--- title: "Consistent character AI video: one face in every clip" canonical: https://withicarus.com/blog/consistent-character-ai-video site: Icarus description: "Why a face drifts between AI video clips, and how to keep one consistent character from first clip to last: one photo, a saved character, the same words." --- # Consistent character AI video: one face in every clip Why a face drifts between AI video clips, and how to keep one consistent character from first clip to last: one photo, a saved character, the same words. _October 11, 2026. Written with AI from the sources listed below, and checked automatically before it went live._ ![Young woman in a camel wool coat looking back over her shoulder, holding the collar](https://icarus.b-cdn.net/blog/consistent-character-ai-video/woman-camel-coat-over-shoulder.png) _Made with Icarus_ To keep the same character across AI video clips, start every clip from the same sharp, front-facing photo of one face. Save that face once and reuse it, and keep the light and scene wording identical from clip to clip. Most drift happens when one of those three things changes between clips. This post explains why a face changes from one clip to the next. It then walks you through the steps that keep one person steady. You also get a quick check for spotting drift and a simple rule for when to re-make a clip instead of trying to fix it. **Key takeaways** - Use one sharp, front-facing headshot as the only face reference. - Save that face as a character and tag it in every shot. - Make stills first, check the face, then turn the good ones into video. - Copy the same light and scene line into every prompt. - If the face is wrong, re-make the clip from the still. Don't patch it. ## Why does the face change from one clip to the next? An AI video model doesn't remember the last clip you made. Each clip is a fresh attempt at your words, so "a woman with short dark hair" can come out as a different woman every time you press the button. Starting from a picture fixes most of this. When you turn an image into a video, [Google's Veo uses the input image as the first frame](https://ai.google.dev/gemini-api/docs/veo), so the clip opens on the face you gave it. When you give it reference images instead, [Google's docs say Veo keeps the subject's look in the output video](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/use-reference-images-to-guide-video-generation). Drift creeps in wherever the model has to invent something it never saw. A head turn shows a side of the face that wasn't in your photo. A new room brings new light, and new light changes skin tone. A big move blurs the features for a moment, and they come back slightly different. Look at these first: - Eye shape and the distance between the eyes - Jawline and chin - Hairline, parting and hair length - Skin tone and freckles or marks - Apparent age The fix is to give the model less to invent. That starts with the photo. ## Start from one sharp, front-facing photo Your face photo does more for consistency than anything else you'll do. In Icarus you need one sharp, front-facing headshot. More angles don't help here: several angles of the same face make the results less consistent, not more. ![Front-facing headshot of a woman with short dark hair in soft, even window light](https://icarus.b-cdn.net/blog/consistent-character-ai-video/front-facing-headshot-window-light.png) _Made with Icarus_ Other tools ask for different inputs. [Google's Vertex AI docs](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/use-reference-images-to-guide-video-generation) describe giving as many as three images of a single person, character or product. Follow whatever the tool you use asks for, and keep that reference the same for every clip in the set. What makes a good headshot: - The face looks straight at the camera and fills most of the frame. - Soft, even light, with no hard shadow across one side. - Sharp focus on the eyes. - No sunglasses, hat brim or hair over the face. - A calm, neutral expression. - No beauty filter. The filter becomes part of the face you're asking for. Take it once, check it at full size, and don't swap it halfway through a project. ## Save the character, then make stills before video Here's the order that holds one face from the first clip to the last. 1. Upload your headshot and save it as a character. [Face locking](https://withicarus.com/features/same-face) then keeps that person across scenes, poses, outfits and lighting. 2. Make the stills. Tag the character in each shot, add your product if you sell one, and describe the scene in plain words. 3. Put the stills side by side and keep only the ones that clearly show the same person. 