從真實影片案例學習 MiniMax H3 提示詞

Maurice Bourdon · @MauriceBourdon

Multiple locally generated text-to-video segments are context-connected into a continuous two-minute visual loop.

Kent Dhani · @kentdhani

A side-by-side comparison highlights how a MiniMax H3 fine-tune changes feather detail and quality during movement.

グラビアのAI美少女 · @DateAIko_SD24

An eight-second H3 experiment using a four-step LoRA on a 12 GB RTX 4070 with 32 GB system RAM.

有趣的80后程序员 · @sunbaolong_2001

A local action workflow combines FastH3 V2, two-stage latent upscaling and a dual-pass face fix.

aicreataro · @aicreataro

A ten-second H3 green-screen character is keyed into an After Effects space with animated cameras and lights.

Alex Shev · @AlexshevPm

A ComfyUI workflow passes ten seconds of video and audio context into subsequent sections before latent upscaling.

Wildminder · @wildmindai

An H3 style LoRA explores retro science fiction with cyborgs, laser guns and synthesizer sound.

CALMDUST aka ちゃんこあ · @core_tan

A WebUI test combines RefMod with three-step Ref2VA generation at 768p on an A100 80 GB.

Stable Diffusion Tutorials · @SD_Tutorial

A work-in-progress reference adapter compresses image or video references into small latents without training.

鈴木憂一 | Highdrama · @yu_ichi_suzuki

An experiment applies TaoMate by block group and resumes a three-step preview to an eight-step render.

Solutions Joe · @SolutionsJoeG

A 40-second restaurant promo turns food and terrace photos into H3 footage, then adds tracked graphics and a sound mix.

石ちゃん🎤LLMでアプリ開発 · @Isichan_Hitori

A local ComfyUI experiment removes token_refiner to test higher-resolution H3 generation within a fixed VRAM budget.

CALMDUST aka ちゃんこあ · @core_tan

A RefMod demo pairs an eight-second H3 result with icon-based reference management and prompt-tag insertion.

えふのじ · @fnoji

H3 green-screen output passes through SAM3, MatAnyone and RTX VSR; retained green outlines aid final keying.

Yogi Seezatnap · @seezatnap

An agent-led experiment gives Opus access to local H3 on an RTX 5090 for roughly four hours of video work.

MindfulReturn 身心修复局 · @MindfulReturn

A local Qwen agent makes prompts and first-frame images, chains ten-second H3 clips and joins the sequence.

LerSent · @LerSentAI

A deployed scrolling outfit-change webpage combines MiniMax Code development with H3 video generation.

Victor · @victor_explore

A RunPod H3 workflow scripts, voices, generates and reviews a reported 27 promotional reels for a baby kick counter app.

CALMDUST aka ちゃんこあ · @core_tan

A WebUI 1.6 example combines Ref2VA, two-step PDMD and RefMod, reporting 124 frames generated in 35.9 seconds.

🟢DeepForest_Studio_AI🟢 · @DeepForest_Inc

A side-by-side test compares INT8 ConvRot, Standard and LynnReal Light VAE processing after the same H3 sampler conditions.

shi3z · @shi3z

A local H3 demonstration on DGX Spark reports roughly twice the speed using training-free Layer Adaptive SLA.

Kamimoto(かみもと) · @sep_is_heim

An RTX 4070 I2V example reports 124-frame TaoMate H3 output in 2:26, with a follow-up publishing its ComfyUI optimizations.

Greg Horvay · @greg_horvay

An Intel B70 H3 workflow reports a 1344×768, 107-frame Ref2VA clip dropping from 42 minutes to about nine.

Inner-Reflections · @InnerRefle11312

A local H3 demonstration explores temporal upscaling as a way to improve generated video results.

ゆずき · @uzuki425

A Japanese H3 experiment shares a memory-optimization table and contrasts optimized execution with slower low-VRAM offloading.

Kamimoto(かみもと) · @sep_is_heim

A 12GB RTX 4070 benchmark library compares Ref2VA and FL2VA outputs and provides downloadable generation workflows.

烁皓 · @eternityspring

A Kung Fu Panda example joins two H3 segments by preserving 22 motion-context frames and trimming the overlap.

rebekah · @bekhovsgun

A stateful backend drives H3 Max Turbo video UI, with mouse and keyboard events triggering fresh prompts and last-frame transitions.

☬ŘØβØ༒ƤŞ¥ĆĦØ☬ · @RobopsychoAi

An RTX 4090 user shares a two-pass latent-upscaler workflow with 20-step and four-step options for local H3 clips.

Andy Orsow · @andyorsow

Andy Orsow demonstrates using H3 Max Turbo on fal for quick concept exploration and early prompt tests before video production.