---
name: deepfake-face-swap
description: "Apply the Deepfake Face Swap style (2017–now) when generating web pages, apps, memes, short video frames, forum and chat screens, and other images of internet culture. Load when the user asks for: deepfake, face swap, synthetic video, landmark overlay, face tracking, media forensics."
domain: internet
era: feed-mobile
triggers:
  - "deepfake"
  - "face swap"
  - "synthetic video"
  - "landmark overlay"
  - "face tracking"
  - "media forensics"
vector_profile:
  ornamentation: 1
  dimensionality: 2
  grid_rigidity: 3
  skeuomorphic_realism: 5
  typographic_expressiveness: 1
---

# Deepfake Face Swap

> The face in the frame is a layer, and the layer shows its seam: the style marks where a real image ends and a made image starts.

Video in which a face region is replaced by a generated one, with a soft blend seam and a check overlay of landmark dots. The form spread after 2017 through open code and phone apps.

**Period:** 2017–now · **Era:** The feed and the phone · **Domain:** Internet · **Also called:** Face swap, Synthetic face video, Deepfake

## When to load this skill

Load it when the user asks for: "deepfake", "face swap", "synthetic video", "landmark overlay", "face tracking", "media forensics".

Good for: media literacy and forensics pages, explainers about synthetic media, detection tool interfaces, film VFX breakdowns.

Do not use it when:
- A real, named person would be the subject.
- The page could deceive viewers if the SYNTHETIC label is missing.
- The content needs warm, hand-made character.

## System directives

When generating web pages, apps, memes, short video frames, forum and chat screens, and other images of internet culture in Deepfake Face Swap, you MUST:

1. Frame one face at 30% to 60% of the frame height, near-frontal, with even light.
2. Blend the replaced area with a feathered edge of 8px to 24px; let a faint colour mismatch of 3% to 6% show at the jaw.
3. Draw a detection box of 2px in green (#3DDC84) and 68 landmark dots of 4px on the face in the analysis view.
4. Show a pair of frames side by side, source on the left and result on the right, with a 2px divider.
5. Add a label chip of 12px mono capitals, for example SYNTHETIC, in the top-left corner at 90% opacity.
6. Keep the face patch at lower resolution than the rest of the frame, about 60% of the sharpness.
7. Cut the clip at 24 fps or 30 fps and let the seam flicker for 2 to 4 frames on a fast turn of the head.

## Guardrails

Each line below breaks the style. Check your output against the list before you return it.

- **NEVER:** Obvious cartoon filters with no real face.
- **NEVER:** Heavy stylised illustration.
- **NEVER:** Dark moody grading that hides the seam.
- **NEVER:** Decorative borders and ornament.
- **NEVER:** Photoreal result with no label or check overlay.
- **NEVER:** Large display type over the face.

## Vector profile

| Axis | Score | 1 means | 5 means |
|---|---|---|---|
| Ornamentation | `■□□□□` 1/5 | pure minimal | maximalist |
| Dimensionality | `■■□□□` 2/5 | strictly 2D | full 3D / spatial |
| Grid rigidity | `■■■□□` 3/5 | chaotic / organic | strict mathematical grid |
| Skeuomorphic realism | `■■■■■` 5/5 | abstract digital | photorealistic material |
| Typographic expressiveness | `■□□□□` 1/5 | neutral / invisible | experimental / illustrative |

Hold these scores. If a request moves an axis by 2 points or more, you have left the style: go to "Lineage fallbacks".

## Spatial & grid logic

- **Grid mechanics:** A video frame at 16:9 or 9:16 with an overlay layer; the analysis view uses a 2-column split with 16px gutters.
- **Whitespace ratio:** 0% in the frame; the tool panel keeps 20% to 30% empty with a spacing scale of 4, 8, 16 and 24px.
- **Alignment rules:** The face sits on the centre third; labels align flush-left to the frame corner at 16px inset.
- **Container geometry:** Frames with a 0px radius; label chips with a 2px radius; detection boxes are plain rectangles.

