The Assembly Edit: Cutting Mismatched AI Shots Together in DaVinci Resolve 21 (Conform, Trims, J/L-Cuts, Optical Flow, Export)

Let's run the news. The big story this week is a new image model from ByteDance, and yes, it belongs in a video show, because it feeds the frames your video model turns into shots.

On July eighth, with the official announcement the next day, ByteDance's Seed team shipped Seedream five point oh Pro. Quick naming warning, because this one trips people up out loud. It's Seedream, like seed plus dream, and that's their image line. Seedance, with a D, is their video line.

Here's what it does. It generates natively up to two K, around two thousand forty-eight pixels on a side. That's actually lower than Nano Banana Pro, which still does native four K and stays ahead on resolution, texture, and text accuracy. So where does Seedream have the edge? Reasoning and editing. It runs a deep-thinking pass on your prompt and can do a real-time web search while it generates. You can edit with real precision, marking the exact region you want changed with a point, a lasso, a box, arrows, or sketches. And here's the production hook. It separates layers, handing you a background plus however many element layers as transparent alpha PNGs. Compositable pieces, not one flat frame. It also writes text onto the image in more than ten languages. ByteDance is pitching the whole thing as production design, not one-off pictures.

On where you can get it, it went live first on Volcano Ark, which is Volcano Engine, and it's rolling out to the Doubao and Jimeng apps, Jimeng also being called Dreamina, plus the BytePlus, fal, and Magnific APIs. Pricing runs from about three cents an image at lower resolution to about nine cents at two K, with fal quoting about six and three-quarter cents up to fifteen thirty-six pixels and about thirteen and a half cents up to twenty forty-eight. Runway added it to their API on July eighth too, with multi-image fusion, the interactive editing, prompts up to four thousand characters, up to ten reference images, fourteen aspect ratios across one K and two K, at five credits for a one K image and nine for a two K. Want to kick the tires? Pull it through Runway or fal and test that alpha-layer export on a reference-driven shot.

Quick leaderboard read from the Artificial Analysis Video Arena, which reshuffles every month. Two movers since our last read. On text-to-video without audio, Alibaba's HappyHorse still holds the top two spots, but Wan two point seven climbed into the top tier at fourth, with Dreamina Seedance two point oh third and Kling three point oh Pro fifth. On the with-audio board, Seedance still leads native audio, and Wan jumped all the way to number two. And on image-to-video without audio, Seedance leads, with grok-imagine-video one point five preview now number two and PixVerse V six third. That grok model isn't new; it went generally available June sixteenth. Only its ranking moved, so treat it as context, not a launch.

And one date for the horizon, not this week. The EU AI Act's Article fifty transparency rules take full effect August second, with C2PA and SynthID both named as accepted ways to mark AI-generated content. Worth putting on your calendar now.

Okay. Today we're doing the assembly edit. You've got a folder full of AI clips, they were never filmed together, they don't match, and your job is to cut them into one sequence that feels like a single piece of video instead of a pile of generations stapled end to end. This is the episode where the shot list finally becomes a movie.

An assembly edit just means you lay every shot down in order, end to end, before you refine anything. It's the rough string of the whole thing. You don't polish, you don't fuss over a single cut, you just get the shots on the timeline in the right sequence so you can watch it start to finish. That string-out is the assembly. Everything after it is refinement.

Let's set the tool. We're in DaVinci Resolve, and I want to anchor you to the version so nothing on screen surprises you. The current release as we record is Resolve twenty-one, which shipped final in June of twenty twenty-six, and the latest little point update in early July is twenty-one point oh point two. Twenty-one added a Photo page, which is a Lightroom-style still editor with camera RAW, plus a batch of AI tools with names like IntelliSearch, CineFocus, an AI Speech Generator, and AI Face tools, an integrated motion-graphics library called Krokodove, keyframe easing, and audio folders in the Fairlight side. Here's the thing though. You need exactly none of those AI-page features to do a first assembly edit. They're optional. I'm mentioning them so you know what those extra buttons are, not because we'll touch them today.

The pages along the bottom of Resolve, left to right, are Media, Cut, Edit, Fusion, Color, Fairlight, and Deliver, plus that new Photo page in twenty-one. Today we live in Media, Cut and Edit, a light touch of Color, Fairlight for audio, and Deliver to export. That's the whole trip.

