These notes are related to a model description and, for the sake of better presentation, are published as an article. Please ignore it if you are not going to use the LoRA
Qwen-Image ( 20B parameters) struggles with NSFW generation - Flux.1 has 12B parameters.
LoRA fine-tuning ( 10-20M trainable parameters, rank 128) improves anatomy generation.
A small dataset of 100–300 images won't do it. A general-purpose NSFW LoRA requires a minimum of 1,500+ handpicked, high-quality images, carefully captioned, and trained with a minimum rank of 128. For specialized scenarios (like specific poses, medical details), a complementary, specialized LoRA would still be needed.
I merged 4 BETA LoRAs, each focused on a different aspect of training the LoRA. Their weights were carefully adjusted to create a unified LoRA. While not perfect, it provides a basics for testing tasks.
The code for merging LoRAs with adjustable weights is provided. You can use it to create a single, easy-to-use LoRA or mix it with your own LoRAs.
If you plan to mix it with your own or any other LoRA, first determine the LoRA structure and check if they are compatible
from safetensors import safe_open
lora_path = "/ComfyUI/models/loras/Qwen/Qwen-Beta.safetensors" # Replace with one of your LoRA paths
with safe_open(lora_path, framework="pt", device="cpu") as f:
for key in sorted(f.keys()):
print(key)If the result is a list of:
lora_unet_transformer_blocks_0_attn_add_k_proj.alpha
lora_unet_transformer_blocks_0_attn_add_k_proj.lora_down.weightyou're good to go. your LoRAs use a naming convention with .lora_down.weight for the down-projection (A matrix), .lora_up.weight for the up-projection (B matrix), and .alpha for the scaling factor, which is slightly different from the PEFT-style .lora_A.weight and .lora_B.weight
Rank-concatenation method: Output size assuming each LoRA has 500 MB x 4 = 2 GB
import torch
from safetensors import safe_open
from safetensors.torch import save_file
"""
Define your LoRAs and their weights, scaled to target ~1.0 strength in ComfyUI. The sum of all weights should not overpower the model. The sum should be in the range of 0.9 to 1.2. Below example weights will overpower the model, you need to minimize the model strength to 0.2 to achieve a good quality image if you run the script as is. Experiment.
"""
lora_recipes = [
("/ComfyUI/models/loras/Qwen/Qwen-NSFW.safetensors", 0.60),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta2.safetensors", 0.65),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta3.safetensors", 0.65),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta4.safetensors", 0.86),
]
# The output path for your new, combined "Meta-LoRA"
output_lora_path = "ComfyUI/models/loras/Qwen/Qwen-Image-merge.safetensors"
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
prefixes = set()
for lora_path, _ in lora_recipes:
with safe_open(lora_path, framework="pt", device="cpu") as f:
for key in f.keys():
if key.endswith(".lora_down.weight"):
prefix = key[:-len(".lora_down.weight")]
prefixes.add(prefix)
print(f"Found {len(prefixes)} unique adapted modules to merge.")
combined_state_dict = {}
for prefix in sorted(prefixes):
A_list = []
B_list = []
new_r = 0
in_features = None
out_features = None
has_alpha = False
for lora_path, weight in lora_recipes:
with safe_open(lora_path, framework="pt", device="cpu") as f:
A_key = f"{prefix}.lora_down.weight"
B_key = f"{prefix}.lora_up.weight"
alpha_key = f"{prefix}.alpha"
if A_key in f.keys() and B_key in f.keys():
A = f.get_tensor(A_key).to(device)
B = f.get_tensor(B_key).to(device)
r_i = A.shape[0]
assert A.shape == (r_i, A.shape[1]), f"Unexpected A shape in {lora_path}"
assert B.shape == (B.shape[0], r_i), f"Unexpected B shape in {lora_path}"
if in_features is None:
in_features = A.shape[1]
out_features = B.shape[0]
else:
assert in_features == A.shape[1], f"Dim mismatch in {lora_path}"
assert out_features == B.shape[0], f"Dim mismatch in {lora_path}"
if alpha_key in f.keys():
alpha = f.get_tensor(alpha_key).item()
scaling = alpha / r_i
has_alpha = True
else:
scaling = 1.0
s = weight * scaling
if s == 0:
continue
sqrt_s = torch.sqrt(torch.tensor(s)).to(device)
A = A * sqrt_s
B = B * sqrt_s
A_list.append(A)
B_list.append(B)
new_r += r_i
if new_r > 0:
combined_A = torch.cat(A_list, dim=0)
combined_B = torch.cat(B_list, dim=1)
combined_state_dict[f"{prefix}.lora_down.weight"] = combined_A.to("cpu")
combined_state_dict[f"{prefix}.lora_up.weight"] = combined_B.to("cpu")
if has_alpha:
combined_state_dict[f"{prefix}.alpha"] = torch.tensor(new_r)
print(f"Saving new combined LoRA to {output_lora_path}...")
save_file(combined_state_dict, output_lora_path)
print("Done! You now have a single, combined LoRA")Linear Combination method: Output size: 500MB Lora x4 = 500MB (risk: cross-term issues)
import torch
from safetensors import safe_open
from safetensors.torch import save_file
"""
Define your LoRAs and their weights, scaled to target ~1.0 strength in ComfyUI. The sum of all weights should not overpower the model. The sum should be in the range of 0.9 to 1.2. Below example weights will overpower the model, you need to minimize the model strength to 0.2 to achieve a good quality image if you run the script as is. Experiment.
"""
lora_recipes = [
("/ComfyUI/models/loras/Qwen/Qwen-NSFW.safetensors", 0.60),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta2.safetensors", 0.65),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta3.safetensors", 0.65),
("/ComfyUI/models/loras/Qwen/Qwen-NSFW-Beta4.safetensors", 0.86),
]
# The output path
output_lora_path = "/ComfyUI/models/loras/Qwen/NSFW_Meta-B4.safetensors"
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
keys = set()
for lora_path, _ in lora_recipes:
with safe_open(lora_path, framework="pt", device="cpu") as f:
keys.update(f.keys())
print(f"Found {len(keys)} unique keys to merge.")
combined_state_dict = {}
for lora_path, weight in lora_recipes:
print(f"Adding {lora_path} with weight {weight}...")
with safe_open(lora_path, framework="pt", device="cpu") as f:
for key in f.keys():
tensor = f.get_tensor(key).to(device)
if key in combined_state_dict:
combined_state_dict[key] += weight * tensor
else:
combined_state_dict[key] = weight * tensor
print(f"Saving new combined LoRA to {output_lora_path}...")
save_file({k: v.to("cpu") for k, v in combined_state_dict.items()}, output_lora_path)
print("Done! You now have a single linear merged LoRA.")Good luck!