我来做点什么吧!ComfyUI教程都来看,别瞎几把搞了。
#7
by deleted - opened
首先对齐diffusion_model,到处都是转换脚本自己去找一个。
节点直接调用官方 comfy.sd.load_bypass_lora_for_models(),与作者的运行时 base(x) + B(A(x)) 语义一致。
所以普通的LORA加载器是调用权重补丁路径,所以根本没鸟用!
正确的加载组件是 Load LoRA (Bypass, Model Only) (for debugging)
源码上真正对应作者语义的是:
class LoraLoaderBypass:
"""
Apply LoRA in bypass mode without modifying base model weights.
Bypass mode computes: output = base_forward(x) + lora_path(x)
采用运行时旁路:base(x) + B(A(x))
实测测试4步必须用minimax_h3_fl2va_int8_convrot.safetensors和minimax_h3_ref2va_int8_convrot.safetensors
搞不定的别喷我
对齐rank吗