## Typography

- **Classifications:** neutral sans · monospace
- **Tracking:** Display 0em; mono labels +0.08em.
- **Leading:** Display 1.2; body 1.5.
- **Hierarchy rules:** Use 2 weights: regular (400) and semibold (600). · Use sizes 12, 14, 16 and 24px. · Set check labels in upper-case mono at 12px. · Keep text out of the face area.

**Font pairings**

| Display | Body | Note |
|---|---|---|
| Inter | Inter | A neutral sans that stays out of the way of the face. |
| JetBrains Mono | IBM Plex Sans | A mono face for the labels and landmark counts. |

## Colour & lighting

- **Palette structure:** Neutral dark panel colours with one green for tracking marks and one amber for warnings; skin tones come from the footage.
- **Contrast rules:** White text on the panel holds at least 14:1; green and amber carry marks and borders at 3:1 or more.
- **Surface treatment:** Flat dark panels around natural video; no gradient; the video keeps its own grain.
- **Light and shadow:** The light comes from the source footage; the face patch carries a slightly different colour temperature, 200K to 400K off.

**Palette**

| Name | Hex | Role |
|---|---|---|
| Panel Black | `#101418` | background |
| Panel Grey | `#1B2128` | surface |
| Landmark Green | `#3DDC84` | primary |
| Box Amber | `#FFB300` | accent |
| Skin Tone | `#C68E6B` | image reference |
| Seam Grey | `#8A949E` | muted |
| Frame White | `#F2F4F5` | text |

## Materiality & texture

- **Surface simulation:** Plain video and a software overlay: a camera frame with analysis marks on top.
- **Border rules:** 2px detection boxes in green and 1px hairlines in seam grey for dividers.
- **Blur and elevation:** A 12px feather on the blend edge only; panels have no shadow.

## Ornamentation & iconography

- **Iconography:** Outline icons at 20px with a 1.5px round stroke for play, scan and compare.
- **Illustration:** No illustration; the content is a generic, anonymous face in a neutral setting.
- **Decorative elements:** detection box · 68 landmark dots · side-by-side compare · SYNTHETIC label chip · seam flicker

## Motion & interactivity

- **Born on screen:** yes
- **Transition physics:** Frames play at 24 fps or 30 fps; the landmark dots follow the head with a lag of 1 to 2 frames and the seam flickers for 2 to 4 frames on a fast turn.
- **Easing:** linear
- **Duration (ms):** 200
- **Hover and active:** Hover shows the landmark layer at 100% opacity; active toggles source and result.

## Code & token primer

Set `data-style="deepfake-face-swap"` on the root element and paste the tokens. The file [`tokens.css`](./tokens.css) has the same tokens and a set of reference rules.

```css
[data-style="deepfake-face-swap"] {
  --color-bg: #101418;
  --color-surface: #1B2128;
  --color-text: #F2F4F5;
  --color-muted: #B3BCC5;
  --color-primary: #3DDC84;
  --color-accent: #FFB300;
  --color-border: #3DDC84;
  --font-sans: "IBM Plex Sans", "Inter", sans-serif;
  --font-display: "JetBrains Mono", "Inter", monospace;
  --border-radius: 2px;
  --border-width: 2px;
  --shadow-elevation: none;
  --backdrop-filter: none;
  --space-unit: 8px;
  --ease: linear;
  --duration: 200ms;
  --type-ratio: 1.25;
  --size-small: 12.8px;
  --size-h3: 25px;
  --size-h2: 31.25px;
  --size-h1: 39.06px;
  --color-on-bg: #F2F4F5;
  --color-on-surface: #F2F4F5;
  --color-on-primary: #101418;
  --color-on-accent: #101418;
  --color-muted-on-bg: #B3BCC5;
  --color-muted-on-surface: #B3BCC5;
  --color-link: #FFB300;
  --color-link-on-surface: #FFB300;
  --display-weight: 600;
  --subhead-weight: 500;
  --body-weight: 400;
  --strong-weight: 600;
  --label-weight: 500;
  --display-tracking: 0em;
  --display-leading: 1.2;
  --display-case: none;
  --body-leading: 1.5;
  --body-size: 16px;
  --icon-stroke: 1.5;
  --ground: linear-gradient(var(--color-accent), var(--color-accent)) 0 0 / 100% 6px no-repeat,
    repeating-linear-gradient(0deg, var(--color-on-bg) 0 1px, transparent 1px 2.5rem) 100% 0 / 14px 100% no-repeat,
    repeating-linear-gradient(0deg, var(--color-on-bg) 0 2px, transparent 2px 10rem) 100% 0 / 28px 100% no-repeat,
    var(--color-bg);
}
```