4. Turn each kept still into a short clip with [photo to video](https://withicarus.com/features/photo-to-video). The model comes with it. Why stills first? A still shows you the face before you pay for motion. An image at 1K costs $0.03 when it holds a saved product or character, so checking the face there is the cheap step. A 10-second clip costs $1.19 at 480p, so you want the face right before you get there. If you shoot a product too, save it the same way. Product pinning keeps its real colour, logo and stitching, so the person and the thing they're holding both stay put. ## Use the same light and scene words in every prompt Change the wording and you change the face. "Golden hour" in one prompt and "warm evening light" in the next can be enough to shift skin tone and the shape of shadows under the eyes. ![Woman in a camel wool coat walking toward the camera on a quiet street in overcast light](https://icarus.b-cdn.net/blog/consistent-character-ai-video/woman-coat-walking-city-street.png) _Made with Icarus_ Write one scene line and paste it, word for word, into every shot in the set. Only change the action. > The same woman in a camel wool coat walks slowly toward the camera on a quiet city street, soft overcast daylight, muted colours, eye-level camera Try this prompt in Icarus: https://withicarus.com/dashboard?prompt=The%20same%20woman%20in%20a%20camel%20wool%20coat%20walks%20slowly%20toward%20the%20camera%20on%20a%20quiet%20city%20street%2C%20soft%20overcast%20daylight%2C%20muted%20colours%2C%20eye-level%20camera For the next clip, keep everything but the action: she stops and looks into a shop window, same coat, same street, same overcast light, same eye-level camera. If you shoot in one setting often, save it as a [studio](https://withicarus.com/features/brand-studios): the light, background, mood and framing in one place, so every shot starts from the same scene. A few habits help in the clip itself: - Keep the face toward the camera. Profile turns are where drift starts. - Choose small camera moves. On video models that offer camera control, a slow push in or hold still keeps the face in view. - Keep clips short, and tell a longer story by cutting several short clips together. - Keep the outfit and hair the same unless the change is the point of the clip. Some tools can lengthen a clip for you. Google's Flow, for example, has [an Extend feature that continues the action from your original clip](https://blog.google/technology/ai/veo-updates-flow). Whatever you use, check the face again at the end of the longer clip. ## Spot drift, and know when to start over Watch each clip twice: once at normal speed, once paused on the first frame, the middle and the last frame. Hold those frames next to your original headshot. - [ ] Same eye shape and spacing as the headshot - [ ] Same jawline and chin - [ ] Same hairline, parting and length - [ ] Skin tone matches the other clips in the set - [ ] The person looks the same age from start to end - [ ] Any product in shot keeps its colour and logo Then follow one rule: fix the still, not the clip. - If the face is wrong in the first frame, the problem is the still. Make a new still, check it, then make the clip again. - If the clip starts right and drifts partway through, make the clip again from the same still with a smaller move and a shorter length. - If it drifts again, change one thing at a time: the camera move, then the action. Never the headshot or the scene line. Re-making beats patching because a patched clip often shows a seam where the face changes. Re-making costs little, too: six clips at $1.19 come to $7.14 at 480p. And if a shot fails outright, the money goes back to your balance by itself. Every price is on the [pricing page](https://withicarus.com/pricing), and you see each shot's price before you make it. ## Questions ### Why does my AI character look different in every video? Each clip is made fresh, and the model doesn't remember the last one. If you start from text alone, or change the photo, light or scene words between clips, it fills the gaps differently each time. Start every clip from the same face photo and the same scene wording. ### How many photos of a face do I need for a consistent character? In Icarus, one sharp, front-facing headshot. Several angles of the same face make the results less consistent. Other tools may ask for something different, so follow what yours asks for. ### Should I make images first or go straight to video? Make stills first. You can check the face cheaply on a still, keep only the ones that look like the same person, then turn those into clips. That way you don't pay for motion on a face that's already wrong. ### Can I fix a face that changes halfway through a clip? It's usually quicker and cleaner to make the clip again from the same still, with a smaller camera move and a shorter length. Patched clips often show where the face changed. ### Can I use the videos in ads? Yes. Every image and video you make in Icarus is yours for commercial use, with no licensing fees and no attribution. ## Sources - [Veo 3.1 and Flow updates (Google blog)](https://blog.google/technology/ai/veo-updates-flow) - [Generate videos with Veo (Gemini API docs)](https://ai.google.dev/gemini-api/docs/veo) - [Generate videos from references (Vertex AI docs)](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/use-reference-images-to-guide-video-generation) --- Site: Icarus (https://withicarus.com) Canonical: https://withicarus.com/blog/consistent-character-ai-video Contact: tanmay@onai.studio