Before we go further, let me define the tool itself, because if you're brand new, the acronym matters. Resolve is an NLE, which stands for non-linear editor. The non-linear part is historical. Back in the tape and film days, editing was linear. You assembled shot one, then shot two, then shot three, physically, in order, and if you wanted to slip a new shot in between one and two, you basically had to redo everything after it. An NLE frees you from that. You can drop any clip anywhere, reorder them, trim them, change your mind, and none of it touches your original files. This is the single most important idea in editing, so let me say it plainly. Your source files are never altered. The timeline is just a set of instructions that points at those files. Editing changes the recipe, not the ingredients. When you cut a clip shorter, the clip on disk is exactly as long as it ever was. You're just telling Resolve which part to show.

A couple more quick definitions so the rest lands. A codec is the compression format a video file is encoded and decoded in, things like H two six four, H two six five, or ProRes. It's how the pixels are packed. That's a different thing from resolution, which is the pixel dimensions, different again from aspect ratio, which is the shape of the frame, and different again from frame rate, which is how many frames play per second. Keep those four separate in your head and a lot of confusion goes away. The timeline is the horizontal canvas where your clips lay out left to right in time, and the vertical lanes stacked in it are called tracks.

What makes Resolve special is that it's Blackmagic Design's free, professional-grade, all-in-one app. Edit, color grading, Fusion visual effects, and Fairlight audio all live in one program, so you're not exporting out to one app for sound and another for color and round-tripping between them. It's all one project.

Let's talk money and machine, because people assume free means crippled. It doesn't. The free version of Resolve is genuinely free. No watermark, no time limit, no nagging. Studio is a one-time two hundred ninety-five dollars, and I want to stress one-time, not a subscription. The free version exports up to four K UHD, that's thirty-eight forty by twenty-one sixty, at up to sixty frames a second. Studio goes way past that, up to thirty-two K and a hundred twenty frames, but you don't need it. Free is single-GPU. Studio unlocks multi-GPU and the DaVinci Neural Engine, which powers Magic Mask, better noise reduction, and most of the AI features, plus Dolby Atmos and multi-user collaboration. For a creator delivering to YouTube or TikTok at ten eighty or four K, the free version is fully sufficient. Don't let anyone tell you otherwise.

On the machine itself, be aware that Resolve twenty-one drops support for Intel Macs. It wants Apple Silicon and macOS fifteen Sequoia, and of course it also runs on Windows and Linux. It leans hard on your graphics card. Figure sixteen gigs of RAM or more, and a dedicated GPU with at least four gigs of video memory for HD work, more if you're going to four K.

Let me give you the quick tour of the seven pages so you know where things live. Media is where you ingest and organize, that's your Media Pool and clip attributes. Cut is the fast editing page, single-click everything, built for speed and turnaround. Edit is the traditional timeline with full precision, the Inspector, transitions, titles, and advanced trimming. Fusion is node-based visual effects, and we skip it today. Color is grading, which is next episode, and today we only borrow one tool from it. Fairlight is audio. Deliver is export. The teaching point that ties it together is that all these pages share one project and one timeline. You can bounce between Cut and Edit as much as you want on the same sequence.

So which do you use, Cut or Edit? The Cut page is built for speed. Everything happens live with a single click, and it has a feature called Source Tape that stacks every clip in your bin into one long scrubbable tape, so you don't have to hunt through folders to find a shot. It's great for social and fast turnaround. But the Cut page cannot unlink audio from video, and that's a real limitation for the split-edit work we'll do later. The Edit page gives you the full timeline, a Source viewer and a Timeline viewer side by side, the Inspector, precise trims, transitions, and the ability to unlink audio and video for J-cuts and L-cuts. My framing is simple. Rough the order fast on Cut, or just drag clips into Edit, and refine on Edit. The real split-edit surgery has to happen on the Edit page.

Now let's actually get clips in, and there's a critical gotcha here that I want you to hear before you make the mistake. The Media Pool is your project's library of imported clips. You import by dragging files in, or right-clicking and choosing Import Media or Import Media Folder. Importing references the files where they sit on disk. It does not copy or move them. So if you move the originals later, you'll have to relink them.

