Video summary

How to Make Videos Like Chloe vs History With AI (Full Guide)

Main summary

Key takeaways

Technology

Summary of the “AI video like Chloe vs History” workflow (technological concepts + features)

Goal

Create a 60-second “history vlog” AI video using a workflow the guide claims can be completed quickly (<20 minutes), producing a cinematic, consistent result.

Core approach (high level)

  1. Plan the story + scenes first
  2. Generate reusable assets (character + locations)
  3. Generate each scene in an AI video model
  4. Edit and export the final video

1) Story + scene planning using Claude (prompt template)

  • Uses Claude to generate a complete plan from a single prompt.
  • The prompt includes:
    • Video length/type: “60-second AI history vlog” for YouTube
    • History location: e.g., Kamakura, Japan
    • Time period: e.g., 1250 AD
    • Creator/persona definition: an AI influencer description (example given: Yuri)
    • Vibe/mood selection
    • Camera style
  • Claude output includes:
    • A scene breakdown (example: 8 scenes total)
    • Dialogue + voiceover directions per scene

Key advantage (claimed): pre-planning a scene list saves time versus building scenes manually.

2) Asset generation in Higgsfield (character consistency)

Character creation (Cinema Studio)

  • Uses Higgsfield (Cinema Studio).
  • Settings:
    • Model: Nano Banana 2
    • Output quality: 4K
  • Process:
    • Upload a reference image
    • Provide prompts for a multi-view character sheet (no text)
    • The character sheet generates multiple angles plus close-ups (eyes/mouth/ears) to help preserve identity in later video generation.

Outfit handling (improves reliability)

Instead of the common beginner approach (combining a separate outfit image), the guide recommends:

  • Create a new character sheet wearing the exact outfit needed for each costume change.

Benefits highlighted:

  • Better texture/detail fidelity (including small elements like sandals)
  • Avoids “weird changes” between clips
  • Cleaner swaps when switching outfits across scenes

Saving assets

  • Download character sheets and import into Higgsfield as elements, categorized as character.

3) Location generation for cinematic, consistent backgrounds

Locations (Cinema Studio)

  • Also done in Higgsfield using “cinematic-only” trained features (described as trained on cinematic footage).
  • Model selector: cinematic locations
  • Example: generate 4 locations, each with distinct cinematic styling, such as:
    1. Market/street (dark documentary-like lighting)
    2. Busy market (prompt emphasizes “respected” clothing/period feel)
    3. Courtyard (natural colors; avoids over-saturated “AI look”)
    4. Hillside temple (colder/serious mood)

Prompt structure:

  • Each location prompt includes place + time period
  • Sometimes the time-period wording is authored using Claude in the same session.

Saving assets

  • Locations are saved similarly to character references, but categorized as locations.

4) Video generation per scene (SeaDance 2.0 + references)

Scene generation flow (Higgsfield Video section)

  • Uses SeaDance 2.0 as the video generator model.
  • Before writing prompts, the workflow attaches:
    • Character reference and/or location reference as needed
  • Resolution/format:
    • 1080p, 16:9
  • Audio:
    • Audio enabled so ambient sound can be generated

Hook scene technique

  • The hook prompt includes instructions for shaky camera motion to mimic handheld travel vlog footage.
  • Emphasis: subtle imperfections improve believability and help avoid “AI slop.”

B-roll vs talking scenes (different reference strategy)

  • B-roll shots
    • May use only location references (no character sheet) to reduce continuity problems.
  • Talking scenes
    • Attach both character + relevant location
    • Use Claude-provided prompts.

5) Handling voice problems (voice cloning / “change voice”)

Issue

  • Generated clips can have random AI voices that don’t sound like the creator.

Fix

  • Use Higgsfield’s Audio tab feature: “change voice”
  • Select the creator’s cloned/custom voice
  • Provide a reference video for voice cloning (upload + generate)

When it’s applied:

  • Only for scenes where the character is speaking (B-roll doesn’t need voice cloning).

6) Final assembly + export (CapCut)

  • Import all generated clips into CapCut
  • Trim/adjust timing
  • Export at 1080p

Result claimed: a continuous “same day” feel—same city, same character, but with consistent speech.

Troubleshooting / tips explicitly mentioned

  • Generate multiple times if pronunciation/timing is inconsistent
    • Example: English → Japanese (“konnichiwa”) changes across runs.
  • Use new character sheets whenever the outfit changes.
  • Add realistic “imperfections” such as camera shake to increase believability.

Main speakers / sources (as stated or implied)

  • Speaker: the tutorial creator (uses “I,” mentions generating “me/Yuri” as the persona)
  • Tools used:
    • Claude: prompt template + scene/dialogue planning
    • Higgsfield: Cinema Studio + SeaDance 2.0 + Audio “change voice”
    • CapCut: trimming/editing and final export

Original video