two camera angles, mediocre audio, tight deadline. I'd been wanting to test DaVinci Resolve 21's new AI features on something real — not a synthetic demo reel, not a 10-second clip of a golden retriever running across a field. An actual project with actual problems.

Here's what I found: some of these DaVinci Resolve 21 AI features are excellent. A few are rough. And one quietly changes the economics of video editing in a way that should make Adobe nervous.

What's Actually New in DaVinci Resolve 21?

Blackmagic Design announced DaVinci Resolve 21 at NAB 2024 with a long feature list, but the AI-powered additions are what got people talking. The DaVinci Neural Engine — Blackmagic's machine learning framework — got major upgrades across tracking, audio, and transcription.

The headline features I tested:

• IntelliTrack AI — neural engine-powered object tracking and stabilization
• Voice to Subtitle — automatic transcription and caption generation
• uTalk — AI-driven audio panning that matches speakers to their on-screen positions
• Color Slice — a new vector grading tool for subtractive color work (not AI, but too good to skip)

Here's the detail most coverage buries: many of these AI features ship in the free version. Not a trial. Not a crippled demo. The actual free DaVinci Resolve 21. GPU-accelerated tools like UltraNR noise reduction require the $295 Studio license, but IntelliTrack and Voice to Subtitle work without paying a cent.

That's not a pricing strategy. That's a siege.

IntelliTrack AI: The Feature That Actually Delivers

Let me start with the one that impressed me most. IntelliTrack AI replaces the older point-based tracker with neural engine-powered object recognition. You draw a rough selection around your subject, hit track, and the DaVinci Neural Engine figures out what you meant — even when the subject rotates, gets partially occluded, or changes scale.

I tested it on three scenarios from my project:

Speaker walking across frame while gesturing. IntelliTrack nailed it. The tracker locked onto the person and held through arm movements, a quarter-turn away from camera, and a brief moment where another person's shoulder crossed in front of the subject. I've burned hours manually keyframing corrections for exactly this kind of shot. IntelliTrack handled it in one pass.

Stabilizing handheld B-roll of a laptop screen. Also solid. The stabilization was noticeably smoother than the previous generation. It correctly identified the laptop as the anchor point rather than trying to stabilize the entire frame. That kind of contextual awareness is what separates machine learning from brute-force warp stabilization.

Fast pan with motion blur. Here it fell over. The tracker lost the subject during a rapid camera whip, and I had to manually set a correction keyframe mid-shot. Not a disaster — the old tracker would have failed completely — but it's a reminder that "AI-powered" doesn't mean "works on everything."

Jose Antunes at ProVideo Coalition called IntelliTrack a significant leap for the DaVinci Neural Engine's capabilities. Having spent real time with it, I agree. For 80% of real-world tracking tasks, this eliminates manual keyframing entirely. That alone saves hours per project.

The gap between "mostly works" and "always works" is where professional tools earn their reputation. IntelliTrack is closer to the latter than anything Resolve has shipped before.

Voice to Subtitle: Good Enough to Be Dangerous

Automatic transcription is one of those features where "pretty good" can actually be worse than useless. If you have to re-check every line anyway, you might as well have typed it yourself.

DaVinci Resolve 21's Voice to Subtitle lands in a weird middle ground. On my project — two speakers, decent lavalier audio, standard North American English — it hit roughly 90-95% accuracy on first pass. It correctly segmented speaker changes most of the time, generated sensible line breaks, and placed subtitles on the timeline with proper timing.

Where it fell apart: one speaker had a slight accent and occasionally dropped consonants. The transcription mangled about one word per sentence for those sections. Technical jargon was hit-or-miss. It got "API endpoint" right but turned "Kubernetes" into "cooper nets." Honestly, relatable.

The workflow is fast, though. I went from raw audio to a working subtitle track in about three minutes for a 12-minute video. Cleanup took another 20 minutes. Compare that to the 45-60 minutes I'd normally spend transcribing and timing captions by hand.

If you're creating content where subtitles are mandatory — and in 2026, that's basically everything — this feature alone justifies the upgrade. It's not replacing professional transcription for broadcast, but for YouTube, internal videos, and social clips, it's more than good enough.

This follows a pattern I keep seeing across AI-powered tools that promise to automate creative work: they rarely eliminate the human entirely, but the best ones kill the tedious 80% and leave you with just the judgment calls.

The latency question matters on the infrastructure side too — if you're building AI-assisted workflows that call external models rather than running them locally, the gap between acceptable and frustrating is surprisingly narrow, which is something covered in depth in LLM Latency Benchmarks 2026: 6 Levers to Hit Sub-500ms TTFT.

