How to keep a character consistent in AI video

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

One mascot, trainer or actor that looks identical across many images and clips. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: Character sheet + N scenes (stills or clips) Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. Character Design (varg-character-design, varg, live) — Start here: How to define the character, its details and clothing, then the sheet and how to pass it to Nano Banana, Kling 4 and Seedance 2.5 (tested)
  2. Consistent Characters (varg-consistent-characters, varg, beta) — Start here: Character profile cards, location card, scene composites approved before video
  3. fal Character Design (fal-character-design, fal, live) — Alternative: Reference and expression sheets, outfits, consistent characters · Needs: fal API key (FAL_KEY)
  4. Character Design Sheet (inference-sh-character-design-sheet, inference.sh, live) — Alternative: Turnarounds, expression sheets and palettes · Needs: inference.sh account, belt CLI (belt login)
    • Claude Code: npx skills add inference-sh/skills --skill character-design-sheet -a claude-code
    • Codex: npx skills add inference-sh/skills --skill character-design-sheet -a codex
    • Cursor: npx skills add inference-sh/skills --skill character-design-sheet -a cursor
    • Gemini CLI: npx skills add inference-sh/skills --skill character-design-sheet -a gemini-cli
    • Details: https://varg.ai/how-to/skills/inference-sh-character-design-sheet.md
  5. Higgsfield Soul ID (higgsfield-soul-id, Higgsfield, live) — Alternative: Trains a face model from photos on Higgsfield · Needs: Higgsfield account, Higgsfield CLI (higgsfield auth login)
  6. varg AI (varg-ai, varg, live) — Adds: Reference-image generation and video on one key
  7. Awesome Seedance Skills (awesome-seedance-skills, LearnPrompt, live) — Adds: The @-reference order for Seedance

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Style?

Use?

Start from?

How to keep a character consistent in AI video: step by step

  1. Make a three-panel map per character (front and back full body with the face blurred, one large frontal portrait) on mid-grey, and a new map for every story state; one empty plate per location; props shot with a human and their size stated in inches
  2. In every video prompt give each @image one job (face from the portrait, outfit from the full-body views, location, prop) and say the sheets guide appearance only
  3. For multi-shot scenes add a numbered storyboard grid as a second ref: it owns order, framing and action, the sheet owns the face
  4. Generate 15-30 s scenes with timed hard cuts, left/right staging and continuity lines; check each cut against the map and re-render any jump

What works now (researched 2026-10-27)

The character sheet is now the standard input: one board with front, side, back, face close-ups and props, fed as a labeled reference to Seedance 2.x, Kling 3 or Vidu Q4 and often shown next to the result. Per-character LoRAs are now mostly for high-volume influencer pipelines. What spreads is a recognizable character who comes back episode after episode.

  1. Make one master sheet before any scene: full body front, three-quarter, side and back, 3-4 face close-ups with expressions, key props and a palette. Paste the same short text description of the character into every prompt. The most-saved consistency post of the year is exactly this: a GPT Image 2 sheet animated with Seedance, 155k views and 1,672 bookmarks, prompt in thread. (Kiki (@Mayz1169) on X, character sheet to Seedance (Jul 2026))
  2. Give every reference one job and say it in the prompt: 'image 1 = face only, image 2 = outfit and hair, image 3 = location'. Say what a reference does NOT control. Seedance binds files with @Image1-@Image9 tags, and uploading a file is not enough: you have to state what it is for. Kling's element binding reads only facial data, so give front and side face views. (Luma, Seedance 2.0 complete guide)
  3. Storyboard in rough sketch style, not photoreal. Photoreal boards pull the face away from the sheet; sketch boards keep it. A creator tested both and shared prompts; 516 bookmarks on 38.9k views (save ratio 0.99). (Ivanna Humeniuk on X, sketch storyboards keep likeness (Jul 2026))
  4. Give a recurring character one signature feature readable at phone size (hair shape, a loud color) and a short name. Avoid masks, capes and props in hands, which break swaps and drift. A widely saved agent prompt for recurring AI characters (873 bookmarks) makes this the first rule of character design. (Chris (@everestchris6) on X, recurring AI character prompt (Oct 2026))
  5. Check every cut: compare the last frame of a shot with the first frame of the next and re-render on any face, hair or outfit jump. Extend long scenes from the last frame, going back to the master sheet often. Kling element binding fails when the character orientation does not match the video, and quality drops when the reference face differs a lot from the first frame. (Kling, Motion Control user guide)

