is heun / linear_quadratic good for the lightx2v rank 256 model (not nvfp4)

#1
by Perfs - opened

sorry for asking there but i seek this info urgently as im stress testing and don't wanna do a grid test of every sampler/scheduler.
my needs are regarding strict preservation of face identity and the first frame prioritely, working on wan animate essentially. appreciate your answer, i could put results here when you answer.

No worries. To be honest, I'm no researcher, just a hobbyist. Unfortunately I lack the hardware resources to give you an answer, since I cannot run inference with such big models. My main motivation to create this quantization was because I wasn't satisfied with the Wan 2.1 NVFP4 models available. My hardware is really limited and cannot run inference with BF16. It also struggles with FP8.

The original weights I quantized were those from https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v. My guess is that they used a rank64 LoRA to produce that distillation, as https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v/tree/main/loras contains a rank64 LoRA, so I couldn't give you an answer for my quantiation, either.

No worries. To be honest, I'm no researcher, just a hobbyist. Unfortunately I lack the hardware resources to give you an answer, since I cannot run inference with such big models. My main motivation to create this quantization was because I wasn't satisfied with the Wan 2.1 NVFP4 models available. My hardware is really limited and cannot run inference with BF16. It also truggles with FP8.

The original weights I quantized were those from https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v. My guess is that they used a rank64 LoRA to produce that distillation, as https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v/tree/main/loras contains a rank64 LoRA, so I couldn't give you an answer for my quantiation, either.

alr, and in general do you know what sampler/scheduler is choosen over when plugging speed lora (4step) in wan workflow, working with wan animate and vace but with animate atm. appreciate any crucial infos

Sign up or log in to comment