What AI Tools Can Make Videos Like haru.vs.history?
If you are asking what AI tools can make videos like haru.vs.history, start with the job split, not one brand name.
Explore Haru.vs.history ProfileIf you are asking what AI tools can make videos like haru.vs.history, start with the job split, not one brand name. The creator has not publicly disclosed any tool list, so the useful answer is a recommendation pool: one lane for Haru's character reference, one for motion and camera, and one for audio and subtitles. That is the stack most likely to reproduce this kind of work.
Methodology: I analyzed 4 of @haru.vs.history's works from 2026-03-20 to 2026-05-07 to map the tool roles that can produce this output - character consistency, historical environments, motion, and speech - and cross-referenced approved capability cards in `research/tool-capabilities/`. Last updated 2026-05-28.
What This Content Looks Like - the Signals Your Stack Has to Match
The selected clips are built around one recurring constant: Haru. She stays recognizable in a navy cardigan, plaid skirt, and low ponytail while the era around her changes from Heian court color codes to a Kyoto night fight and then to Meiji street scenes.
The second constant is speech. These videos are explanation-led, not silent spectacle, so the stack has to keep Japanese narration, reaction beats, and subtitles readable at the same time. That means the first technical problem is not motion, it is continuity.
The clip uses Haru's modern cardigan as the contrast point inside a hall of purple, red, and light blue court robes. It turns rank into a color test, so the output has to keep those hues separated and readable. It also pushes into a forbidden orange-yellow Crown Prince color, which means the model has to preserve fine color categories, not just a vague "historical" look.
The alley fight depends on a low-light lantern read: Ikeyado signage, blue haori figures, sword movement, and a breathless voiceover all have to stay legible while the camera shakes. If the scene gets muddy, the historical moment collapses, because the viewer needs to understand both the chaos and the explanation at the same time.
Key Insight: All 4 selected works keep Haru recognizable while the historical environment changes completely around her.
Takeaway: Build the character board first, then vary era plates and camera angle.
Bottom Line: Haru identity appears in 4 of 4 selected works, so reference consistency is the first technical requirement.
Tools That Can Produce This Kind of Work
I split the recommendation into four jobs because the videos split that way in practice: image plates for Haru, video generation for the historical motion, motion control for repeated gestures or walks, and audio/post for narration and ambience.
For the editorial formula behind that split, see the companion G3 breakdown How to Make Videos Like haru.vs.history. This page stays on what can produce this output, not on the formula itself.
| Role | Recommended tools | What each is good at | Distinctive signature | Alici alternative |
|---|---|---|---|---|
| Image generation (character / period plates) | Nano Banana Pro · Seedream · GPT Image 2 · Midjourney v8.1 | Nano Banana Pro is the safest pick for multi-reference character consistency and photorealism; Seedream is better when you want filmic light and texture; GPT Image 2 is useful for storyboard sets and text-heavy reference boards; Midjourney is good for moodboards and period costume exploration. | — | Nano Banana Pro · Seedream · GPT Image 2 · Midjourney text-to-image (partial) |
| Video generation (motion / scene continuity) | Veo 3.1 · Kling 3.0 · Hailuo 2.3 | Veo 3.1 is strongest when you want native synchronized audio and vertical output in one pass; Kling 3.0 is the best general pick for multi-shot motion stability; Hailuo 2.3 is useful for emotional close-ups and facial micro-expression. | — | Veo 3.1 · Kling 3.0 · Hailuo 2.3 |
| Motion / camera control | Kling 3.0 Motion Control · Runway Gen-4.5 | Kling Motion Control transfers a clear reference performance onto a still character; Runway gives you Motion Brush and tighter camera steering, but it is not on alici. | — | Kling 3.0 Motion Control |
| Audio / post-production | ElevenLabs · OpenAI gpt-4o-mini-tts · ElevenLabs SFX v2 · Stable Audio 3.0 | ElevenLabs is the voice-layer leader, gpt-4o-mini-tts is useful for controllable delivery, ElevenLabs SFX v2 handles precise foley, and Stable Audio is better for longer ambience beds. | — | none on alici |
One model I would not make central here is Seedance 2.0. Its card is explicit that it is strong for stylized characters and camera-driven motion, but not for realistic human faces, which makes it a weaker fit for Haru's semi-photoreal schoolgirl look.
Veo 3.1 is the cleanest one-pass option if you want motion and sound together, but I would still treat subtitles as a post layer because its card warns about in-frame text drift. Kling 3.0 is the best general motion choice when Haru has to walk, turn, or react while staying coherent.
