Hello, I mentioned this in the comments last time, but I'm organizing and sharing some basic tips I think about.
First
https://github.com/LeonQ8/ComfyUI-ALLinONE-MinimaxH3
This is an all-in-one node.
When you open ComfyUI, there are basic workflows, but it's convenient to use this instead of opening them one by one.
The second is AI utilization.
ComfyUI MCP has recently been supported, and tools like Antigravity IDE allow you to modify and create nodes if you give permission. You can also use Claude or Codex desktop. When something goes wrong, if you copy the error and throw it at AI, they'll mostly fix it automatically. In the past, installing Sage Attention on Windows was difficult and everyone struggled, but now we've entered an era where if you ask an AI to do it, they just do it.
The all-in-one node above was also created for convenience, and you can add more individual convenience and develop (!) it.
I've also seen people who have developed it really amazingly (!!!).
The third is rule application.
In fact, just like many people already use it for coding or projects, you can register skills, throw a rulebook, and tell them to produce prompts that follow rules when creating prompts.
https://github.com/MiniMax-AI/MiniMax-H3
You can basically throw down the skill MD files provided by the official GitHub, and it's also good to validate the prompts created by AI and build your own data.
The fourth is size.
No matter how much 2MP is supported, you can't create R2V for 15 seconds on a 5090.
I do most creations at 1MP and use an upscaler.
You can create a production workflow to match your computer's specs.
If you have a Pro 6000...
Finally, it's about licensing.
For various reasons, it's an excluded country at launch, but
https://platform.minimax.io/h3-license
You can go here and apply.
This information might not be useful for those already using it, but for those who want to try it, it would be good to just know about it.
Anyway, better models can come out, and like previous models, if it's forgotten, methods for this model won't be needed anymore, but personally, it seems like it will be quite a memorable model.
I made the video below both to test continuity that appears as a one-take and with my wishes in it.
Scenes that transition with cuts are easy to edit later, but having the camera continue like this to the end means inserting scenes with prompts...
I converted the mouth shape lip-sync to 60fps and made slight time adjustments separately.
I used a separately specified LoRA for the background to give it an oil painting feel.
Turbo LoRA types only work up to 1, but I heard you can do multiple style LoRAs.
