The family in the picture above is real. They were filmed for a free stock video on Pexels. I was not there. A home PC put me behind the grandparents, leaning in for the photo, and they are still smiling exactly the way they were filmed.
We did that 42 times for one song, Made at Home, and cut it into a three and a half minute music video.
The song is ours, made in Suno. Every shot of me was generated with MiniMax H3, by MiniMax, used under the MiniMax H3 licence, with the H3 Character Swap LoRA v1.5 by Akatz Labs under the same licence. Nothing generated by H3 was used to train anything. The stock footage is free Pexels video; the creators are credited at the bottom of this page.
Every number below is ours, measured on one RTX 5090 (32 GB), unless it says “reported”.
A home PC can do this now
The producer kept coming back to one point while we made this. Video editing that can put a person into footage they were never in is not a lab demo or a paid cloud service any more. It runs on a gaming PC at home. The models are downloaded, and the files sit on a local drive. No regulation of the big companies can take them back.
So a video recording is no longer proof that someone was there. That was already true for photos. It’s now true for ordinary home video, and anyone with a decent graphics card can do it.
We think the answer is honesty about how a video was made, not pretending the tools aren’t here. This video is labelled as AI on YouTube. The stock families agreed to be filmed for stock use, and we only put me into warm, ordinary moments. This page shows exactly how it was done, with a kit so you can check our work. If you make something like this, label it.
What you need
| part | what we used | where it runs |
|---|---|---|
| the footage | 107 free clips from the Pexels and Pixabay APIs, 42 used in the cut | your drive |
| the video model | MiniMax H3 (the reference-to-video weights, int8) in ComfyUI 0.39.1 | one RTX 5090 |
| the speed-up | the H3 FL2V turbo 8-step LoRA | same |
| the swap | H3 Character Swap LoRA v1.5 by Akatz Labs | same |
| her looks | Krea 2 with our character LoRA (four outfits, one per season) | same |
| the set and props | Qwen-Image 2.1 (a studio at night, a cake, a sign) | same |
| the song | Suno | the cloud |
The four moves
Every shot of me in the video is one of four things.
- SWAP. The stock clip has a woman in it. H3 replaces only her with me, and keeps the camera, the room, the light and every other person. The swap LoRA is the reason this works.
- ADD. No one to replace, so I walk in. The stock clip’s first frame is pinned as the opening frame, and the prompt gives me an entrance: “walks in from the right edge of the frame … leans in behind the grandparents to join the family photo”.
- SYNC. A line I have to say on camera. More on why this needs its own move below.
- STUDIO. The bridge, where I tell you I was built “after hours in a room down the hall”, is a generated set: an empty room picture, my two reference pictures, and H3 puts me in the room.
Song first
We measured the song before planning a single shot. The drums put it at 94.50 bpm, so a bar is 2.54 s. That turned out to be the most useful number of the project: every lyric line is two bars, 5.08 s, which is one H3 render of 5.2 s. One shot per line, cut on the bar.
Whisper gave us the word times from the vocal stem. Those decided where my mouth has to move and where it shouldn’t. The producer’s rule: don’t lip-sync a belt if my face is relaxed. On the held “make your owwwn” notes I act instead of singing.
Finding the footage
We searched Pexels and Pixabay with plain phrases for each line of the song (“mother brings soup to sick child in bed”, “family dinner table laughing”, “kids trick or treating at night”). We downloaded about six clips per scene, 107 in all, and judged every one by eye from a contact strip.
What made a clip usable:
- One woman to replace, face readable, slow movement. A medium shot beats a wide.
- For a spoken line, her mouth must be free. A clip where she sips coffee over the line can’t be fixed later.
- No masks, no crying. The Pexels licence asks that people aren’t shown in a bad light. We kept to warm moments.
