The example images are from the official WanAI.
We know the Lightx2v_Lora is extracted from the Lightx2v checkpoint that was released. So why not just use the Lightx2v checkpoint directly?
Usually checkpoints give better results than LoRAs.
We can skip the extra step of loading a LoRA, so it's faster and, in my opinion, the quality is better. Unfortunately, the originally released checkpoint is too bulky, and I couldn't find a related quantized model.
So, I decided to handle it myself, choosing the Q4_K_M quantization level because it's more balanced, faster, and still guarantees good generation quality.
Lightx2v (Checkpoint):
https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v
https://huggingface.co/lightx2v/Wan2.1-I2V-14B-720P-StepDistill-CfgDistill-Lightx2v
Description
FAQ
Comments (6)
Apparently you haven't seen what the Light X2V rank256 version can do with a Q8 checkpoint. It honestly makes even this Q4 checkpoint look like garbage in comparison. With GGUF there's no point in using Q4 unless you're running models from a slow hard drive or small system RAM, because OOM issues are a thing of the past due to SysRAM/VRAM swapping. There's no real speedup from running Q4 vs Q8.
The fact about Q4 is they're pretty much equivalent to using an FP8 checkpoint, and at that point you might as well use an FP8 checkpoint, due to the fact that LoRA's are faster with FP8 checkpoints than with quantized ones, due to the overhead associated with GGUF.
Compared to a Q8+LoRA, this model garbles the image and creates more of an almost stop motion effect like frames are missing. So this definitely isn't superior to using the LoRA.
One problem I noticed with the baked in LoRA, is it behaves as if you set the strength to 1.2 on a regular checkpoint, which is too high. So again, I think there is actually better quality from just using the LoRA.
In fact, the quality difference is soo bad, that running Q8 + LoRA at 3 steps is still better quality than running the Q4 at 4 steps, while being significantly faster.
So it's really Q8 or bust. In fact, I would say the ultimate model would be a Q8 version of this, with the full UMT5 clip quantized and VAE also baked into it. Doesn't matter how big it is, because it'll all be swapped between RAM anyways.
TLDR; you're not just quantizing WAN 2.1, you're quantizing the LightX2V optimizations that allow for lower steps, so that's why the LoRA works better, unless you use Q8 instead of Q4.
Your opinion is mostly correct, but not entirely. You talk about swapping, but that only works when you have high system memory. What you're overlooking is that not everyone in the world has a high-spec computer like you. This model is a realistic, runnable option for people with low-spec systems. In my view, this model has plenty of value for low-spec users, and I believe that’s exactly what the uploader intended. You’re seriously undervaluing this model’s worth.
use q8 u said, are you gonna give me the money for a new graphic card?
Hey PlK, Does this Q8 work good with a 3090ti? lightning t2v 4 step lora and it isn't bad quality but its still a bit noisy at 4 steps.
it works, faster and with good results, thank you
I tested this model while generating a few videos. Compared to the FusionX model that I'm using the most these days, I feel it's heterogeneous and overactive. In some ways, it's good or bad. I'm not sure in terms of quality. I can't evaluate it because my potato PC can't produce high-definition images. Compared to FusionX, it feels like it's not just a matter of good or bad, but a difference in characteristics.