The Stable Diffusion community (via Kohya_ss GUI or EveryDream2) famously splits LoRA weights into two files:
Inside lora_unet_up.pt, the actual weight tensor is frequently keyed as up-param.bin. If you unpack these archives, you will find dictionaries where "up-param.bin" maps directly to the up-projection matrix for the U-Net attention layers.
If you find a folder with up-param.bin but no adapter_model.bin, you can convert it by constructing a state dictionary:
state_dict =
"base_model.model.model.layers.0.self_attn.q_proj.lora_B.weight": load_up_0,
"base_model.model.model.layers.0.self_attn.v_proj.lora_B.weight": load_up_1,
# ... match all modules
peft_model.load_state_dict(state_dict, strict=True)
The name up-param.bin is derived from the internal architecture of neural network layers, specifically Linear (Dense) layers that are modified using low-rank decomposition. up-param.bin
To understand the "Up," we must first recall the basic forward pass of a linear layer:
Output = Input × Weight_Matrix + Bias
In Parameter-Efficient Fine-Tuning (PEFT), specifically LoRA, we don't modify the original Weight_Matrix directly. Instead, we inject a pair of smaller matrices: A (Down) and B (Up).
The mathematical update is:
W_final = W_original + (Up_Matrix @ Down_Matrix) The Stable Diffusion community (via Kohya_ss GUI or
Hence, up-param.bin contains the weights of the "Upscale" (or Up-projection) matrix. If you see both down-param.bin and up-param.bin, you are looking at a classic LoRA adapter in its raw, unpacked form before being merged into the base model.
up-param.bin never works alone. It is part of a triad:
If you delete up-param.bin while keeping down-param.bin, the adapter is dead. The output dimension math will fail catastrophically. Inside lora_unet_up
You do not need special software to interact with this file—just Python. Here is a standard forensic workflow to inspect the file without loading it into a full model pipeline.
Step 1: Verify the file type
file up-param.bin
If it returns data, it is likely a raw PyTorch pickle. If it returns NumPy data, it is a raw array.
Step 2: Load and inspect (Python snippet)
import torch
import numpy as np
2006-2026 © ВПК.name. Все права защищены.
При перепечатке ссылка на ВПК.name обязательна.
ВПК.name on english
Пользовательское соглашение.
№ 21-PP в реестре распространителей информации.
Хостинг сайта: TimeWeb.