Updated: Jun 2, 2025
concept
i2v version updated to v3.0
Improved the issue of slow motion when characters walk, fixing it to a normal walking speed
i2v version updated to v2.0
This version has optimized the stability of walking, avoiding strange walking movements and adding some hand movements
The i2v_v2.0 version also adds some hand movements, such as:
adjusted her hair with handOr
hands on hipsI tried using Wan2.1 to generate videos of characters walking behind their backs but was unsuccessful. There were always instances of people walking backwards or making strange movements. So I trained this Lora, which will improve the success rate of generating walking back videos.
Trigger word:
walking from behindPerhaps increasing the number of sampling steps can lead to improvement
Discussion
Hi, @ivever! Can you please share how have you trained this? With what software? Did you use images or videos, with which sizes and duration? Did you train for 720p i2v? How much vram was required for training, how much time?
I'm asking because I want LoRAs for animated character walking cycle animations(front, right, back, front right, back right etc) to create sprite sheet assets for video game from one input image.
I've noticed that even with 0.85 strength, the LoRA tends to make the subject to wear shoes (specially high heels), and with colored women, it tends to use the same color skin as color of the shoes. So if anybody plans to use this with barefoot subjects, my recommendation is to add "barefoot" in the + prompt and "shoes" and "high heels" in the -.
For me it often morphs the womans skin into a beige dress and it changes the background, looks like the source videos it was trained on maybe? idk cfg is around 5... gguf Q8 models and 720p resolution