license: apache-2.0
tags:
- text-to-image
- lora
- diffusers
- template:diffusion-lora
base_model:
- Wan-AI/Wan2.2-I2V-A14B
Wan2.2 I2V LightX2V 2Step
Fast LightX2V LoRAs for Wan2.2 I2V image-to-video.
Tuned for 2 steps high + 2 steps low setup.
| File | Load on |
|---|---|
Wan2.2_I2V_LightX2V_2step_high_noise.safetensors |
High-noise |
Wan2.2_I2V_LightX2V_2step_high_noise.safetensors |
Low-noise |
Strength: 1.0 on both.
Why use this?
Standard LightX2V workflows run 4 steps per stage. These LoRAs are tuned so you can run 2 steps per stage and still get usable motion and prompt follow at roughly half the 4Step time.
On my hardware, 81 f/s:
| Setup | Resolution | Steps | Avg. time |
|---|---|---|---|
| Default LightX2V | 512×848 | 4 + 4 | ~89 s |
| This LoRA | 512×848 | 2 + 2 | ~37 s |
The time we spend on a 512×848 clip with 4Step LightX2V is about what it needs for 720×1280 with this pair.
Best when you need many previews, prompt tests, or seed hunting, and care more about iteration speed than maximum fine detail.
Text-to-Video style: input image | Prompt
Recommended sampling
Schedulers that work well
| Sampler | Scheduler |
|---|---|
dpmpp_2m_sde |
sgm_uniform |
dpmpp_2m_sde_heun |
sgm_uniform |
sa_solver_pece |
sgm_uniform |
Strength: 1.0 high · 1.0 low
Steps: 2 high · 2 low
Euler / Simple
euler + simple still works, but reliability drops. Expect more duds, often about 2 good out of 4 runs. Prefer the samplers above when you want fewer re-rolls.
Quality expectations
From a small random-seed test set (9 generations):
| Outcome | Count | Notes |
|---|---|---|
| Clean motion, on-prompt, no body deformation | 6 / 9 | Usable “hero” candidates |
| Mild motion issues | 1 / 9 | Often still salvageable |
| Strong motion error | 2 / 9 | Reroll seed |
So: most seeds are fine, a minority will need a new seed. That trade-off is kinda acceptable for speed.
Compared with full 4-step LightX2V you may lose some fine detail and consistency. For final hero frames, you can still switch back to a 4-step stack.