從真實影片案例學習 MiniMax H3 提示詞

Blaine Brown · @blizaine

Maestro demonstrates H3 reference editing of longer stock footage by splitting it into windows for character swaps and effects.

ミチル｜今日もAIが気になる · @michiru_aixbtj

A robot added to an empty gallery picks up a ball and passes behind a pillar, while a requested one-second run takes about three.

えふのじ · @fnoji

A creator shares H3 source clips and a glasses variant while reporting persistent oversized eyes and character-detail inconsistency.

Idorenyin Joseph | AI Video creator · @IddyJosephh

A widely viewed two-video comparison labels the first result H3 Max and the second Gemini Omni Flash.

Day of the Answer · @DayoftheAnswer

A creator labels a ComfyUI H3 result as Blender to Anime, combining 3D-source design with an anime treatment.

えふのじ · @fnoji

A character-reference preparation study pairs an anime output with a published sheet, without using Blender or other DCC tools.

Md Ismail Šojal 🕷️ · @0x0SojalSec

A locally generated celebrity likeness demonstrates identity conditioning with a reported 155 MB LoRA and no reference images.

akatz · @akatz_ai

An original, green-mask and person-removed comparison demonstrates an H3 LoRA used to prepare character-swap training samples.

1つ目クロスケ · @HitotsumeKuro

Two H3 anime outputs hold seed and references constant while the author investigates prompt-driven character duplication.

Graalito · @Graalitoo

An H3 Max video-reference experiment achieves synchronized dual viewpoints once, while its author reports many failed attempts.

元荒川ちゃん❤ · @mtArakawa

An H3 group-flight test adds individual character instructions after unspecified actions made everyone move identically.

WBK · @WBK_SZ

A creator switches an action sequence from R2V to I2V after repeated difficulty and retains visible roughness in the local H3 result.

ふとし · @mmmpa

A single-photo 360° LoRA test exposes an invented tail as H3 fills in character details absent from the reference view.

wasabi_vaio · @WasabiVaio

Krea2 helicopter stills guide H3 animation of a large science-fiction helicopter formation.

𝐙𝐞𝐧𝐠 💜 · @zeng_wt

A pastel surreal base frame made with Muse Image is animated with H3 Max in a dream-state study.

buchi_ · @mix_buchi_

A kimono character study is generated with H3 Ref2V using the Singularity model according to its creator.

monoradio · @monoradio102

A manga prototype combines H3 and ORB360 LoRA for adjustable viewpoints, with character references used to stabilize faces.

FoldingEnd A.I. · @EndFolding79421

A ComfyUI H3 workflow test publishes realistic and anime character sheets with face, body and detail views.

huangserva · @servasyy_ai

A side-by-side RTX 4090 test compares three-view Ref2VA orbit guidance with a single-image first/last-frame loop.

Kent Dhani · @kentdhani

A local RTX 5090 test replaces three Jurassic World raptors at once rather than swapping one isolated character.

buchi_ · @mix_buchi_

A frog-shaped robot is animated with H3 Ref2V using Singularity and an LLM-written prompt.

Satsuki AI Log · @SatsukiAILog

A background-and-person Reference Edit test improves placement but changes the face and skin texture, according to its author.

Jerrod Lew · @jerrod_lew

One reference image becomes draft images in Magnific One, H3 Max video clips and an edited spec ad with ElevenLabs audio.

akatz · @akatz_ai

H3 Ref2VA transfers performances with a character V1 LoRA, prompting and masking in a side-by-side development reel.

Kamimoto(かみもと) · @sep_is_heim

An anime-character test combines start and end images with a Transformation LoRA and a prompt.

さんかくてん · @sankakuten91256

A labeled comparison tests H3, Seedance and Wan against previs, depth, line art and a character reference in five-second sections.

バスケットマン（基本無課金 多垢100連ガチャ） · @kaimakulink

A creator uses Gemini to reconstruct an H3 anime-to-live-action workflow from a LoRA description, showing a graph and a short result.

LinacoWSL · @breath_in_code

A local 20-second H3 clip uses one person image and one chibi image, with prompting rather than a storyboard or LoRA.

kongo jun / 混合 順 · @jun_kongo

A creator publishes a video tutorial about using a FaceSwap LoRA with MiniMax H3 in ComfyUI.

なまなり あい · @AiNamanari33743

A sunset-train experiment pairs an H3 character result with a simplified previsualization across multiple angles.