To fix flicker, morphing and drifting faces in an AI video, ask the tool to change less: one simple action, a slower or still camera, a shorter clip, and steady light. Then give it something fixed to hold on to, such as a starting image.
A reference picture of the face, or a start and end frame, works the same way. In this site’s reading, most of these glitches come from asking for more change than the model can keep consistent.
- Ask for less motion. One action, one slow camera move or none.
- Shorter is steadier. Build long scenes from short clips.
- Give it an anchor: a starting image, a reference picture, or a start and end frame.
- Describe what stays the same, rather than writing “no morphing”.
Think of a flip-book where every page is redrawn. The more that changes from page to page, the more chances to get a hand, a face or a wall slightly different. Simpler action, and a photo of the character pinned to the desk, mean fewer chances to slip.
Which fix for which glitch?
The plain point: find the glitch you can see and try its first fix. The causes are this site’s reading unless a section below quotes a vendor.
| You see | Likely cause | First fix |
|---|---|---|
| Flicker | Cause: light or texture redrawn slightly differently each frame | Fix: one steady light source, still camera, shorter clip |
| Morphing | Cause: too much changing at once | Fix: one action, slower move, a starting image |
| Face drifting | Cause: nothing fixes what the face looks like | Fix: reference image, same description every time |
| Extra limbs mid-clip | Cause: complex body movement, bodies overlapping | Fix: one person, simple pose, whole body in frame |
| Warping hands | Cause: busy finger movement | Fix: one simple hand action, or frame the hands out |
| Background swimming | Cause: long clip or big camera move | Fix: still or slow camera, start and end frame |
Flicker: the light or texture pulses
The plain point: brightness, colour or fine texture jumps from frame to frame. Ask for steady light and a still camera.
Why it happens: in this site’s reading, the model redraws fine detail in every frame, and busy texture, changing light words or a moving camera give it more to redraw differently. ByteDance’s Seedance guide lists light among the things to hold fixed: “keep the same face, age, hairstyle, wardrobe, product geometry, and lighting direction across every shot.”1
The fix: name one steady light source, and leave out changing light (candles, strobes, headlights) unless you want it. Test with a still camera.
Morphing: things melt or change shape
The plain point: a cup turns into a bowl, a logo smears. Ask for less change, and start from a clear picture.
Why it happens: this site’s reading is that the model is inventing more motion than it can keep consistent. Vendors warn about overloaded prompts too. Runway says over-specifying “can paradoxically lead to unexpected or unnatural results”, and Kling says “it is challenging to generate complex physical movements”.2
The fix: cut to one action and slow it down. Start from an image of the object, and on Seedance add a short “Preserve” line, as its template does: “Preserve [identity, wardrobe, logo, object geometry]”.3
Faces drifting: the person slowly becomes someone else
The plain point: the face, hair or clothes change. Show the tool the face, and describe it the same way every time.
Why it happens: in this site’s reading, words alone leave room for a slightly different face in every frame. Google says Veo 3.1 accepts up to three reference images to “preserve the subject’s appearance in the output video”.4 ByteDance gives the same advice for Seedance: “Pin identity early” with two or three portrait stills tagged in the prompt.1 On Runway, the References feature works on still images, which you then animate; Runway says it keeps characters consistent “from just a single reference image”.5
The fix: add a clear, front-facing reference where your tool allows it. Without one, write a fixed description and paste it unchanged into every prompt, as Google advises: “Copy and paste the entire, unchanged character description into your prompt for every new scene or action.”6 Templates are on identity lock prompt templates.
Extra limbs that appear mid-clip
The plain point: a third arm swings in, or a leg vanishes and comes back doubled. Ask for one person, a simple movement, the whole body in view.
Why it happens: in this site’s reading, limbs go wrong most when they cross, overlap or leave the frame, because the model has to guess what is hidden. Kling notes that video models are “not sensitive to numbers”, which makes it hard “to maintain consistency in counts”.7 Applying that to arms and legs is this site’s reading, not Kling’s. When Kling copies a movement from a video with its Motion Control, it asks for a single character with the “entire body and head clearly visible and not obstructed.”8
The fix: one person, nothing in front of the body, and a slow, simple action. Leave spins and fights until that works.
Warping hands
The plain point: fingers bend, merge or change number as the hand moves. Give the hands one simple, visible job, or keep them out of the shot.
Why it happens: this site’s reading is that hands are small, have many parts and move fast, so every frame is a fresh chance to draw a finger wrong.
The fix: one slow hand action and what it touches, or frame from the chest up. The hands section of negative prompts by problem has lists for tools that read them.
Background swimming
The plain point: walls ripple and trees wobble as if under water. Hold the camera still, keep the clip short, and fix the start and end.
Why it happens: in this site’s reading, a moving camera makes the model redraw the whole background in every frame, and the longer the clip, the further it can wander. ByteDance makes the same point about long clips: “Start/end frames: give the model both so the full 30 seconds doesn’t wander off compositionally.”1
The first two sentences are Runway’s own rewrite of “No camera movement”.9
Four bigger fixes that help with any glitch
The plain point: when a line does not work, try these bigger changes, in this order, one at a time.
- Shorter, simpler clips. Google says a short video should be “a single, focused moment”; Kling suggests subject movement “suitable for a 5-second video”.6 Join short clips afterwards.
- Image-to-video instead of text-to-video. A starting image fixes the look of the first frame. Google: “Think of your source image as the first frame of your film.” Runway asks for an input image “high quality and free of visual artifacts”.6 See how to turn a photo into a video.
- First and last frame. Give both the start and the end picture, and the model only has to invent the route. Google calls this interpolation for Veo 3.1.10 See first frame and last frame.
- A lower motion setting. Where a tool has one, use it while you fix a glitch. Midjourney’s default Low Motion “is more likely to result in still scenes, low camera motion, slow motion, or subtle character movements.”11
Change one layer at a time. ByteDance’s own advice: “When a generation misses, change one control layer at a time.”1
Less change, more anchors. Shorten the clip, simplify the action, then pin the look with a starting image, a reference or a start and end frame.
Questions people ask
Does writing “no morphing” or “no flicker” help?
Not on Runway, which says negative phrasing may produce the opposite result (source above); describe what stays the same instead. Tools with a negative field can take a short list; the video section of negative prompts by problem says which.
Why does my AI video get worse towards the end?
In this site’s reading, every extra second is more frames the model must keep consistent, so drift adds up. ByteDance tells you to give both start and end frames on long clips so they do not “wander off compositionally” (source above). Chain short clips instead.
Where to go next
- “Something else went wrong.” Read why your generation came out wrong.
- “I need lists for my tool’s negative field.” See negative prompts by problem.
- “Start again with a clean prompt.” Follow how to write your first AI video prompt and camera words for AI video.
- “What do these words mean?” See image-to-video, keyframe, negative prompt and seed.
- “Take me through it step by step.” Follow make your first AI video.
- Related: How to Upscale AI Video to 4K.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does the page give as the main idea behind most fixes for flicker and morphing?
- Ask for less change and give the tool something fixed to hold on to
- Ask for more motion so the glitches blur away
- Add a long list of words starting with no
Show the answer
Answer: Ask for less change and give the tool something fixed to hold on to. The page says to ask for less change and then anchor the clip with a starting image, a reference or a start and end frame.
2. What is the page’s first fix for a face that drifts during a clip?
- Switch to a wider lens
- Use a reference image and describe the person the same way every time
- Make the clip longer so the face settles
Show the answer
Answer: Use a reference image and describe the person the same way every time. The page says to add a reference image where the tool allows it, or paste the same fixed description into every prompt.
3. On Runway, why does the page suggest writing what stays the same instead of no morphing?
- Runway shows negative words as on-screen text
- Runway only reads the first word of a prompt
- Runway says negative phrasing may produce the opposite result
Show the answer
Answer: Runway says negative phrasing may produce the opposite result. The page quotes Runway saying negative phrasing is not supported and may produce the opposite result.
