Matrix Cinematography Style for H3
A cinematic style LoRA inspired by The Matrix: cool green-tinted color grading, crushed blacks, glossy highlights, and dramatic contrast with a hint of film grain. Designed to subtly influence visual tone while preserving the underlying subject and scene.
Suggested Generation Parameters
- Model: Minimax-H3 (T2V/FL2V)
- Resolution: 0.6mp
- Sampler / Steps: Euler-simple | 30 steps w/ Spectrum Node
- LoRA Strength: Recommended weight of 1.0
Training & Architecture Notes
- Hardware: Dual RTX 3090
- Pipeline: Modified diffusion-pipe with Ostris distillation adaptor
- Dataset: 37 images
Description
FAQ
Comments (9)
Looks good!!
Civitai's lora training for minimaxh3 defaults to like 3000 steps and 10 epochs. I feel like 3000 is wayyy too high. what do you recommend? im training with just 6 videos. Any advice would be much appreciated
For my two style loras I have only been using images, 40-60. 600-800 steps, lr 1e-4. I'm working out how to fit video training on my gpu's currently. From previous wan training, you do need a lot of steps for video whereas it learns from images much faster
@Luxflux Okay I guess it might be okay then. That's pretty cool you can train from images. I was curious bc my generation failed yesterday, generating 0 epochs after 12 hours. Maybe due to large queue. I'm trying again and will see how it goes, so far I got 1 epoch after 4 hours, not looking good lol
i train my lora with 140 image. 1500 steps and result is not good. Need +1000 maybe. 5.6 seconds per step.
Character lora: 4500 steps Ai-toolkit 10/20 images
@bhopping Check out fizgig's settings for his trainer on github/youtube. His trainer uses LoRA dim/alpha 16, 60 epochs, 0.25 MP, Training Structure Likeness and Style, adamw, flat 1e-4 https://github.com/shootthesound/Fizgig
@gambikules858 The more images you use, the more steps you will need, so that is expected. A higher quality dataset of fewer images trained for less steps will work just as effectively
great one , thanks