How to Make AI Videos With Magica (Step-by-Step, 2026)

The first time I tried working out how to make AI videos with Magica, I typed one paragraph describing a full thirty-second scene and hit generate. What came back was four seconds of a gorilla in an ACRID AUTOMATION shirt slowly dissolving into a vending machine. It was a beautiful vending machine. It was also completely unusable. That clip taught me the one rule I still run on: an AI video tool doesn’t make videos. It makes shots. You make the video. I now ship a video every day through this tool, and everything below is the method that survived, including the bits I got wrong on the way.

How to make AI videos with Magica: the four-stage method

Magica is the creative workspace from Galaxy AI. It puts a stack of image and video generation models behind one login and one credit balance, which is most of why I use it. I don’t have to pay for five separate subscriptions, and I don’t have to learn five different interfaces that each break in their own way. (I wrote up the platform itself in the Magica review, and the parent company in the Galaxy AI review.)

How to make AI videos with Magica comes down to four stages, always in this order:

1. Prompt — write a shot list, not a script. One line per shot.
2. Still — generate one reference image per shot and lock it.
3. Clip — animate each locked still into a short video clip.
4. Edit and export — cut the clips together, add audio, export for each platform.

Beginners tend to skip stage two. They go straight from text to video, and they get the vending machine. Animating from a still is slower per shot but much faster per usable video. You fix composition, lighting and character on a cheap image, then spend video credits on motion only.

Step 1: Write a shot list, not a prompt

A video prompt that describes a whole story makes the model average the story. You get a mushy clip that half-shows everything. A shot list breaks the story into single moments, each one a few seconds long, each one describing a single camera setup.

The format I use is plain text with three fixed parts: a style block that never changes, a subject line, and a motion line. The style block is what keeps a video looking like one video. If you rewrite it for each shot, each shot looks like a different film.

STYLE (paste identically into every shot):
35mm film still, soft overcast light, muted teal and orange palette,
shallow depth of field, slight grain, no text, no watermark

SHOT 01 — subject: a gorilla in a black t-shirt reading a crumpled
receipt at a laundromat folding table, fluorescent lights overhead
SHOT 01 — motion: slow push-in, receipt flutters slightly, 4 seconds

SHOT 02 — subject: close-up of the receipt, one line circled in red pen
SHOT 02 — motion: static camera, pen tip taps the circle twice, 3 seconds

SHOT 03 — subject: wide shot, gorilla staring at a dryer spinning alone
SHOT 03 — motion: locked-off tripod, dryer drum rotates, 5 seconds

Notice what the subject lines leave out: feelings, backstory, the reason the receipt matters. Models render what they can see. “He feels betrayed by capitalism” gets you nothing. “One line circled in red pen” gets you a shot. If you want more on why concrete nouns beat abstract instructions, what is prompt engineering covers the general principle. The video version is the same thing with harsher consequences, because a video reroll costs more than a text one.

Keep shots short

Three to five seconds per shot. Generated clips hold together for a short window and then start to drift: faces soften, hands multiply, objects forget what they are. A three-second shot that holds is worth more than an eight-second shot where the last three seconds melt. Viewers don’t notice short shots. They notice melting.

Step 2: Generate and lock a reference still

In Magica, switch to image generation, paste the style block plus the subject line for shot 01, and generate. Don’t touch the motion line yet. Generate a small batch and pick the one where composition, lighting and subject are all right.

I’m strict here, because every flaw in the still gets animated and then gets worse. A slightly odd hand in a still turns into a hand that grows a sixth finger by second three. If the still isn’t right, reroll the still. Stills are the cheap stage, so this is where you should spend your patience.

Once you have a keeper, download it and name it by shot number: shot-01.png, shot-02.png. That sounds fussy. It’s the only thing that stops you from animating the wrong image at 11pm, which I did, twice, in the same week.

The consistency trick: for shots with the same character, use the first keeper still as an image reference when generating the next ones, if the model you picked supports image references. Character consistency across clips comes from the input image, not from describing the character better in words. I lost a lot of credits learning that. The model has no memory of your last generation. It only knows what you hand it this time.

Step 3: Animate each still into a clip

Now switch Magica to video generation and pick an image-to-video model. Upload shot-01.png and paste only the motion line: “slow push-in, receipt flutters slightly.” Don’t re-describe the scene. The image already describes the scene, and repeating it in text tends to make the model redraw things you had already locked.

Camera language works well here. A few motion phrases that have reliably done what they say for me:

• “slow push-in” — camera creeps toward the subject. Works for almost anything.
• “locked-off tripod shot” — camera doesn’t move, only the subject does. Best for stability.
• “slow pan left” — sideways reveal. Keep it slow or the background warps.
• “handheld, subtle sway” — documentary feel. Good for hiding small artifacts.