Here's the part to do first, before you import anything. Set your project resolution and frame rate. Go to Project Settings, that's the little gear in the bottom right, then Master Settings, and set your Timeline Resolution, something like nineteen twenty by ten eighty or thirty-eight forty by twenty-one sixty, and your Timeline Frame Rate, twenty-four, twenty-five, thirty, or sixty. Do this before you import a single clip. Why the urgency? Because of the gotcha. The moment any clip lands in your Media Pool, the Timeline Frame Rate control greys out and locks. That's by design, not a bug. If you already imported and now you're stuck, you have two ways out. You can delete every clip from the Media Pool, which re-enables the control, set your frame rate, and re-import. Or, mid-edit, you create a new timeline, uncheck "Use Project Settings," set the frame rate on the Format tab of that new timeline, create it, and copy your clips over into it. The clean move is just to decide your delivery frame rate first and set it before anything comes in. Pick your delivery frame rate on day one.

Now the heart of the problem with AI footage. These generators spit out wildly inconsistent files. One clip is twelve eighty by seven twenty at twenty-four frames. The next is ten eighty by nineteen twenty, vertical, at thirty frames. A third one is sixteen frames a second, and each one is a different codec. Your timeline is one resolution and one frame rate, so Resolve has to conform everything, which just means normalizing every clip to that single timeline standard. Let's walk through how.

For resolution and aspect ratio mismatches, the setting is Input Scaling. Go to Project Settings, then Image Scaling, then Input Scaling. You've got a few choices, and they matter. "Scale entire image to fit" is the default, and it fits the whole frame inside your timeline, adding letterbox bars on the sides or top so nothing gets cropped. "Scale full frame with crop" fills the frame edge to edge and crops off whatever overflows. "Stretch frame to all corners" distorts the image to fill, and you want to avoid that one, especially on faces, because it makes everybody look squished or stretched. There's also a center crop with no resizing. My advice for a mixed pile is to set a sensible global default, usually scale to fit, and then override the oddballs one at a time. To override a single clip, select it, open the Inspector, go to Retime and Scaling, then Scaling, and set Crop, Fit, Fill, or Stretch just for that clip. So if you've got a stack of horizontal clips and three rogue verticals, set the global default for the majority and hand-fix the three.

Frame-rate mismatches have two paths, and it's worth knowing both because they do different things. The first path is the Retime Process, specifically Optical Flow. In the Inspector, under Retime and Scaling, there's a Retime Process setting. Optical Flow analyzes the motion in the clip and generates brand-new in-between frames to bridge the gap between the clip's frame rate and the timeline's. The alternatives are Nearest, which just drops or duplicates frames and tends to stutter, and Frame Blend, which dissolves between frames and can leave ghosting. Optical Flow looks the best, but it's expensive to render, so use your render cache so playback stays smooth. The second path is Clip Attributes. Right-click the clip in the Media Pool, choose Clip Attributes, go to Video, and change the Video Frame Rate. This reinterprets the clip project-wide. You can force a sixteen-frame clip to be read as twenty-four frames, for example. Here's the difference in one line. Optical Flow smooths the motion by inventing frames. Clip Attributes just relabels the clip's frame rate. Different tools for different situations.

Let's talk about the timeline itself. Tracks stack. Video tracks go top to bottom, and higher covers lower, so whatever's on V two plays over whatever's on V one underneath it. Audio tracks, A one, A two, and so on, sit below. To add a track, right-click a track header and choose Add Track, or just drag a clip above the top track and Resolve makes a new one. The playhead is that vertical line marking your current frame. On the Edit page you've got two viewers. The left one is the Source viewer, showing your raw clip, and the right one is the Timeline viewer, showing your edited sequence. You mark an In point and an Out point on the source with the I and O keys, then push that trimmed piece to the timeline. And snapping, which is when clip edges and the playhead magnetically stick to cut points so you land exactly on an edit, toggles with the N key. Get comfortable turning it off and on.

