For consistent scenes - train your character or product lora models in LoRA Pilot. Oh, by the way - cost of this 90s video was less than $2 (training+inference). https://t.co/SWEeWgJJZ2
I run Diffusion Lab, a Skool community for people working with ComfyUI and generative AI. Tutorials, workflow breakdowns and help when something gets stuck.
If you want to understand how things work, join us:
https://t.co/s9nftZmpqg
Different H3 model branches, so treat this as a practical workflow comparison. The video description links to the workflows and all 12 takes. What makes you reject a shot first: character drift, missed action or camera movement?
I’ve put together a 12-take LoRA vs RefMod comparison using R01, my orange robot character. RefMod held the face and claw closer in the tested close-ups, but missed some camera instructions.
https://t.co/vWuq37otX6
In Making R01 #8, I repair the endpoint, keep the good first frame and rerun the saved first/last-frame ComfyUI workflow. Then I check the movement and neighboring cuts.
I made 3 LoRA Pilot tutorials with R01, our orange robot:
- First session
- Camera prompts in ComfyUI
- Uploading and captioning a character dataset
I build LoRA Pilot. It’s free and open source
https://t.co/ix8wU0mdMk
This doesn't train a LoRA. It encodes images that H3 can use as references during generation. The download includes a six-view orange robot RefMod and an example Comfy UI workflow.
GitHub: https://t.co/N9ptbt3mPr
For those of you creating videos using Minimax H3 model I've put together RefMod Pilot, a free download for making and using H3 character references in ComfyUI.
https://t.co/qedzh7BQbs
A paper plane, clouds and a coastline. Let It Fly is a 45-second AI short made with LoRA Pilot, the free, open-source workspace I’m building for LoRA training and image/video generation.
Watch: https://t.co/cBUCbMzqiq
Showcase: https://t.co/tDzz1mwdSW
For consistent scenes - train your character or product lora models in LoRA Pilot. Oh, by the way - cost of this 90s video was less than $2 (training+inference). https://t.co/SWEeWgJJZ2
The most revealing trace of a civilization may not be the thing it built for eternity. It may be the ordinary thing it made too many times to disappear.
https://t.co/UZtXc01BJ8
Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work.
https://t.co/OoM83SyISN
Meet Doug. 74, bedridden, losing his sight… and still designing the future of human-AI interaction. This might be the most important project of my life.
https://t.co/PiDettoFuC