I've been running some of these inference tasks locally rather than through cloud APIs, and the tradeoffs around setup cost versus ongoing savings are worth understanding — 10 Self-Hosted AI Tools That Replace SaaS [2026 Tested] covers exactly that ground if you're evaluating whether to bring transcription or other AI workloads on-premise.

That same dynamic — AI tools doing impressive work while still requiring human judgment at critical moments — plays out in software development contexts too, as I explored recently in AI Code Review Tools 2026 Compared: What Catches Bugs vs What Fakes Confidence, which benchmarks how well these systems actually perform on real codebases versus how confidently they present their results.

uTalk: The Sleeper Feature Nobody's Talking About

This is the DaVinci Resolve 21 AI feature I think is most underrated. uTalk lives in the Fairlight audio page, and it does something clever: it analyzes your video to identify where speakers are positioned on screen, then automatically pans the dialogue audio to match.

In a two-person interview, the speaker on the left gets panned slightly left. Speaker on the right, slightly right. It creates a subtle stereo image that makes dialogue feel more spatial and immersive — the kind of thing film mixers have been doing manually for decades.

I tested it on a segment where two people were talking across a table. The result was surprisingly natural. Not aggressive stereo separation — just enough spatial positioning that the audio felt connected to the image in a way it hadn't before.

The limitation is predictable: uTalk needs reasonably clean audio separation. If both speakers are on the same mic with heavy bleed, it can't magically pull them apart. It's working with spatial information it infers from the video, not performing source separation. Intelligent panning, not intelligent mixing.

For podcast-style videos and interview content, though, uTalk is a real quality upgrade that takes about five seconds to apply. That's my favorite kind of AI feature. The effort-to-improvement ratio is absurdly good.

Color Slice and the Grading Toolbox

Color Slice isn't an AI feature, but it landed alongside the AI announcements and it's too useful to ignore. It's a new vector tool on the Color page that enables subtractive color adjustments — think pulling specific hues out of the image rather than pushing them around.

The Film Look Creator, which has been part of Resolve's toolkit for a while, pairs well with Color Slice for quickly dialing in cinematic grades. Together they get you 70% of the way to a professional color grade in minutes, with fine control for the remaining 30%.

I used Color Slice to desaturate the slightly green-tinted fluorescent lighting in my office footage while keeping skin tones intact. In previous versions, I'd have reached for qualifier-based secondary corrections — doable, but fiddly. Color Slice made it a single adjustment. The subtractive approach just feels more intuitive for corrective work like this.

Nino Film noted on cineD that Color Slice and the new grading tools represent key creative additions for filmmakers. I think that's actually understating it. The Color page was already Resolve's crown jewel. These additions widen the gap further.

The Free vs. Studio Question

Here's the math that matters. DaVinci Resolve 21 Free includes IntelliTrack AI, Voice to Subtitle, uTalk, Color Slice, and the full editing, grading, effects, and audio toolkit. Free. $0.

DaVinci Resolve Studio costs $295. One time. Not per month. Not per year. You buy it once and get every update until a major version bump.

Adobe Premiere Pro runs on a monthly subscription. The exact price depends on your plan and region, but individual subscribers are typically paying north of $260 per year — and that's every year, forever. After five years you've spent over $1,300 for the privilege of continuing to access your editing software.

I've been thinking about the economics of tool choices a lot lately — it's a pattern that shows up across the entire software industry, where the cost structure of your tools quietly shapes every decision you make. In video editing, Blackmagic's pricing model is doing real damage to the subscription crowd. When the free tier includes professional-grade AI features, the conversation shifts from "can I afford Resolve?" to "can I justify paying for anything else?"

Studio is worth it if you need UltraNR noise reduction (excellent for low-light footage), resolutions above 4K, or multi-GPU rendering. For everyone else, the free version is shockingly complete.

Who Should Actually Upgrade?

After putting DaVinci Resolve 21's AI features through a real project, here's my honest take:

Upgrade immediately if: You're already using Resolve and you do any amount of tracking, captioning, or dialogue work. IntelliTrack and Voice to Subtitle will pay for themselves in time savings on your first project.

Seriously consider switching if: You're on Premiere Pro and tired of the subscription treadmill. The learning curve is real — Resolve's node-based color grading is different from Lumetri — but the AI features and the price make the switch increasingly hard to argue against.