Now the core edits, the ways you actually get a clip from source onto the timeline. The umbrella idea is three-point editing. You set three of the four points, source In, source Out, timeline In, timeline Out, and Resolve calculates the fourth for you. The one you'll use most in an assembly is Append, which adds the clip to the end of the timeline, and the shortcut is Shift plus F twelve. That's the fastest way to lay a shot list down in order, just append, append, append. Insert drops a clip at the playhead and ripples everything after it later to make room, that's F nine. Overwrite drops at the playhead and replaces whatever's underneath, keeping the timeline the same length, that's F ten. There's also Replace on F eleven, Fit to Fill on Shift plus F eleven, Place on Top on F twelve which puts the clip on the track above for cutaways and B-roll, and Ripple Overwrite on Shift plus F ten. For today, Append is your workhorse.

Trims and the blade. The Razor, or Blade, slices a clip at the playhead. Press B for the blade tool, or use Control or Command plus B to blade at the playhead without even switching tools. There's a Trim Edit Mode on the T key that's smart, it auto-switches between ripple, roll, slip, and slide depending on where your cursor is on the clip. Let me define those four trims because they trip everyone up. A ripple trim is when you drag a clip edge and the neighbors slide over to close or open the gap, so the total length of your timeline changes. A roll trim moves the cut point between two clips, one gets shorter as the other gets longer, and the total length stays the same. A slip trim changes which part of the footage shows without moving the clip at all, the position and duration stay fixed while the in and out content slides underneath. And a slide trim moves a clip earlier or later while its own content and duration stay fixed, and the neighbors absorb the change. Ripple changes length, roll moves a cut, slip changes content in place, slide moves a clip in place. Also key, ripple delete removes a clip and closes the gap in one move, that's Shift plus Delete. Plain Delete leaves a hole. And Dynamic Trim, on the S key, lets you trim live using the J, K, and L keys.

Which brings me to J, K, and L, the most important keys on your keyboard for editing. J plays backward, K pauses, L plays forward. Tap L or J repeatedly to speed up, two times, four times, eight times. Hold K and tap L or J to jog a single frame at a time. Spacebar plays and stops. I and O set in and out. Up and Down arrows jump edit to edit. Left and Right step one frame. N toggles snapping, A is the selection tool, M drops a marker. Learn J, K, L first. Once your hands know them, you review footage at the speed of thought.

Alright. The clips are in, they're conformed, they don't stutter. Now the actual craft, the part that separates a video that reads as one scene from a video that reads as a bunch of AI clips glued together. This is durable stuff. It'll outlive every model we talked about in the news.

Your shot list is your assembly order. This is the callback to the storyboarding episode. That shot list you built is the sequence you Append in. String it out end to end first, in order, then refine. That string-out is the assembly edit, full stop.

The single most powerful trick for hiding that two AI clips were never filmed together is match-on-action. That means you cut during a movement, so the motion carries across the join and the viewer's eye follows the action instead of noticing the cut. Find a common gesture across two shots, someone raising a hand, turning their head, reaching for something, and cut on the peak of that motion. The brain is so busy tracking the movement that it glides right over the edit. When two of your clips share any kind of motion, cut mid-motion.

Next, J-cuts and L-cuts, which editors call split edits. A split edit is when the audio and the picture cut at different moments instead of at the same instant. An L-cut is when the picture cuts to the next shot but the outgoing audio lingers over it, so you're already seeing speaker B while you still hear speaker A finishing a sentence. A J-cut is the reverse, the next clip's audio starts early, under the current picture, so you hear the next scene before you see it. To do this on the Edit page, right-click the clip and choose Unlink, or press Control or Command plus Shift plus L, then ripple or roll just the audio edge on its own, then Link it back when you're done. On the Cut page, which can't unlink, you fake it using Trim, then Ripple Start to Playhead or Ripple End to Playhead, with Control or Command plus Shift plus the left or right bracket. Why do split edits matter so much for AI footage? Because they glue clips together sonically. When the sound flows across a cut, mismatched visuals feel continuous, and you hide that tell-tale seam where both the picture and the sound slam-cut at the exact same frame. That simultaneous cut is one of the biggest giveaways of amateur editing, and split edits kill it.

Cutaways and B-roll are your next patch. A cutaway is a brief shot of something else, hands, a detail, the environment, the sky, inserted over the main action. If you've got an ugly join between two shots, or the lip-sync drifts, which is a callback to the dialogue episode, drop a cutaway over the trouble spot. Put it on V two, above V one, using Place on Top with F twelve, so your base audio underneath keeps running uninterrupted. The viewer looks at the hands for a second, and when you cut back, the problem is gone and they never knew it was there.

