Hello,
I would like to introduce version 1.0 of the MiniMax H3 workflow.
Without the stress of writing complicated prompts,
all you need to do is provide several reference images and a simple idea,
and this workflow will create a high-quality video for you.
It actively uses a local LLM, and I added several custom nodes to make it easier to use.
Although generation takes some time, the results were produced at the highest quality settings.
To better understand how the workflow works,
I think watching the videos will be helpful.
This workflow is based on the default workflow, which provides the best compatibility and highest quality,
with additional improvements for convenience and ease of use.
--
As you probably know, there are currently many performance-optimized options, including Turbo LoRAs, distilled LoRAs, and others.
However, the model was released only a few days ago, and there are still serious quality issues,
so version 1.0 does not include any generation-speed optimization patches.
I do not think that fast but unusable videos have any value.
Until better models and better techniques become available,
let’s take our time and create great videos.
In the next version, I plan to update the workflow with performance optimizations.
ㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡ
안녕하세요
Minimax H3 워크플로우 버전 1.0을 소개합니다.
복잡한 프롬프트 입력에 대한 스트레스 없이
여러장의 참조 이미지와 간단한 아이디어만 입력하면
완성도 높은 영상을 만들어주는 워크플로우 입니다.
로컬 LLM 을 적극적으로 활용하며, 사용 편의성을 위해 몇가지 커스텀 노드를 추가하였습니다.
워크플로우가 어떤 방식으로 작동하는지
영상을 확인해보신다면 확인에 도움이 되실 것 같습니다.
호환성이 완벽하고 품질이 가장 높은 기본 워크플로우를 중심으로
편의성을 개선한 워크플로우 입니다.
--
아시다시피 현재 turbo lora, distilled lora 등등 많은 성능 최적화 모델들이 있지만
모델이 공개된지 며칠밖에 되지 않은 상황속에서 아직은 품질문제가 심각하기 때문에
버전 1.0 에서는 생성속도 최적화 관련 패치를 넣지 않았습니다.
단지 빠르기만하고 사용이 불가능한 동영상은 아무 의미가 없다고 생각합니다.
좋은 모델이 나오고 좋은 기술이 나올때까지는 느긋한 마음으로
멋진 영상을 만들어 봅시다.
다음 버전에서는 성능 최적화가 포함된 워크플로우를 업데이트 할 예정 입니다.
Tutorial
Description
Initial release of the MiniMax H3 anime workflow.
- Designed and tested for an 8 GB VRAM environment using an RTX 3060 Ti with 32 GB of system RAM
- Reference-to-video generation using multiple reference images and a simple idea
- Optional local LLM prompt generation
- Native synchronized audio output
- Custom nodes added for easier setup and operation
- Quality-first default workflow with no Turbo or Distilled LoRA speed optimizations in v1.0
FAQ
Comments (14)
Stop interpolating anime!
DONT INTERPOLATE ANIME PLEASE
I BEG
I'll handle it
im kinda shocked, this is good, using turbo lora 6 steps, im impressed ngl, thanks for this
edit: any plans to add ref video also?
비디오와 오디오가 따로 저장되는 것 같은데 합쳐서 저장되게는 안될까요? 좋은 워크플로우 감사합니다!
I think convenience would be further improved if we could edit and save prompts generated by LLMs.
저장 됩니다.
Png 결과물을 다시 불러와보세요
@Deno2026 The simultaneously generated PNG file did not contain the LLM's output results. Could this be because I was using a custom setup with koboldcpp instead of LM Studio?
@tomgm777320 I'll check again!
@Deno2026 I’ve tested this a few times in a koboldcpp + gemma4 31b environment, and it seems the PNG file contains the prompt from the previously generated video (which means the prompt is empty for the very first run). The video itself is generated correctly.
It might just be an issue with my specific setup, though...
@tomgm777320 지금 새로운 버전으로 테스트해보시겠어요?
More interpolating anime