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Three-panel character map (blurred-face trick) (any)

Turns one photo or generated portrait into the master sheet: front and back full body with the face blurred, one big frontal portrait as the only face reference. Adapted from Ivanna Humeniuk on X, prompt maps for a two-actor Seedance 2.5 scene (Oct 2026).

Create a three-panel character reference sheet of the {PERSON} in @image1 in a single wide 16:9 image. Preserve the exact facial identity, eye colour, hair colour, skin tone and natural facial proportions from @image1.
The left half contains two full-body views side by side: front and back, identical in scale, fully visible from above the hair to below the shoes. The right half contains one large head-and-shoulders portrait facing directly into the camera. All three views share one continuous neutral mid-grey studio backdrop, matching soft lighting and the same {FILM_LOOK} colour treatment.
In the full-body front view, replace the facial features with a smooth, softly blurred skin-toned oval. Keep the hairstyle and head silhouette intact. The large portrait is the only detailed face reference.
Hair: {HAIRSTYLE}. Outfit: {TOP}, {BOTTOM}, {SHOES}, {ONE_ACCESSORY}, with realistic fabric texture and seams. Build: {HEIGHT_AND_BUILD}. The same outfit, fit, hem lengths and hairstyle appear in all three views.
Front view: upright, square to the camera, arms relaxed, hands empty. Back view: the identical stance turned fully away. Portrait: head upright, chin level, both eyes into the lens, lips closed, natural skin texture with visible pores.
Soft frontal studio light, soft contact shadows under the shoes. The composition consists entirely of the three views and the seamless backdrop. No text, no labels.

Seedance scene with labeled references (seedance_2_5)

A 15-30 s dialogue scene where each @image has one job (face, outfit, location, prop), with timed hard cuts, fixed voices and continuity locks. Adapted from Ivanna Humeniuk on X, Scene 1 animation prompt (Oct 2026).

SCENE CONTEXT: A {SECONDS}-second scene in {LOCATION}, {TIME}. {WHAT_HAPPENS}. Play it as {TONE}. {CHAR_A} wants {GOAL_A}; {CHAR_B} wants {GOAL_B}.
ACTIVE REFERENCES: @image1 - {CHAR_A}: use the portrait for the exact face and the full-body views for outfit and hair. @image2 - {CHAR_B}: same rule. @image3 - the {LOCATION} layout, furniture and light. @image4 - {PROP}: keep its shape, colour and scale. Sheets guide appearance only; the output is full-frame footage.
SPACE AND PROPS: {CHAR_A} screen-left, {CHAR_B} screen-right; camera stays on one side of the dialogue axis, matching eyelines. {PROP_START_STATE}. {WHAT_EACH_HAND_HOLDS}.
CAMERA: 16:9, real time, hard cuts at the listed times, stable focal length within each shot.
0.0-{T1} s, {FOV} deg FOV. {SHOT}: {ACTION}.
HARD CUT {T1}-{T2} s, {FOV} deg. {SHOT}: {ACTION}. {CHAR_A}: "{LINE}" {INTENTION}.
HARD CUT {T2}-{T3} s, {FOV} deg. {SHOT}: {REACTION}. End on a reaction, not a line.
ACTING: behaviour, breaths, each listens while the other speaks.
LIGHT: {PRACTICALS}, shoot from the shadow side, fixed {KELVIN}K white balance.
AUDIO: {CHAR_A} VOICE: {AGE, PITCH, ACCENT, RHYTHM}. {CHAR_B} VOICE: {SAME_FIELDS}. Use the dialogue exactly. {ROOM_TONE_AND_FOLEY}.
CONTINUITY: {PROP_STATES_THAT_PERSIST}. Same faces, hair, clothes, screen direction and room geometry throughout.

