안녕하세요
Minimax H3 R 2 V 워크플로우 입니다
알리바바의 8Step Turbo lora 를 사용했고
Sage attention 설치가 필수 입니다.
저사양 PC 부터 고사양 PC 까지 안정되고 빠른 결과물을 출력 합니다.
처음에는 0.4 메가픽셀부터 시작해서 해상도와 길이를 늘려가면서 테스트해보세요
튜토리얼
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 (16)
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
Please add an option to bypass the LLM if I want to type in my prompt or refine an existing prompt.
Great work! Thanks for the custom nodes!