Wait if: You're primarily doing motion graphics or heavily After Effects-dependent work. Fusion is powerful but it's not After Effects, and the ecosystem integration isn't there yet.

The AI features in Resolve 21 aren't gimmicks. IntelliTrack is legitimately best-in-class for its price tier (which, again, is free). Voice to Subtitle is a solid 90% solution. uTalk is a small touch that makes a real difference. And Color Slice, while not AI, rounds out the most complete color toolset in any NLE.

Blackmagic isn't trying to match Adobe feature-for-feature. They're doing something smarter — making the AI-powered editing workflow free, betting that once you're inside the Resolve ecosystem, you'll eventually buy Studio or their hardware. Same playbook that worked for open-source developer tools: give away the core, monetize the professional tier.

The question isn't whether DaVinci Resolve 21 is good. It's whether Adobe has any answer for a competitor whose free tier keeps getting better every single year. After this release, I honestly don't know what that answer looks like.

Photo by Amjith S on Unsplash.

DaVinci Resolve 20 vs 21: What Actually Changed?

If you're coming from Resolve 20, the upgrade question is legitimate: is version 21 a meaningful leap or a point release dressed up in NAB press releases? The short answer is that it depends heavily on how you edit. For anyone doing interview content, documentary work, or corporate video — the kind of project I was cutting — the AI additions in version 21 represent a genuine workflow shift. For pure narrative editors who mostly live in the Cut page, the differences are less transformative.

The Concrete Differences That Matter Day-to-Day

The DaVinci Neural Engine in version 20 already had capable tools — magic mask, smart reframe, speed warp — but they operated more like discrete features than a connected intelligence layer. Version 21's additions feel more integrated. IntelliTrack AI isn't just a better tracker; it's a different philosophy about what tracking should require from the editor. Voice to Subtitle wasn't meaningfully present in version 20 in the same form. uTalk is entirely new. These aren't refinements to existing tools — they're additions that open up new categories of work.

• IntelliTrack AI: replaces the older point-based tracker with neural object recognition — version 20 had no equivalent
• Voice to Subtitle: automatic transcription-to-timeline is a version 21 addition, not an upgrade from a prior system
• uTalk: didn't exist in version 20; spatial audio panning from video analysis is new to the Fairlight page
• Color Slice: the subtractive vector grading tool is version 21-only — version 20 editors relied on qualifier-based secondaries for similar corrections
• Neural Engine performance: Blackmagic describes the underlying engine as significantly upgraded, which benefits existing tools like magic mask and speed warp even if their interfaces look the same

The honest version of the comparison: if you skipped version 20 entirely and jumped from version 18 or 19, the gap feels enormous. If you updated regularly, some of version 21's improvements will feel incremental on familiar tools — but the three new AI features are genuinely additive. There's no version 20 equivalent of Voice to Subtitle generating a subtitle track from your timeline in three minutes. That's not iteration. That's a new capability.

DaVinci Resolve 21 Free vs Studio: Which License Do You Actually Need?

The licensing question comes up constantly, and it's worth being precise about it because most coverage gets fuzzy at exactly the point where the details matter. The free version of DaVinci Resolve 21 includes IntelliTrack AI, Voice to Subtitle, uTalk, and Color Slice — the features covered throughout this review. That's not a limited trial. Those tools are fully functional in the $0 tier. The $295 Studio license unlocks a different set of capabilities, and the distinction is more nuanced than 'free is crippled, Studio is complete.'

What Studio Actually Adds

The Studio license is primarily about GPU-accelerated processing for computationally expensive tools, collaboration features, and a handful of additional effects. UltraNR — Blackmagic's AI-powered noise reduction — requires Studio. So does the full version of the Film Look Creator in certain configurations, Dolby Vision output, and multi-user collaboration workflows. If you're cutting alone on a modern consumer GPU, you can go a long time before hitting a wall that only Studio removes.

• Free includes: IntelliTrack AI, Voice to Subtitle, uTalk, Color Slice, full edit/color/audio/effects pages, H.264 and H.265 export
• Studio adds: UltraNR AI noise reduction, GPU-accelerated processing for heavy Neural Engine tasks, multi-user collaboration, Dolby Vision and HDR10+ support, additional Resolve FX plug-ins
• Studio is a one-time purchase (~$295), not a subscription — that single payment covers all future version upgrades
• Both versions work on Mac, Windows, and Linux

For most solo creators and small teams, the free version covers the AI features that actually move the needle on productivity. The case for Studio strengthens if you're delivering broadcast content with strict noise requirements, working collaboratively across a network, or processing long-form content where GPU acceleration meaningfully reduces render times. At $295 as a one-time cost — no annual renewal, no seat fees — it's also one of the more rational purchasing decisions in professional software. But don't let the Studio upsell pressure you into thinking the free tier is a demo. For the work described in this review, free handled everything.

