Hey, it's Cindy ๐ฑ You commented AISTUDIO, so here is the whole thing: the four-stage workflow, where to run each step and what it costs.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
Read this before you start, it saves you an evening. This does not retouch your footage, it regenerates it. A video model reads your original clip for motion and framing, then paints a new scene over the top. That means your face, your clothes and your hands are all being redrawn, so identity drift is the thing you are fighting the whole way through. It works beautifully on b-roll, cutaways, wide shots and short 5 to 15 second inserts. It is much harder on a long piece to camera with lip sync, so start with a cutaway.
the mechanic
How it actually works
Four stages, and the order matters. You film normally. You pull one still frame out of that clip and restyle only that frame, which is fast, cheap and easy to iterate on. That restyled frame becomes the look reference. Then a video model takes your original clip plus the restyled frame and re-renders the footage inside the new scene, keeping the motion and camera move from your real video.
The reason you restyle a single frame instead of prompting the video model directly: a still costs a few cents and takes seconds, so you can try eight versions of the lighting before you commit. Video generations cost real money per second, so you want the look locked before you ever press generate on a clip.
do it
The six steps
๐ฅ Film it like normal, but film it clean. Any room, phone on a tripod or propped on books. Two things genuinely matter: stay in one position rather than drifting around the frame, and get one light source on your face (a window works). Shoot 10 to 15 seconds, because that is the ceiling most video models will regenerate in one pass. Leave a little headroom, since re-rendering can shift the crop slightly.
๐ผ๏ธ Pull one frame out. Scrub to a moment where you are looking roughly at the lens and your hands are not mid-blur, then export that single frame. On a Mac, open the clip in QuickTime, pause, and use Edit then Copy, or take a full-quality screenshot. On a phone, pause and screenshot. Save it at the highest resolution you can, because the video model will read detail from it.
๐ง Have Claude write the restyle brief. This is the step everyone skips and it is the one that decides the result. Upload the frame to Claude and run prompt one below. Claude reads your actual frame, names what is fighting the studio look, and writes you a proper image-editing prompt with camera, light, set and wardrobe direction in it.
๐จ Restyle that single frame. Open Reve, upload your frame, switch to editing, and paste the prompt Claude wrote. Generate three or four versions changing only the lighting between them. Pick the one where you still look like you. That is the only acceptance test at this stage, because everything downstream inherits your face from this image.
๐๏ธ Regenerate the video against that frame. Open your Seedance route (options below), upload your original clip as the video reference and the restyled still as the image reference, set the duration to match your clip, and paste prompt three. The model keeps your motion and camera move and rebuilds the scene around you. Generate twice at low resolution to check the look holds, then once at 1080p for the keeper.
โ๏ธ Cut it back in. Drop the regenerated clip into your edit as a cutaway rather than the whole video. The strongest version of this trick is a normal talking head that cuts to two or three regenerated inserts. Full regeneration of a long piece to camera is where it starts to look uncanny, and where the cost stops making sense.
the prompt pack
Prompt 1: read my frame, write the brief
Upload your exported frame to Claude first, then paste this. It works in any Claude, no Claude Code needed.
๐ฉโ๐ป diagnose the frame
I have attached one frame from a talking-head video I filmed in a normal room. I want to restyle this frame so the finished clip reads as if it were shot on a real set, and then use it as the look reference for regenerating the video.
First, look at the frame properly and tell me what is actually giving away that this was filmed at home: the direction and quality of the light, the colour temperature, what is behind me and how far away it is, the depth of field, the camera height relative to my eyeline, and anything in the wardrobe or the background that a set designer would remove. Be specific and blunt, name the objects.
Then tell me which of those I can fix in an image edit without changing my face, and which ones would require reshooting. I want the honest split, not a list of everything you could theoretically change.
Do not write the restyle prompt yet. Just give me the diagnosis and the split.
the prompt pack
Prompt 2: the studio restyle prompt
Run this straight after prompt one, in the same conversation, so Claude is writing against your real frame rather than a generic room.
๐ก write the image restyle prompt
Now write me the image-editing prompt for this frame. I am pasting it into an image model that edits an uploaded photo, so write it as a single instruction paragraph addressed to that model, not as advice to me.
Direct it like a cinematographer would. Specify: the key light (direction, height, hardness and colour temperature in kelvin), a fill and a practical light if the scene needs one, the background (what it is, how far behind me, and how out of focus it should read), the lens and aperture look you want the depth of field to imitate, the camera height and angle relative to my eyeline, and two or three set-dressing objects that suit the mood of [describe your content in one line].
Then add an explicit preservation clause at the end of the prompt: keep my face, hair, skin tone, body position, head angle, gaze direction and expression exactly as they are in the source image, keep the framing and crop identical, and change only the lighting, background and set. State that no elements should be added between me and the camera.