Nano Banana Pro is the safest first stop for Haru's face because the card emphasizes multi-reference character consistency. Seedream is the better pick when the historical street or interior needs more filmic texture, while GPT Image 2 is the fastest way to build a storyboard board with readable text and layout.
Hailuo 2.3 deserves a place in the pool because these videos are emotional, not just procedural. It is strong on facial micro-expression, but because it ships silent, it still wants a separate audio layer.
Key Insight: The selected set splits into reference, motion, camera, and audio jobs, so no single model covers the formula cleanly.
Takeaway: Start with the hardest failure
Bottom Line: A 4-role stack covers all 4 selected works, and only one of those roles changes materially across clips
What's Hard About Making This Kind of Content
The hard part is not getting a generic "history video" look. It is keeping one modern protagonist readable while the era, lighting, and camera grammar keep changing.
A scene like the Shinsengumi alley asks for low light and sword motion; a scene like Meiji Ginza asks for street-scale crowd flow and architecture; an Edo rule explainer asks tiny styling details to stay legible. If any one layer drifts, the whole clip feels like a different series.
What's harder to do well - Keep Haru's face, hair clip, cardigan, and skirt stable across court halls, alley fights, and city streets. - Keep Japanese narration clear while the camera moves and the scene gets noisy. - Keep era-specific color, clothing, and architecture distinct so Heian, Edo, and Meiji do not collapse into one generic preset. - Handle both intimate selfie shots and wider crowd scenes without losing subject placement. - Make the speech read as natural rather than pasted on top of the visuals.
This is the clearest sign that the workflow is built as a production brief, not a single render. The title itself makes hair a social marker, so the visual system has to preserve small grooming differences while Haru stays on-camera and explains the rule set. It is also a reminder that the speaking face matters, because the clip is trying to teach a rule, not just show a costume.
The Meiji street scene asks for a different kind of control: brick streets, Westernized facades, and busier crowd motion. The background has to carry modernization without turning Haru into a blurred extra, which is a harder continuity test than a closed room.
Key Insight: In 4 of 4 selected works, the scene only works if the model can carry Haru through a harder environment than the last one.
Takeaway: Test your stack on the worst scene first
Bottom Line: The reproduction bar is high because the stack has to lock identity, speech, and era at once.
Where the Recommendation Falls Short
I can recommend the tool roles that fit this formula, but the finished clips do not let me prove the exact stack.
- exact_tool_stack - The creator has not publicly disclosed the setup, so this page stays at the role level.
- specific_model_version - The finished clips do not uniquely identify one model version.
- custom_trained_identity_methods - I cannot tell whether Haru comes from a reference pack, private character system, or manual curation.
- post_processing_pipeline - I cannot tell whether grading, timing, subtitles, or cleanup happened in a separate editor.
- voice_or_audio_pipeline - I cannot tell whether speech came from native video audio, TTS, dubbing, or separate sound design.
I also keep Seedance 2.0 out of the core pool for this specific creator. The card says it is reliable for stylized characters, but not for realistic human faces, and Haru's look depends on a believable schoolgirl face rather than a cartoon avatar.
That boundary is part of the answer, not a weakness in the analysis. The right recommendation is a pool with clear job boundaries, not a fake certainty about the exact stack.
Key Insight: The finished clips show what works, not the exact stack that made them.
Takeaway: Use the pool as a starting map, then test the smallest combo that preserves identity and audio.
Bottom Line: The evidence is strong enough to recommend roles, not strong enough to name the exact stack.
FAQ
What AI tools can produce videos like haru.vs.history?
Start with Nano Banana Pro or Seedream for Haru reference frames, then use Veo 3.1 or Kling 3.0 for motion. Add Kling 3.0 Motion Control when a specific gesture or walk has to survive across scenes, and finish with ElevenLabs or gpt-4o-mini-tts if the speech needs a separate polish pass.
Which tool should I start with if I want Haru's face to stay consistent?
Nano Banana Pro is the safest first pick because its card emphasizes multi-reference character consistency. Seedream is the second option when the historical scene needs more filmic texture and natural light.
Do I need native audio, or can I add voice later?
You can add voice later, and in this format that is often cleaner. Veo 3.1 can keep synchronized audio in one pass, but Hailuo 2.3 ships silent and Kling or Runway workflows still benefit from a separate audio layer.
Can I make this style without premium tools?
Yes, but the fidelity drops. GPT Image 2 and Midjourney can help with boards, and lower-cost voice and SFX tools can cover the rest, but the safest path is to keep the stack role-based rather than all-in-one.