Two continuity rules to keep in mind. The one-eighty-degree rule imagines a line running between your two subjects, and you keep the camera on one side of that line across your cuts, so screen direction stays consistent. Cross the line and a character seems to teleport to the other side of the frame. Related is the eyeline match. If character A is looking screen-right at B, then B should be looking screen-left back at A. Here's the catch with AI. The generator has no continuity memory, none. It doesn't know where anyone was looking in the last clip. So you have to pick and prompt your shots to respect the line yourself. That's on you, not the model.

Pacing and rhythm. Vary your shot lengths. Hold your establishing shots longer to let the viewer take in the space, and cut your action tighter and faster. A pile of AI clips that are all the same length feels mechanical and robotic, because it is. The variety is what makes it feel like a human made choices.

And now my favorite AI-specific fix, the AI stare and the dead frames at clip ends. This is a callback, and it's a real pattern. These generators tend to decelerate at the end of a clip. The motion slows down, subjects freeze or drift or go a little rubbery as the clip runs out of prompt to work with. You'll see the last two to eight frames go soft, static, or warpy, and there's often a soft ease-in at the head of the clip too. The fix is to ripple-trim the decelerating tail off every clip, and trim the soft head, or just cut away right before the tail hits. If the good motion is in the middle, use a slip trim to slide your usable window inward and center it. Do this on every single clip, and you strip out most of that tell-tale AI limpness that shows up right at your cut points. This one habit alone will make your edits look dramatically less AI.

Now, color. We are not grading today. That's next episode. But your AI clips are going to arrive with different exposure and different white balance, and if you cut straight from a bright clip to a dark one, the picture flashes light to dark at the cut and it's jarring. So we do a quick match, a bridge, just enough that the cut stops flashing. The tool is Shot Match, and it lives on the Color page. Select a clip you like as your reference, then right-click the clip you want to change and choose "Shot Match to This Clip." Resolve automatically matches the color, contrast, and brightness as a starting point. Even lighter than that, you can use the Color Wheels and hit Auto Balance, or just nudge Lift, Gamma, and Gain by hand. Keep the whole thing to one question. Does the cut stop flashing? That's it. No looks, no LUTs, no style. The reason a light match is all you need here is the callback to the consistency episode, where you already locked your style and look at generation time. So we're just evening out the exposure jumps, and the full grade waits for next time.

Audio. Some AI generators emit synced sound with the clip. Many text-to-video clips are silent. Put your audio-bearing clips on their own tracks, and know that the silent ones will need sound added later. When video and audio import together, they come in linked. Unlink them, right-click or Control or Command plus Shift plus L, when you need to edit them separately, whether that's for a J or L-cut or to delete bad generated audio while keeping the picture. Link them back to re-couple. Use a scratch track, which is just a placeholder audio track for temp music, temp voiceover, or temp sound so you can feel out your timing. Label it, put it on A three, call it SCRATCH, and mute or delete it before final. And keep your stems separate. Stems means your dialogue, your music, and your sound effects live on their own separate tracks and never get mixed down into one. This is a callback to the scoring episode. If you keep them separate, the scoring pass can add or swap music cleanly without touching your dialogue. Resolve twenty-one's Fairlight folders help you group those stems tidily.

Delivery. Head to the Deliver page. You've got Render Settings and a row of presets, YouTube, Vimeo, social, ProRes Master, H two six four, H two six five, Custom, and they auto-pull your timeline's resolution and frame rate. Quick codec refresher. H two six four is the universal one, most compatible, plays everywhere, that's your standard MP four. H two six five, also called HEVC, is newer, gives you about the same quality in a smaller file and better ten-bit handling with less banding, but it's slightly less universal. Bitrate is the amount of data per second, and higher bitrate means better quality and a bigger file. You set it under Quality, "Restrict to" a value. Good targets, ten eighty is roughly fourteen to twenty megabits, four K at thirty is roughly forty-four to fifty-six, four K at sixty is roughly sixty-six to eighty-five. Here's a real tip. Resolve's YouTube preset defaults its bitrate low, so override it upward, so your quality survives YouTube's recompression when you upload. Turn on Frame Reordering for B-frames, turn on Network Optimization, and set the encoding profile to High for H two six four. For vertical social, ten eighty by nineteen twenty, that's nine by sixteen, H two six four. For YouTube long-form, nineteen twenty by ten eighty or thirty-eight forty by twenty-one sixty, frame rate matching your timeline. Set your filename and destination, hit Add to Render Queue, then Render All. Remember the free version tops out at four K sixty, which is plenty for YouTube and social.

