從真實影片案例學習 MiniMax H3 提示詞

Kamimoto(かみもと) · @sep_is_heim

A four-step H3 pipeline enlarges latents after step three and decodes with X2 VAE; its author reports 116.6 seconds for a 14-second clip.

AIossansan · @toraaiuser2

A creator passes images directly from Codex to ComfyUI for H3 animation, removing repeated manual prompt copy and paste.

めたまむ☺︎ 絵本アプリ「えほんのこや」 · @metamom_mama

A local app promo turns hand-drawn art into clay-style composites, animates them with H3 and inserts phone-screen footage in post.

iGavroche · @AiGavroche

A Strix Halo developer reports Veda sparse attention reducing an eleven-second H3 run from 691 to 366 seconds in ComfyUI.

⛏️🌱Don Kackman 🌱⛏️ · @Dkackman

An all-local demonstration uses Qwen 3.8 27B in OpenCode to prompt MiniMax H3.

Hong.Aigc · @newdvn2025

An open-source AutoDL skill connects H3 shot planning, six- or nine-image references, first/last frames and image-plus-audio references.

FoldingEnd A.I. · @EndFolding79421

An H3 study compares an experimental cyberpunk LoRA on and off at twenty and eight steps, then adds a thirty-step nightclub render.

ドンメカ｜メカノ姫β · @DON_Mechaaa

A single DGX Spark Ref2VA test animates one official Jensen Huang photo with local audio at 768p, reportedly taking about 25 minutes.

thefirmament17 · @thefirmament17

A same-still, same-seed RefMod comparison reports 130s versus 277s at 1408×800, with different available acceleration stacks.

Nina小娜学AI · @geesehowardt7

A local H3 workflow combines a Hybrid model, Selflift sampling enlargement and X2 VAE with a published video example.

Kent Dhani · @kentdhani

An RTX 5090 comparison reports 136s for DMAD at four steps versus 247s for Turbo v4 at eight, with fast-motion smearing.

中崎工房 | AIで1時間の仕事を5分で終わらせる人 · @nakazakifam

A psychology explainer combines local H3 T2V clips, LTX avatar video, code-led editing and separate speech and music tools.

aicreataro · @aicreataro

Python redraws H3 animation as manga panels with ink and color stages, animated lettering, camera paths and page turns.

三味松 · @shamimatsu

A thirty-second dance test uses four eight-second Extender segments and reports seam degradation, darkening and style drift.

kimutaka · @kimutaka_ai

A twenty-second ad used 46 generated video seconds, with H3 Max/Turbo selection, image revisions and a music retry affecting cost.

ヘミ · @Hemi3AI

A 16 GB RTX 5060 Ti test reports 126.75s versus 105.35s with INT8, 5% SLA, LBH enlargement and X2 output.

久田誠良(Seira Hisada) · @SeiraHisada

A short film uses H3 for visuals and character dialogue, with automated frame/audio checks, Colab rendering and human revision requests.

Ark · @quimedesu

A local single RTX 3090 ComfyUI example reports twenty minutes for each ten-second H3 Q6 generation.

AI Bard Guild · @IsekaiBardGuild

Tripo3d meshes use front and side H3 action videos as references for Opus-led rigging and motion, including difficult winged anatomy.

沉默的未知 · @hylwr

A 24 GB RTX 3090 pipeline uses Codex to research a science topic, create script and storyboard, then render local H3 clips with subtitles.

みみみ · @mmm_tech

A small one-second DGX Spark test took 36s but missed night styling and melted a robot; the five-second run shut the device down.

Ark · @quimedesu

A local RTX 3090 example reports a five-minute render using H3 Singularity INT8 with Turbo LoRA.

bigiver · @bigiver66

A seven-stage local pipeline moves from topic and script to Qwen stills, H3 animation, cloned narration and voice-aligned ffmpeg subtitles.

ざすこ（道草 雑草子） · @zasuko_michiksa

A game-development stream documents FastH3-based eight-direction motions, OK/NG review, reference-overlay alignment and in-game checks.

HMLG · @HMLGGGG

A 24 GB RTX 3090 and 64 GB RAM rig reports 23 minutes for a 576×1024, twenty-step, ten-second H3 preview before paid cloud finals.

LBan · @LBan2049

A creator tests short-drama production on a rented RTX 4090 ComfyUI server and reports a low hourly rental cost.

元気ファイン · @genkifine_hero

A Unity scene switches among 64 extracted H3 orbit views according to camera position, with added lighting instead of a 3D character mesh.

有趣的80后程序员 · @sunbaolong_2001

A FastH3 Trim workflow uses four anchors over ten seconds and a two-stage latent upscale to address glossy skin artifacts.

ITOK · @ITOKtw

A ComfyUI H3 still-image workflow describes repeating the latent five times and decoding the third to reduce artifacts.

布留川英一 / Hidekazu Furukawa · @npaka123

A Colab experiment runs H3 from Codex and reports 6 minutes 32 seconds for a ten-second 960×544 anime clip.