How the DaVinci Neural Engine Works (And Why It Matters for AI Features)

The term 'AI features' gets applied to everything from simple macros to genuinely sophisticated machine learning systems. In DaVinci Resolve 21, the engine underneath the AI tools is the DaVinci Neural Engine — Blackmagic Design's on-device machine learning framework that has been evolving since its introduction several versions back. Understanding roughly what it does helps explain why some features work well and others have predictable failure modes.

The Neural Engine runs inference locally on your hardware — it's not sending footage to a cloud server for processing. That matters for two reasons: speed and privacy. On a modern GPU, most Neural Engine tasks complete faster than real-time or close to it. For editors working with sensitive client footage, the fact that nothing leaves the machine is meaningful. The tradeoff is that performance scales with your hardware. On a lower-end integrated GPU, tasks like IntelliTrack or Voice to Subtitle will run noticeably slower than on a dedicated GPU — though they'll still complete correctly.

What the Neural Engine Is Actually Doing in Each Feature

• IntelliTrack AI: runs object recognition to understand what the subject is, not just where it is — this is why it can maintain a lock through partial occlusion where a point tracker loses the thread entirely
• Voice to Subtitle: performs speech-to-text inference on the audio track, including speaker segmentation and timing alignment — accent sensitivity and technical vocabulary are the known weak points of most models in this category
• uTalk: analyzes video frames to infer spatial positioning of speakers, then uses that information to drive audio panning — it's doing computer vision work to serve an audio outcome
• Speed Warp (carried over): optical flow analysis that generates interpolated frames for slow motion — one of the earlier Neural Engine tools, still among the best at this task in any NLE

The pattern across all of these is consistent with how on-device machine learning typically behaves: strong performance on common cases, degraded performance at the edges. A speaker with a slight accent, a tracking shot with aggressive motion blur, audio with heavy bleed between mics — these push the Neural Engine toward its limits in predictable ways. That's not a criticism unique to Blackmagic. It's the current state of on-device inference. What makes Resolve 21 notable is that Blackmagic has identified the right problems to apply it to — the tedious, repetitive tasks where 'pretty good, most of the time' is genuinely better than the manual alternative.

GPU requirements: what you need for Magic Mask, subtitles, Voice Isolation, and NR

I’m going to be blunt: most “minimum requirements” pages are basically fan fiction. They don’t map to your footage, your codecs, your timelines, or your patience.

Here’s how I’d tier hardware for Resolve 21 AI features, in creator terms.

The 3 GPU tiers that matter

Tier 1: “It runs, but you’ll manage it”

• Expect proxies and aggressive render cache.
• Neural Engine features will work, but you’ll wait.
• VRAM is often the limiter.

Tier 2: “Comfortable Studio workflows”

• Most AI features are usable without constant babysitting.
• Still benefits a lot from proxies for long-form.

Tier 3: “Client-scale, high volume”

• Neural Engine analysis becomes “background time,” not “stop the world.”
• Multicam + subtitles + masking becomes feasible without constant compromises.

Practical VRAM guidance by feature (rule-of-thumb)

I’m not going to invent benchmark numbers here. But you can still make good decisions with rules of thumb:

• Magic Mask / object masking: usually the most VRAM-sensitive. Long clips at 4K push you into “comfortable” territory fast.
• Super Scale upscaling: heavy compute. Plan it as an offline step. Don’t expect it to feel interactive.
• Voice Isolation: compute-heavy but usually less VRAM-explosive than masking.
• Transcription/subtitles: often feels more CPU and implementation dependent. It’s less about VRAM spikes and more about how quickly your system can chew through audio.

How to tell when you’re GPU-bound (without becoming a hardware nerd)

• If playback is fine but Neural Engine analysis crawls, you’re probably GPU throughput-bound.
• If analysis starts and then fails, or Resolve gets flaky when you stack effects, you may be VRAM-bound.
• If transcription is slow while GPU looks idle, you may be CPU-bound or on a path that isn’t accelerating on your hardware.

If you’re on Apple hardware, unified memory changes the mental model. It behaves like a big shared pool. Stuff that “should” crash often just… loads. But throughput still matters. The same confusion shows up when people jump into Apple Silicon for a local LLM: memory feels magical, speed becomes the trade.