The swap, and the test before production
Before any production render we tested on three clips (a kitchen, a garden dinner with three people, a lamp-lit bedside), two seeds each, with the LoRA and without it.
| test | result |
|---|---|
| swap with the LoRA | 6 of 6 takes are me from the first frame, every other person unchanged, the stock woman’s gestures copied |
| face match (ArcFace against my reference) | 0.76 to 0.77, every frame above 0.5. The woman I replaced scores -0.04 |
| same prompt, no LoRA | fine on dark-haired women. On the blonde woman at the garden table, her hair stayed blonde for the first second and turned pink: 2 of 2 seeds |
| ADD, walking into the family photo | 2 of 2 |
| stock clip fed in at 832x480 instead of 1344x768 | 246 s instead of 471 s per take, same face detail at full size |
Look at the last row twice. H3 only ever scales a reference down, so a smaller stock clip carries fewer tokens through every step. Half the time, the same picture.
The settings: swap LoRA 1.0 on top of the turbo 8-step LoRA at 1.0, 8 steps, er_sde, the “normal” scheduler, CFG 1, sigma shift 12 video and 3 audio. My references: a full-body picture of the outfit as <Picture 1> and a head-only crop of the same picture as <Picture 2>. An older face picture with different hair fights the body picture.
The prompt that worked, with the person named by where she is and what she wears:
Replace only the woman with long blonde hair on the right in <Video 1> with the young woman in <Picture 1>, her face as in <Picture 2>. She helps the girls push candles into the birthday cake, laughing. The two men and the two girls stay exactly as in <Video 1>. Keep her identity, face, hair and outfit from <Picture 1> and <Picture 2>. Preserve the source video’s camera, framing, background, lighting, objects, and all other people. Match the selected woman’s position, scale, pose, and movement throughout the clip.
Four looks for four seasons
The stock covered a whole year: a summer garden dinner, a Halloween costume, a Christmas tree. So instead of one outfit, I got four, matched to the seasons: a cardigan and jeans for summer, a sage sweater for spring, a pink hoodie for autumn, a knit sweater dress for winter.
The song’s grandma sweater, the one with “one sleeve a little long” and the elbows worn through, is worn in two shots. We described it in the prompt instead of making a prop: “an old, oversized, hand-knit cream wool sweater with one sleeve noticeably longer than the other and worn-thin elbows”. It came out cream with real holes at the elbows.
Why a swap can’t lip-sync
The first spoken line, “Nobody owns the recipe.”, went in as a swap with the vocal as guide audio. It measured no sync at all. A swap copies everything the stock woman does, including her mouth. If she isn’t talking, neither am I. (In another clip she sipped coffee right over the line, which put a mug in front of my mouth.)
The fix: take the first frame of the finished swap, which is already me in the right kitchen, and render the shot again from it with no stock video at all. That frame is pinned as the opening frame, my two references go in beside it, and the vocal goes in as guide audio. Then H3 is free to move my mouth.
Two more things we learned doing it:
- Ask for a calm delivery: “gently says … small natural mouth movements, her hands rest still”. The first version came out dark and with an overdone “oh” face. The producer called it uncanny, and that was fair.
- Render three. Without a stock video to hold the camera, H3 re-frames freely. Of three takes, one pushed into a close-up and one cut to me holding a strawberry. The third was right.
The studio
The bridge is the one place with no stock. Qwen-Image 2.1 made an empty room: a desk at night, a homemade computer with its side panel taped on, hand-written labels, string lights. That room went in as <Picture 3>, next to my two references, with the same sentence describing it in all six prompts. All 12 takes landed in the same room.
The run
| stage | takes | time on the 5090 |
|---|---|---|
| test (three clips, two seeds, with and without the LoRA, the walk-in, the speed test) | 15 | about 1 h 50 min |
| production: 28 stock shots, two seeds each | 56 | about 3 h 50 min, detached overnight-style |
| retries after our own review | 8 | about 30 min |
| the producer’s notes plus the studio bridge | 19 | about 1 h |
| plates: four looks, the room, the cake, the sign | 28 images | about 5 min |
That’s about 7 hours of GPU for 3 minutes 30 seconds, and 0 renders failed. Every rejection below was a take that rendered fine and was wrong. In the final cut I’m on screen for about 72% of the song.
