NSFW-Wan-UMT5-XXL-V2

Extended UMT5 text encoder for Wan 2.2 video. 767 new tokens + LoRA on encoder layers 21–23.

Repo layout

encoders/
  nsfw_wan_umt5_xxl_v2_fp16.safetensors              # main fp16
  nsfw_wan_umt5_xxl_v2_fp32.safetensors              # highest precision
  nsfw_wan_umt5_xxl_v2_fp8_e4m3fn_scaled.safetensors # fp8 quantized
  nsfw_wan_umt5_xxl_v2_bf16.safetensors              # bf16 balanced
  nsfw_wan_umt5_xxl_v2_int8-convrot.safetensors      # int8 convrot (best quality among INT8 variants)
  nsfw_wan_umt5_xxl_v2_int4_convrot.safetensors      # int4 convrot (fastest)
  nsfw_wan_umt5_xxl_v2_int8.safetensors              # int8 standard
  
new_tokens_v4.json                                   # 767 extended token strings
config.json                                          # vocab_size 257067
tokenizer/                                           # extended spiece
load_example.py                                      # non-ComfyUI load script
ComfyUI-UMT5-Extended/                               # ComfyUI custom node

Vocabulary

Metric	Stock UMT5	This model
`shared.weight` rows	256,384	257,067
New tokens	—	767

Word	Stock (pieces)	Extended (pieces)
`thighjob`	4	1 (id 256937)
`titfuck`	2	1 (id 256955)
`fellatio`	2	1 (id 256511)
`paizuri`	2	1 (id 256687)
`deepthroat`	2	1 (id 256444)
`cowgirl`	2	1 (id 256409)
`blowjob`	1	1 (already stock)
`penis` / `cock`	1 each	1 each

Training stats

621 hand-written captions, 192 chars avg
LoRA: r=8, layers 21–23 q/v, 3.5M trainable params
V2 vs NSFW-API cosine (same prompt): blowjob 0.91, thighjob 0.88, titfuck 0.84

ComfyUI

Download & unzip ComfyUI-UMT5-Extended/ -> ComfyUI/custom_nodes/
encoders/*.safetensors -> ComfyUI/models/text_encoders/UMT5_extended/
Restart ComfyUI
CLIP Loader -> wan -> nsfw_wan_umt5_xxl_v2_fp16.safetensors (or fp8)

Custom node: ComfyUI hardcodes vocab_size: 256384. Extended files need 257067. The node sets this from shared.weight shape automatically.

Manual (no node): edit comfy/text_encoders/umt5_config_xxl.json -> "vocab_size": 257067 (revert to 256384 for stock encoders).

Not using ComfyUI

No custom node. You need vocab_size=257067 and the extended tokenizer.

pip install torch safetensors transformers
python load_example.py

rzgar/NSFW-Wan-UMT5-XXL-V2