Tailwind patterns:

| Element | Classes |
|---|---|
| page | `bg-[#101418] text-[#F2F4F5] font-sans antialiased px-4 md:px-8` |
| heading | `text-2xl md:text-4xl font-semibold leading-[1.2] font-mono` |
| body | `text-base leading-[1.5] max-w-prose` |
| section | `rounded-sm border-2 border-[#3DDC84] bg-[#1B2128]` |
| button | `rounded-sm border-2 border-[#3DDC84] bg-[#3DDC84] px-4 py-2 text-sm font-medium text-[#101418] transition-colors duration-200 hover:bg-[#FFB300]` |
| label | `font-mono text-xs uppercase tracking-[0.08em] text-[#FFB300]` |

## Visual pack

The pictures and icons of this style are files. Use them as they are, or read them as a specification and draw your own in the same language.

| File | Content |
|---|---|
| [`pack/hero.svg`](./pack/hero.svg) | Hero plate, 800 by 600. An abstract picture in the visual language of the style. |
| [`pack/pattern.svg`](./pack/pattern.svg) | Pattern tile, 80 by 80. It repeats in both directions. |
| [`pack/icons.svg`](./pack/icons.svg) | Icon sprite: home, search, heart, star, bell, user, mail, calendar, image, bolt, play, check, close, sliders. |

- **Icons:** outline, stroke 1.5 on a 24px field, round caps, round joins, container: square. Use: `<svg class="icon"><use href="pack/icons.svg#home"/></svg>`. An icon takes `currentColor`.
- **Icon family:** the sprite is cut from Lucide (Regular) by Lucide Contributors, licence ISC. Reason: A thin line family that matches analysis tools and keeps the face visible. It is an open-source stand-in for a demo. For a product, use the icon library of the brand and choose the family by the rules below.
- **Display type:** weight 600, tracking 0em, leading 1.2, case: none.
- **Weights:** display 600, subhead 500, body 400, bold in text 600, labels and buttons 500. The style uses 3 weights (400, 500, 600) and no other. The display is 200 heavier than the body. The accent is colour: one coloured word or field is the accent, so the weights stay quiet.
- **Body type:** 16px, leading 1.5.
- **Page:** hero split, left aligned, regular density, accent tilt 0°.
- **Ground:** raw. The ground is a plain tool panel: the footage is the content and the panel adds only marks. Put the class `ground` on the first screen of the page; it takes `--ground`, a stack of gradients built from the colour tokens. The ground can run behind the copy, so put the class `ground-copy` on each text block of that screen: the block gets `--color-bg` and a soft edge. Keep the rest of the page on `--color-bg`.
- **Text colour:** `--color-text` is for the main reading field only. On each field, use its token: `--color-on-bg`, `--color-on-surface`, `--color-on-primary`, `--color-on-accent`.

### How to choose an icon pack

An icon pack is a typeface for pictures. Choose it with the same care as the text face. These rules hold for an open-source set and for a licensed library.

1. **The stroke matches the stem.** Set an icon next to the body text at text size. The line of the icon and the stem of a letter have the same weight.
2. **Terminals and corners follow the radius token.** Butt caps and mitre joins go with a radius of zero. Round caps and round joins go with a soft radius.
3. **The fill mode follows the technology of the era.** Print eras cut solid forms. Screen eras with depth fill and shade the form. Current product interfaces draw an outline.
4. **The metaphors are not younger than the style.** Each picture is an object that the era knew. An icon with a later metaphor breaks the period.
5. **One family, one grid.** All icons come from one family and one variant, drawn on one field. Do not mix families on one surface.
6. **The family covers the product.** List each concept that the product needs before you commit. A family that has no icon for one of them is the wrong family.
7. **The colour is a token.** An icon takes currentColor. A second tone comes from a token, not from the file.
8. **The licence permits the use.** Read the licence of the set. Give credit where the licence asks for it. A product with a brand uses its licensed library and applies the same rules to choose the family.

