Best way to use H3 ref2video. Not even joking. it just works and is easier to manage than any replacing attempts fighting the Model
Just draw your composition and let Minimax figure it out instead of using references from the internet
Proof1 community thread
Key points
- I noticed when trying to replace people, the Model struggles a lot when it is physically not straightforward to fit something in.
- Like when your hair has a tied bun behind your head, Minimax struggles to lie you on the floor.
- You have to tell Minimax to untie the hair.
- Which gets really difficult when it is very complex with a lot of minor conflicts between the original and replacement subject
Sources (1)
- [1]Best way to use H3 ref2video. Not even joking. it just works and is easier to manage than any replacing attempts fighting the Modelr/StableDiffusion (top, day) · Oct 11, 04:12 PM
Just draw your composition and let Minimax figure it out instead of using references from the internet
I noticed when trying to replace people, the Model struggles a lot when it is physically not straightforward to fit something in.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026How to pick the right models and dramatically speed up image and video generation speeds & What I wish I knew when starting on AMD hardware with ComfyUI
- Oct 10, 2026Krea2 Turbo Distill 2 step LoRA - FINAL checkpoint released (chk51195)
- Oct 8, 2026The Epoch Brief - October 8, 2026
- Oct 8, 2026Text is so 2023
- Oct 8, 2026Minimax Gaussian Mechanisms for Continual Machine Unlearning
- Oct 8, 2026WorldGuide: Goal-Directed Video World Model for Procedural Task Execution