Storyboard grid, then video (seedance_2_5)

Two stages: a numbered storyboard sheet (gpt_image_2_5) that owns order and framing, then Seedance with the character sheet owning only the face. Adapted from LearnPrompt/awesome-seedance, storyboard-grid-to-video skill (14 verified cases).

STAGE 1, storyboard sheet (image model):
Single-page storyboard sheet, {ASPECT}, {N} numbered panels, rough pencil previs style. Header card: title "{TITLE}", TOTAL VIDEO TIME: {SECONDS} SECONDS, {N} SHOTS, audio: {AUDIO_DIRECTION}.
1. {SHOT_SIZE} - {ACTION_ALREADY_IN_MOTION} - {WHAT_THE_PANEL_IS_FOR}
2. {SHOT_SIZE} - {ACTION} - {PURPOSE}
... one line per panel up to {N}
Every panel must contain visible motion. Same character and set in every panel. No timestamps, no extra characters, no watermark. Footer: {N} shots x {SECONDS_PER_SHOT}s = {SECONDS} seconds.

STAGE 2, video (@image1 = character sheet, @image2 = the storyboard):
@image1 is the EXACT main character reference: it controls only face, hair, body and outfit. @image2 is the EXACT storyboard: it controls locations, actions, compositions and sequence. Do not copy any pose or layout from @image1.
Rules: follow the panels from 1 to {N} in order; one shot per panel, about {SECONDS_PER_SHOT} s each; no skipped or added steps; the character and set stay identical.
Look: {LIGHTING}, {CAMERA_MOVEMENT}, {AUDIO}. No subtitles, no text, no watermark.

Six-view model sheet on flat grey (tested on six models) (seedream_5_pro)

The identity sheet every later image and video call receives: three full-body views and three large face close-ups, no text. Crop the face and the full body out of it for Nano Banana edits and Seedance @image1/@image2. Adapted from varg, character design skill (tested 2026-10-27).

Character reference sheet for AI video, flat light-grey studio background, even soft studio light, no text, no labels, no logos. Top row: three full-body views at the same scale on one ground line: front, side profile, back, arms relaxed, empty hands. Bottom row: three large head-and-shoulders close-ups: front, three-quarter, side profile. The same {PERSON} in every panel: identical face, hair, body, outfit and proportions.
CHARACTER: {NAME}, {AGE}-year-old, {HEIGHT_AND_BUILD_WITH_VISIBLE_CUES}, {FACE_FEATURE_BY_FEATURE}, {HAIR}, {SIGNATURE_DETAIL}, {EXPRESSION}.
WARDROBE: {EACH_GARMENT_WITH_CUT_COLOUR_MATERIAL_FIT}, unbranded.
Photoreal, visible skin pores and natural skin texture, no retouching.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Character sheet, then the action clip: A clip where the character sheet or the reference board sits next to or under the animation, so viewers can check that the face holds. Why it spreads: It is the most-saved character post format of 2026: 1,672 bookmarks on a desert-oracle sheet, 1,269 bookmarks on 38k views for a two-actor scene shown with its 12-reference board. People save it to copy the prompt. Keep: Show the sheet or the reference board on screen next to or under the clip; one sheet layout (full body, turnaround, close-ups) and the same short character description repeated in the video prompt; one 15-30 s clip. Swap the character, world and action. Make it with fal Character Design (fal-character-design).
  2. Recurring mascot series: The same mascot or animal character in a new scene every episode, numbered so people binge the page. Why it spreads: A dino counting-song mascot kept identical across 9 scenes got 2,683 bookmarks on 101.8k views. A Japanese creator ran a 40-episode numbered cat-chase series on Seedance (Apr-Oct 2026), and the same recurring cat-and-girl cast reached up to 476k views on a single clip. Keep: One mascot with one signature feature, one set, one new action per line or episode, numbered parts, captions drawn in code. Swap the mascot, set and song. Make it with Consistent Characters (varg-consistent-characters).

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials