Fix Malfoid AI Faces, Eyes & Frame Drift

Why it happens

Image models build a face from many small details. When the face is small in the frame, when the style mixes realistic and cartoon cues, or when the prompt asks for too many things at once, the model runs short of detail for the eyes and mouth first. Mismatched eye colours, lopsided pupils and a smeared mouth are the most common results. Painterly styles are more sensitive than anime, because the model is trying to add realistic texture without becoming a photo.

How to fix it

1. Use the Portrait or Avatar layout so the face fills more of the frame.
2. Pick one art style and do not mix style words in the description.
3. Name the eye colour once, for example “pale grey-green eyes”, and never list two colours.
4. Choose 2K when the face is small, such as in full-body or comic layouts.
5. Keep the description to hair, eyes, outfit and one action, and move everything else to a later image.

Why it happens

A video model invents every new frame from the ones before it. Fast movement, a turned head or a small face gives it room to reinterpret her features, so by the last second she can look like someone else. Prompts that ask for several actions make it worse, because the model rushes between poses and redraws the face at each one. Longer clips drift more simply because there are more frames to get wrong.

How to fix it

1. Start from a close, front-facing image with even lighting.
2. Ask for one action per clip, such as a smirk or an eye roll, not a sequence of moves.
3. Use the slow push-in or static camera instead of a pan or handheld move.
4. Choose 5 seconds when the action is short; extra seconds give the model time to drift.
5. Make the next clip from the same image and join the clips in your editor.

Why it happens

Hands are the hardest part of any image, and two characters touching doubles the problem: the model has to decide which fingers belong to whom. Dancing, holding hands and hugging are where fingers fuse, extra hands appear or one arm vanishes behind the other person. Video makes it worse, because the hands move and each frame has to solve the puzzle again.

How to fix it

1. Leave a little space between the two characters in the starting image.
2. Describe poses that keep hands visible but apart, such as one on a book and one on a railing.
3. For dance clips, prefer a side-by-side pose and a short action.
4. In Duo images, ask for the two of you side by side rather than embracing.
5. If a clip still fuses hands, cut just before the contact or crop in your editor.

Why it happens

TikTok and Reels re-compress every upload. A 720p clip that the app scales up, an export at a different aspect ratio, or a file that has already been compressed once loses detail in fine lines, which is exactly where anime hair and eyes live. Screen recordings and re-saved downloads lose even more.

How to fix it

1. Generate at 1080p (Seedance 2.0 or 2.5) when the clip is your final post.
2. Edit on a 1080 × 1920 (9:16) timeline so nothing is scaled or letterboxed.
3. Export at a high bitrate and upload the original export, not a re-saved copy.
4. Turn on high-quality uploads in the app's settings where it is offered.
5. Skip sharpening or blur filters after export; they exaggerate compression artefacts.

Why it happens

Every new prompt starts from nothing, so the model re-imagines the face each time. Even with identical words, hair length, face shape and outfit details change from run to run. One image looks right; the next looks like her cousin. That breaks a comic, a series or a profile picture you want to keep using.

How to fix it

1. Generate one image you like and press Use as reference to make the next one from it.
2. Change only one thing per run: the pose, the scene or the outfit.
3. Keep a short list of the details you repeat in every description.
4. For comics, use the 4-panel layout so all panels are drawn together.
5. Save your best image and start every new scene from it.

Most broken results come from prompts that ask for too much at once. A short, ordered prompt gives the model fewer chances to improvise: who she is, what she wears, one expression, one action, one place and one style, in that order. Anything else can wait for the next image.

Malfoid, long straight platinum-blonde hair, pale grey-green eyes · black robe, grey vest, emerald tie · smug smirk · leaning on a stone pillar · candle-lit corridor · anime illustration

The prompt library uses this order for every prompt, so loading one of those is the quickest way to start from a clean base and change one part at a time.

Some results are not worth rescuing. If the face is wrong in the starting image, every clip and every reference image built on it will carry the same mistake, and each new run will copy it more faithfully. Start over with a fresh image when the eyes or face shape are wrong, when the outfit has the wrong colours, or when the character simply does not read as Malfoid. Fix and reuse an image when only the pose, the background or the expression is off.

Failed generations return their credits automatically, but a finished image you do not like still counts as a generation, so a minute spent on the prompt usually saves credits.

• One art style, one eye colour, one action.
• The face fills at least a third of the frame.
• For video: a close starting image, 5 seconds, slow push-in.
• For a series: start from your best image with Use as reference.