[English] This is a quantized GGUF version of the NSFW-Wan-UMT5-XXL text encoder. It is designed to be used with the Wan 2.1 video generation model in ComfyUI.
File Guide :
Pruned Model fp16 (3.4GB) = Q4_K_M (Recommended for 12GB VRAM)
Pruned Model bf16 (3.8GB) = Q5_K_M (Higher Quality)
Why use this?
Uncensored: Fixes the issue where the official Wan 2.1 model refuses to generate NSFW or specific prompts (it plays dumb).
Lightweight: GGUF format allows you to run this huge T5 encoder with much less VRAM (Q4/Q5 versions).
How to use:
Place the
.gguffile in yourComfyUI/models/text_encoders/folder.Load it using the CLIPLoader (GGUF) node.
Set the type to
wan.
Credits & Disclaimer:
Original Weights: NSFW-API/NSFW-Wan-UMT5-XXL on HuggingFace.
Conversion: Quantized by MomusAki.
If the original author (NSFW-API) has any objections to this quantization release, please contact me and I will take it down immediately.
[中文] 这是 NSFW-Wan-UMT5-XXL 文本编码器的 GGUF 量化版本。 专为 ComfyUI 中的 Wan 2.1 视频生成模型打造。
文件说明:
Pruned Model fp16 (3.4GB) = Q4_K_M (Recommended for 12GB VRAM)
Pruned Model bf16 (3.8GB) = Q5_K_M (Higher Quality)
主要作用：
解除限制： 完美解决了 Wan 2.1 官方模型对 NSFW 或特定敏感提示词“装傻”、“听不懂”的问题。
节省显存： 相比原始的 safetensors 格式，GGUF 版本（Q4/Q5）大幅降低了显存占用，让 12G/16G 显卡也能流畅运行。
使用方法：
将
.gguf文件放入ComfyUI/models/text_encoders/文件夹。在 ComfyUI 中使用 CLIPLoader (GGUF) 节点加载。
Type 选项选择
wan。
致谢与声明：
原始权重： 源自 HuggingFace 上的 NSFW-API/NSFW-Wan-UMT5-XXL。
量化制作： 由 MomusAki 转换。
本模型仅为方便社区使用的量化版本。如果原作者 (NSFW-API) 认为此发布不妥，请联系我，我会立即下架。
Description
FAQ
Comments (19)
Great! thank you so much!!
Q8?
To be honest, the Q8 version is almost the same size as the original BF16, so I feel it might be better to just use the original if you have the resources.
But I'm open to it! If I see more than 5 requests for Q8 in this thread, I'll bake one and upload it for you guys.
+1
the fact is that for some hardware gguf is better managed than saftensors, so it's not allways a question of size...
@MomusAki plz bake one.
+1 Q8 please
you mentioned DualCLIPLoader, does this have to run along side with the umt5-xxxl encoder? or can it be loaded by itself.
This GGUF file IS the UMT5-XXL encoder itself. You don't need to run it alongside the original one. It replaces the original huge file.
You can simply use the CLIPLoader (GGUF) node to load it by itself. If you use DualCLIPLoader, just select this file for the T5 slot and leave the other slot empty (or load CLIP-L if your specific workflow needs it for I2V vision).
@MomusAki In this case description provided is inconsistent with reality. It literally says to use DualCLIPLoader node.
@wisieeen293 It was my oversight; I have already revised the description.
can this work with normal checkpoint which are not guff models?
You need to use a node that supports GGUF, for example, by replacing the CLIPLoader node with CLIPLoader(gguf).
@MomusAki i did but the erros is till happening check my other comment
can u add worflow as well as beacause in my case its showing
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
gguf qtypes: Q4_K (144), F32 (49), F16 (24), Q6_K (25)
Attempting to rebuild sentencepiece tokenizer from metadata..
!!! Exception during processing !!! only 0-dimensional arrays can be converted to Python scalars
Traceback (most recent call last):
File "C:\Users\sache\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 530, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\sache\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 334, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\sache\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 308, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "C:\Users\sache\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 296, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "C:\Users\sache\Documents\ComfyUI\custom_nodes\gguf\pig.py", line 632, in load_clip
return (self.load_patcher([clip_path], get_clip_type(type), self.load_data([clip_path])), get_device('default'))
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\sache\Documents\ComfyUI\custom_nodes\gguf\pig.py", line 613, in load_data
sd = load_gguf_clip(p)
^^^^^^^^^^^^^^^^^
File "C:\Users\sache\Documents\ComfyUI\custom_nodes\gguf\pig.py", line 483, in load_gguf_clip
sd['spiece_model'] = tokenizer_builder(path)
^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\sache\Documents\ComfyUI\custom_nodes\gguf\gguf_connector\tkn.py", line 51, in tokenizer_builder
spm.trainer_spec.eos_id = get_field(reader,
^^^^^^^^^^^^^^^^^
File "C:\Users\sache\Documents\ComfyUI\custom_nodes\gguf\gguf_connector\tkn.py", line 12, in get_field
return field_type(field.parts[field.data[-1]])
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: only 0-dimensional arrays can be converted to Python scalars
Prompt executed in 13.66 seconds
@MomusAki its works with the simple gguf and the diffusion models has to be gguf as well as not the safetenor model
they being 2 in 1 sort of breaks wan2gp