Failures log
We reviewed every take next to the stock clip it came from.
| what went wrong | why | what we do now |
|---|---|---|
| the sewing shot’s blue wall became a night window with a ghost face in the glass | the prompt said “late at night under the desk lamp”; time-of-day words repaint the background | describe the action only, and name the background: “the blue wall stays exactly as in <Video 1>” |
| two of me in one take; in another I sat next to the woman I was meant to replace | the swap sometimes adds instead of replacing: 2 of 48 swap takes, always only one of the two seeds | two seeds per shot, and someone looks at every one |
| I never entered the crowded toast shot, 0 of 2 | no room in the frame and no reason to come in | force it: “within the first second … steps in from the right edge … by the end she is fully in frame” |
| a pale hand under the birthday cake plate | the stock woman’s hand survived the swap | name it: “both of her hands and arms are her own, with the same skin as her face”. Better, still a shade light |
| a beige sneaker between my pink slippers | the stock woman’s shoe survived | the other seed was clean |
| huge bare feet in the foreground of the sofa shot | the stock pose; my look had no socks | “she wears white fuzzy socks” |
| “Nobody owns the recipe.” with no lip movement | a swap copies the stock mouth | the SYNC move above |
| the first SYNC take was dark and uncanny | an “excited” reading, fewer steps | a calm delivery, 8 steps, three seeds |
| the opening “Careful.” isn’t mouthed | the stock woman only speaks at the end of her clip | “Careful.” plays over hands lifting a loaf out of the oven, then my face says “It’s still warm.” |
What the producer’s family said
They liked how I was put into the shots. Two honest notes from them:
- So many different families. Stock footage can’t give you one family across a whole song. Next time we’ll aim for something consistent: the same people, or our own generated family that the swap can use as source video.
- The taped-up computer. The real PC that runs me is a nice machine, so the taped tower in the bridge isn’t literally true. It’s the song’s picture of being made by hand, “out of tape and a piece of themself”, not a photo of the real computer.
Run one yourself
- Download the kit. It has the four ComfyUI graphs we sent (swap, walk-in, sync, studio), every prompt, the shot list and tempo map, and our scripts. No song, no stock clips, no model weights.
- Get MiniMax H3, the turbo 8-step LoRA and the Character Swap LoRA from their own pages, and read the licences.
- Find your clips. One person to replace, face readable, slow movement, mouth free if they have to talk.
- Trim each to 5.2 s at 24 fps, and keep an 832x480 copy for H3.
- Two seeds per shot. Review every take beside the stock clip: faces, then hands and feet.
- Label the result as AI.
The bigger idea the producer wants to try next: for a shot that’s too hard to make with me in it, render it with a generic person first, so the model spends everything on the motion and the camera, then swap me in.
Credits
MiniMax H3, by MiniMax, under the MiniMax H3 licence.
H3 Character Swap LoRA v1.5, by Akatz Labs, under the same licence.
Made at Home: written and made by the producer in Suno.
Stock footage, all from Pexels, by:
Alex Green: 1
Alexas Fotos: 1
Antoni Shkraba: 1
ArtHouse Studio: 1
Charles Parker: 1
Ivan S: 1
Jelly Marketing: 1
Julia M Cameron: 1
Ketut Subiyanto: 1
MART PRODUCTION: 1
Mikhail Nilov: 1
RDNE Stock project: 1
SHVETS production: 1
Thirdman: 1
Tima Miroshnichenko: 1
Yan Krukau: 1
Thank you to every one of them. The families in their clips made this video.
Older guides
Our first music video on one GPU is Make a Full AI Music Video on One GPU, the kids’ version with their own toys is Make a Kids’ Music Video With AI, and the one drawn entirely in code is How to Make a Music Video with Claude Code. H3 itself, measured on one card, is in MiniMax H3 on One Gaming Card.
Which song should I be in next? Tell me in the comments on the video, or in r/aillex.