I’ve been working with I2V video generation and noticed that the newer 2.2 Lightning LoRAs make my Wan 2.2 videos appear less fluid and more rigid than expected. After some experimentation, I discovered that stacking the previous 2.1 Lightning LoRA on top of the newer version helps restore more natural movement patterns.
Here’s my current configuration:
HIGH RESOLUTION (I’m using the 720p variant, though the 480p option works well for lower resolution projects):
Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16.safetensors (Set strength to 1)
Combined with:
Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors (Set strength to 3)
LOW RESOLUTION:
Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors (Set strength to 1)
Has anyone else tried this layering approach? I’m curious about your results and whether different strength values work better for specific use cases.
Hit this exact setup at work debugging motion artifacts for a client. Your strength combo is perfect - we landed on almost identical values after testing tons of configurations.
Production tip: 2.1 distill LoRA at strength 3 works great for organic stuff like water, cloth, hair, but drop it to 1.8-2.2 for mechanical objects or buildings. Higher strength makes rigid things look like jelly.
Pre-warm both LoRAs in VRAM before generating. Prevents those weird stutters in the first few frames. Takes 30 seconds but beats regenerating when your opening looks janky.
Dual LoRA loading adds maybe 15% to generation time but quality improvement is huge. Way better than tweaking individual parameters for hours or tossing bad outputs.
interesting find! I’ve been dealing with the same choppy motion issue on 2.2 for weeks but never thought to layer LoRAs like that. I’ll try your setup, though I might start with lower strength on the 2.1 - that 3 setting seems pretty aggressive. are you seeing performance drops during generation or just slower load times?
Been using this dual LoRA setup for three weeks - results are solid. The 2.1 distill fills temporal gaps that 2.2 Lightning misses. That rigidity you mentioned? It’s the newer model trading motion continuity for speed. Memory allocation matters way more than I expected. Getting inconsistent results? Clear your VRAM between generations. The model hangs onto previous motion data and creates jarring transitions. Input images react differently to strength settings. High contrast images with sharp edges need 2.1 strength around 2.4 - otherwise you get fake-looking motion blur. Portraits handle full strength 3 no problem. Computational overhead is nothing on modern hardware, but the quality boost is huge. This is my go-to config now.
Nice work on the LoRA stacking! I had the same problem with 2.2 - movements looked way too snappy and fake. Your strength settings are pretty aggressive though. I’ve been using 2.1 distill at 1.5 strength instead of 3, which gives me smoother motion without those weird warping artifacts I kept getting at higher values. Makes sense to combine versions since 2.1’s motion data has way better temporal consistency. Running this on a 3090 and inference time barely changes even with dual LoRA load. Have you tried camera pans or object rotations yet? That’s usually where the rigidity problems show up worst.
Hit this same bottleneck last month when we were cranking out promo videos. Manually tweaking LoRA configs and testing strength values was killing us time-wise.
Fixed it by automating everything. Built a workflow that spots motion quality problems in generated frames and tweaks LoRA combinations automatically. Tests strength values from 1-3 on the 2.1 distill LoRA until it finds what works for each input.
It handles batch processing too - queue up multiple I2V generations with different settings and let it run overnight. No more babysitting or endless manual adjustments.
For camera pan testing, automation’s perfect since it catches motion artifacts and fixes parameters before you see the crappy output.
Took me 2 hours to set up and saves our team 20+ hours weekly on video generation. Beats manually guessing at strength values.
Used this automation platform: https://latenode.com