The mood I want is [warm and lived-in / clean and editorial / moody and cinematic]. Give me the prompt as one block I can copy, then three one-line variants that change only the lighting setup so I can generate all four and compare.
The preservation clause is doing most of the work. Without it, image models quietly "improve" your face, and by the time it reaches the video stage you are looking at a stranger. If version one already drifted, re-run with the clause moved to the front of the prompt, models weight the opening more heavily.
the prompt pack
Prompt 3: the video regeneration prompt
Once you have your restyled still, this is the prompt you paste into the video model alongside your original clip and that image.
๐ฌ write the regeneration prompt
Now write the prompt for the video model. I am giving it two references: my original video clip, and the restyled still we just made. Write it as one instruction paragraph addressed to the video model.
The prompt must say clearly which reference governs what. The video reference governs motion, timing, camera movement, body position and performance, and none of that should change. The image reference governs the environment, lighting, colour grade and set, and the finished clip should look like the still moving. Say that explicitly, in that order, because ambiguity here is what causes drift.
Include an identity lock: the subject's face, hair, skin tone and wardrobe must match the image reference exactly for every frame, with no morphing, no changes in facial structure between frames, and no extra people or hands entering the shot.
Describe the scene in the restyled still in about twenty words so the model has a text anchor as well as a visual one, and specify the grade in plain terms, for example warm highlights with soft shadows and no crushed blacks.
Finish with a short negative list of what must not appear: text or watermarks, warped hands, changing background objects, flickering light, and any change to the camera move.
the prompt pack
Prompt 4: fix it when it drifts
Your first generation will be about eighty percent right. Rather than guessing at what to change, hand the result back to Claude.
๐ง diagnose the bad take
Here is a frame from the clip the video model produced, and here is the restyled still I gave it as the reference. Compare them and tell me exactly where the generation went wrong.
Go through it in this order: identity (did my face, hair or wardrobe change), motion (did the camera move or my body position change from the original), environment (did the background objects shift or the light change mid-clip), and artefacts (hands, edges, text, flicker).
For each problem you find, tell me whether the fix belongs in the image reference, the prompt wording, or the reference weighting, and give me the exact revised line to change. Do not rewrite the whole prompt, give me the specific edits so I can keep everything that already worked.
where to run it
The tools, and what they cost
The image step and the video step are separate, and you can mix and match. Here is the honest version of each.
Reve, for the frame restyle Editing an uploaded image with a written instruction is exactly what it is built for, and edits are cheap enough to iterate on properly. There is a free tier plus paid plans. Open Reve.
Higgsfield, for Seedance The route I use for the video step, since Seedance is in the model list and the references upload in one place. Open Higgsfield.
fal, if you want per-second pricing Seedance 2.0 reference-to-video takes up to 9 images, 3 videos and 3 audio clips, 4 to 15 seconds, 480p to 1080p. With a video input it prices at roughly $0.18 per second, so a 10 second clip lands near $1.80. The model page has the current rates.
Dreamina, the free-credits route Seedance video-to-video runs here on daily free credits (225 a day is the number they publish). Worth knowing: the free consumer tier is for private, non-commercial use under their terms, so anything you actually post needs a paid plan. The tool page.
The reel showed this running inside Lovart, which bundles the image canvas and the video generator on one page. That works and the steps are identical. I split it because doing the image step in a dedicated editor gives you far more control over the one thing that decides everything, which is whether you still look like you.
craft
6 things that make it read as a set
Motivate the light Studio light comes from somewhere. Put a window, a lamp or a softbox in the prompt and say which side it is on, so the shadows have a reason to fall the way they do.
Push the background back The home giveaway is a wall right behind your head. Ask for depth: something six to eight feet back, softly out of focus, with a little separation light on your shoulders.
Pick a colour temperature Say the number. Around 3200K reads warm and lived-in, around 5600K reads clean and daylit. Mixed temperatures are what make a room look accidental.
Dress the set, lightly Two or three objects that belong to the story you are telling. A full shelf of props reads as clutter and gives the model more to hallucinate.
Keep the wardrobe from the source Let the model restyle the room, not your clothes. Wardrobe changes are the fastest way to lose your own likeness between frames.
Regrade the whole edit to match If one insert is warm and cinematic and the rest of your video is phone-flat, the insert reads as fake. Match the grade across every clip in the timeline.
good to know
Two honest limits
Lip sync is the hard part. Regenerating footage where you are speaking to camera often softens or smears the mouth, which is why the cutaway approach wins. If you do need a speaking shot, keep it under about 6 seconds, generate several takes, and choose on the mouth rather than on the pretty background.
Check the terms of whichever platform you land on before you put a regenerated clip in something you are paid for. Free consumer tiers are commonly personal-use only, and that is a licence question rather than a technical one, so the model quality has nothing to do with it.