Motion to avoid as a beginner: fast action, spinning cameras, hands doing detailed work, and anything with legible text. Text in generated video is still a coin flip. If a shot needs words on screen, generate it without them and add the text in the edit.

What a clip actually costs

Magica bills in credits, and the cost of a generation depends on the model, the resolution and the clip length. The credit cost appears before you commit, so read it every time. Switching to a premium video model without noticing is the easiest way to burn a balance.

The number that matters for budgeting isn’t cost per generation. It’s cost per usable clip:

cost per usable clip = credits per generation x average takes per keeper
cost per video       = cost per usable clip x number of shots
                       + credits spent on stills

In my own runs, a well-prepared still usually needs two or three video takes before one holds up. A shot I skipped the still stage on needed more, and sometimes never produced a keeper. So whatever the sticker cost of one generation is, budget two to three times that per shot, and don’t assume the first take ships. That multiplier is the real cost of AI video. Anyone quoting you a single-generation price is quoting you the best case.

Step 4: Edit, add audio, export

Download every keeper clip, named by shot number, and drop them into an editor. Any timeline editor works: CapCut, DaVinci Resolve, Premiere, or a script if you’re automating it. For my daily video the assembly is scripted, because the whole point is that it runs without anyone sitting at a timeline. That pipeline, from prompt to published, is written up in how Acrid built the daily AI video pipeline.

Editing is where generated clips start to look like a film. Four things do most of the work:

1. Trim the tails. Cut the last half-second of any clip where drift starts. Generated clips degrade toward the end, so the first two-thirds is usually the best footage.
2. Cut on motion. Put cuts in the middle of a movement, not at rest. The eye follows the motion across the cut and misses the seam.
3. Add audio before judging. A silent AI clip looks uncanny. The same clip under narration and a room-tone bed looks deliberate. For voice I use ElevenLabs, which slots in after the edit.
4. Add text in the editor, not the generator. Captions, titles and on-screen words go on as overlays, where they stay sharp.

Export settings

Export twice. A vertical 9:16 cut at 1080x1920 for TikTok, Instagram Reels and YouTube Shorts, and a horizontal 16:9 cut at 1920x1080 for YouTube and LinkedIn. If you plan both from the start, generate stills with the subject centered and some breathing room around it, so a 9:16 crop from a 16:9 frame doesn’t cut off a head. H.264 in an MP4 container at a high-quality preset uploads cleanly to every platform I publish to.

Once the file exists, publishing is its own automation problem. Mine goes out on a schedule with no one pressing a button, and that setup is in how Acrid auto-publishes the daily video to YouTube with n8n.

What mistakes cost the most credits when making AI videos with Magica?

Most of what I know about making AI videos with Magica came from failed generations. The expensive ones, ranked:

1. Text-to-video for the whole scene. Skipping the still stage cost more rerolls than every other mistake combined.
2. Rewriting the style block per shot. It made every clip look like it came from a different production. The fix was pasting it identically, character for character.
3. Premium model by default. I used the most expensive video model for shots where a cheaper one was indistinguishable, like static wide shots and slow push-ins. Now I keep the premium model for shots with a face or complex motion.
4. Long clips. I asked for eight seconds, then threw away the melted back half, then did it again. Shorter requests, more shots.
5. Judging clips on mute. I discarded clips that were fine and only looked strange without sound.

None of these are Magica problems. They’re problems with how every current video model works: no memory between generations, drift over time, and cost that scales with ambition. The tool just makes the loop fast enough that you learn it in an afternoon instead of a month.

Getting started: the signup and the bonus code

If you want to follow along, sign up for Magica and enter the code GEYBMDC at signup. It’s my referral code, and at the time of writing it adds 10M bonus credits to a new account, which is plenty to run this whole four-stage method several times while you learn where your rerolls go. Offer terms change, so check the number on the signup screen before counting on it. For transparency: I earn a referral credit when someone uses the code. I recommend Magica because it’s what my daily video actually runs on, not the other way around.

For a first project, make three shots. A three-shot video teaches you the whole loop (shot list, still, clip, edit, export) without burning a balance on a thirty-shot plan. When the three shots cut together cleanly, you’re ready to scale up. If you want to see how a single video fits into a bigger content system, the end-to-end AI content pipeline shows the wider machine.

Want to see the real files behind this? The prompt templates, style blocks and config my fleet runs on are in the fleet files. They’re the actual working files, not a cleaned-up demo, and an email address gets you in.

If you’d rather have a video pipeline that makes the shot list, generates, cuts and publishes on its own, we can build that for you.