Let me give you the one pitfall to burn into memory, because it's the one that will bite you. Frame-rate mismatch making your cuts stutter. The symptom is sneaky. Playback looks fine inside any single clip, but every cut hitches, motion judders, and worst of all, it renders out that way, so you can't unsee it. The cause is clips at a different frame rate than the timeline, or the wrong project frame rate to begin with. You recognize it because the judder concentrates at cut points and in motion. The fix is what we already covered. Set your timeline frame rate deliberately before you import, remembering it locks once media is in, so to change it later you make a new timeline with Use Project Settings turned off. Then push any off-frame-rate clips through Optical Flow, or reinterpret them with Clip Attributes, and render-cache the Optical Flow clips so they play smooth. There's a related trap too, timeline-resolution mismatch, which makes everything look soft or cropped, and that's wrong Input Scaling. Fix it by setting your true delivery resolution and choosing the right Fit, Fill, or Crop. But the frame-rate stutter is the big one. Lead with it.

Let me give you the whole thing as one copyable workflow, start to finish, pile of clips and a shot list going in, finished export coming out. Decide your delivery first, say YouTube at nineteen twenty by ten eighty at twenty-four, or ten eighty by nineteen twenty for social. Make a new project, and in Master Settings set your Timeline Resolution and Frame Rate before importing anything, set Input Scaling to scale entire image to fit, and set your default Retime Process to Optical Flow. Go to the Media page and import all your clips, tagging them to your shot-list numbers. Create your timeline. Then do the assembly pass, going down the shot-list order and appending each clip with Shift plus F twelve, end to end, building your string-out. Next, trim the AI tells, ripple off the decelerating tails and the soft heads, slip to center the good motion, and ripple-delete the dead space with Shift plus Delete. Then refine the joins, match on action, and where you've got an ugly join or drifting lip-sync, place a cutaway on top on V two with F twelve, add your J and L-cuts by unlinking the audio and leading or lagging it, and check your one-eighty and your eyelines. Handle audio, cull the bad generated sound, put dialogue, ambience, and temp music on separate stems, keep a labeled SCRATCH track, and even out your levels. Do the light color bridge, Shot Match the obvious jumps and nothing more. Then deliver, H two six four MP four, resolution and frame rate matching your timeline, bitrate around twenty megabits for ten eighty or fifty to sixty for four K, Frame Reordering and Network Optimization on, name your file, Add to Render Queue, and Render All.

One last thing on choosing which generator to shoot with in the first place, which is a different question from editing. For that, look at the Artificial Analysis Video Arena, the text-to-video leaderboard. It's an Elo ranking built from blind human A-B votes, people picking which clip looks better without knowing which model made it. As of mid twenty twenty-six it's topped by Dreamina Seedance two point oh, Wan two point seven, and Kling three point oh, and the with-audio and without-audio boards rank differently. But here's the durable advice. Treat the models as basically interchangeable and benchmark them on your own shots. The leaderboard is a shortlist, not a verdict. What wins the arena might not win on your specific scene.

And to close, a quick word on alternatives to Resolve, one line each. Resolve is our default finisher, free, professional, all-in-one edit, color, Fusion, and audio, the steepest learning curve but the most capable free tool there is. CapCut is free, the fastest thing going for vertical social, with AI auto-edit that gives you a rough cut in about fifteen seconds, Smart Cut for pulling out dead air, and the best auto-captions in the business, though it's less precise and it's tied to the cloud and an account. Adobe Premiere Pro is the industry-standard subscription NLE, maximum control and a big collaboration setup, but it's paid and heavier to run. Honestly, a lot of pros use CapCut for volume and social and reach for Premiere or Resolve for their hero and client work. Pick the one that matches the job in front of you, and go make something that cuts together clean.