V4: the usual cfg1, 8-10 steps usually, not picky about samplers works with euler, lcm, sde samplers. Q4KM works, however more sensitive to banding sometimes, if you have the Vram use the Q8.
V3: similar as V2 but work usually with 8 steps and cfg 1.
V2: use Euler or Lcm with simple, beta at 10-12 steps, I tend to use cfg 1.3
A bit of an expirment to se how it goes and get a bit more face variation than the normal Qwen.
Contains part of lightx2v_Qwen-Image-Lightning-4step-8step-Merge plus 2 self trained Loras
Description
FAQ
Comments (6)
By chance are you considering quantitized gguf options?
Hello, I would like to know how you trained such a large model? The VRAM required for fine-tuning such a massive model must be enormous, right? Did you use LoRA fine-tuning and then integrate it into the main model?
My English is not very good, so I am using translation software to communicate. I apologize for any inconvenience and truly look forward to your response. Thank you!
Yes exactly, I did train some LoRAs with AI toolkit and then merged those in with ComfyUI.
先练LORA ，把LORA融合进大模型里。要找完全从大模型作微调的去看看那个白转，几乎没人这么干，成本很高，而且，从大模型微调后，现有的LORA很可能不适配了。得不偿失。