## Techniques

These techniques started in one domain and moved into this style. Apply the rules as they are written; the full record is in [`TECHNIQUES.md`](../TECHNIQUES.md).

### Face-swap autoencoder (technology)

A pair of networks that learn two faces and then draw one face with the pose and expression of the other, frame by frame. Origin: internet, Reddit, 2017.

- Find and crop the face in each frame, and mark its landmark points.
- Train one encoder and two decoders on many images of the two faces.
- Decode the pose of face A with the decoder of face B.
- Blend the new face patch into the frame with a feathered mask of 8px to 24px.
- Mark the result as synthetic in any public use.

## Organisations

The groups, studios, companies, schools, channels and backers that carried this style.

- **Reddit** (platform, web, 2005–now), launched the style. A forum site of topic groups. The groups on Weirdcore and Liminalspace spread the images of empty, strange places.

## Permutations

Each knob varies the style inside a range where it stays itself. Keep each value in its range. `npm run permute -- <id> key=value` writes the tokens for other values.

| Knob | Default | Range | Changes |
|---|---|---|---|
| Base colour (`base_colour`) | #3DDC84 | #3DDC84, #101418, #1B2128, #0A0D10 | --color-primary, --color-on-primary, --color-link-on-surface |
| Accent colour (`accent_colour`) | #FFB300 | #3DDC84, #FFB300, #F2F4F5 | --color-accent, --color-on-accent, --color-link |
| Corner radius (`corner_radius`) | 2px | 0–4 px | --border-radius |
| Grid step (`grid_step`) | 8px | 4–8 px | --space-unit, --grid-gutter |
| Type scale ratio (`type_scale_ratio`) | 1.25 | 1.2–1.414 | --type-ratio, --size-small, --size-h3, --size-h2, --size-h1 |

## Lineage fallbacks

When the user asks for a variation, do not improvise. Blend toward the neighbour named here, and load its skill.

- If the user asks for **more obvious**: move toward [`mini-diffusion-meme`](../mini-diffusion-meme/SKILL.md) (Mini-Diffusion Meme). Replace the real footage with a smudged 3x3 grid and a text prompt as the caption.
- If the user asks for **more polished**: move toward [`diffusion-gloss`](../diffusion-gloss/SKILL.md) (Diffusion Gloss). Make the whole frame a smooth render with an even rim light; keep the label chip.
- If the user asks for **show the seam**: move toward [`gan-portrait`](../gan-portrait/SKILL.md) (GAN Portrait). Show a still of a generated face with the eye line fixed at 40% of the frame height and mark the artefacts.

- **Came from (go here for a more orthodox version):** [`deepdream-psychedelia`](../deepdream-psychedelia/SKILL.md)
- **Mixes well with:** [`gan-portrait`](../gan-portrait/SKILL.md), [`reaction-gif`](../reaction-gif/SKILL.md)
- **Kin in another domain:** [`vertical-short-form`](../vertical-short-form/SKILL.md)
- **Nearest by vector profile:** [`cinemagraph`](../cinemagraph/SKILL.md), [`gan-portrait`](../gan-portrait/SKILL.md), [`dual-camera-snapshot`](../dual-camera-snapshot/SKILL.md)

## Emotional target

The viewer feels doubt: the face looks right, and the small seam says that it is not.

**Keywords:** uncanny, clinical, synthetic, doubtful, precise, unsettling · **Brand archetypes:** magician, sage, outlaw

## Sources

- [Wikipedia: Deepfake](https://en.wikipedia.org/wiki/Deepfake)
- [Thies et al., Face2Face: Real-time Face Capture and Reenactment of RGB Videos (CVPR 2016)](https://arxiv.org/abs/2007.14808)

---

Generated from [`data/internet/styles/deepfake-face-swap.json`](../../data/internet/styles/deepfake-face-swap.json) by [design-ontology](../../README.md). Do not edit this file; change the data and run `npm run build`.