- 1Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 13 Oct 2026“For continuity: keep the same face, age, hairstyle, wardrobe, product geometry, and lighting direction across every shot.” ↑a“Pin identity early: attach two or three portrait stills for each lead and tag them in the prompt (@Image1, @Image2) to keep faces stable across shots.” ↑b“Start/end frames: give the model both so the full 30 seconds doesn’t wander off compositionally.” ↑c“When a generation misses, change one control layer at a time.” ↑d
- 2Runway Academy, Prompting Guide, undated, read at source 13 Oct 2026“This over-specification can paradoxically lead to unexpected or unnatural results”. Kling AI, Text-to-video prompt guide, 24 Nov 2025, read at source 13 Oct 2026: “At the current stage, it is challenging to generate complex physical movements”, with a bouncing ball and a high throw as examples; “Keep the visual content as simple as possible”. ↑
- 3Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 16 Sep 2026 (as cited on this site’s Seedance cheat sheet): template clause “Preserve [identity, wardrobe, logo, object geometry].”↑
- 4Google AI for Developers, Generate videos with Veo 3.1 in Gemini API, read at source 13 Oct 2026“Veo 3.1 now accepts up to 3 reference images to guide your generated video’s content.” “Provide images of a person, character, or product to preserve the subject’s appearance in the output video.” ↑
- 5Runway, Creating with Gen-4 Image References, re-read directly at source 13 Oct 2026 (also cited on this site’s character consistency page): “Gen-4 References excels at generating consistent characters across different lighting conditions, locations, and treatments, all from just a single reference image.”↑
- 6Google Cloud, Best practices for Veo on Vertex AI, last updated 9 Oct 2026, read at source 13 Oct 2026“Copy and paste the entire, unchanged character description into your prompt for every new scene or action.” ↑a“For short videos, dedicate each prompt to a single, focused moment.” Kling AI, Text-to-video prompt guide, 24 Nov 2025, read at source 13 Oct 2026: subject movement should be “straightforward and suitable for a 5-second video”. ↑b“When using the image-to-video feature, the quality of your source image is important.” “Think of your source image as the first frame of your film.” Runway Academy, Prompting Guide, undated, read at source 13 Oct 2026: “For best results, ensure that the input image is high quality and free of visual artifacts.” ↑c
- 7Kling AI, Text-to-video prompt guide, 24 Nov 2025, read at source 13 Oct 2026“Current large video models are not sensitive to numbers, making it difficult to maintain consistency in counts”. ↑
- 8Kling AI, Motion Control user guide, 5 Mar 2026, read at source 12 Oct 2026 (as cited on this site’s Kling cheat sheet): “entire body and head clearly visible and not obstructed.”↑
- 9Runway, Gen-4 Video Prompting Guide, re-read directly at source 13 Oct 2026 (also cited on this site’s negative prompts page): “Negative phrasing is not supported and may produce unpredictable or even opposite results. ❌ No camera movement. The camera doesn’t move. NO MOVEMENT ✅ Locked camera. The camera remains still.”↑
- 10Google AI for Developers, Generate videos with Veo 3.1 in Gemini API, read at source 12 Oct 2026 (as cited on this site’s first frame and last frame page): “Veo 3.1 lets you create videos using interpolation, or specifying the first and last frames of the video.”↑
- 11Midjourney, Video, re-read directly at source 13 Oct 2026 (also cited on this site’s camera grammar page): “Low Motion (default): is more likely to result in still scenes, low camera motion, slow motion, or subtle character movements.”↑
Read at source 13 Oct 2026: Dreamina (ByteDance), Seedance 2.5 prompt guide (18 Aug 2026); Kling AI, Text-to-video prompt guide (24 Nov 2025); Google AI for Developers, Generate videos with Veo 3.1 in Gemini API; Google Cloud, Best practices for Veo on Vertex AI (last updated 9 Oct 2026); Runway Academy, Prompting Guide.
Reused from this site’s detail pages, with the read dates recorded there: Runway’s Gen-4 Video Prompting Guide (negative prompts page, 12 Oct 2026) and Gen-4 Image References (character consistency, 12 Oct 2026), and Midjourney’s Video page (camera grammar, 12 Oct 2026), all three first read in a browser and re-read directly at source on 13 Oct 2026, when Runway’s and Midjourney’s help centres loaded without one; Kling’s Motion Control guide (Kling cheat sheet, 12 Oct 2026); the Seedance template (Seedance cheat sheet, 16 Sep 2026); Google’s interpolation line (first frame and last frame, 12 Oct 2026).
The causes marked “this site’s reading”, the fix table, the flip-book comparison and the prompts to paste are this site’s own, not vendor claims and not measured tests. No vendor document read for this page explains why flicker happens.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK