adbrasiandClaude Opus 4.7
ef936e10b9
model.onnx do deepghs/pixai-tagger-v0.9-onnx tem dois outputs: - output 0: (N, 1024) embedding do encoder EVA-02 - output 1: (N, 13461) logits das tags (head fundido) O codigo pegava outputs[0] (embedding) e quebrava com shape mismatch contra is_character (13461,). Agora seleciona o output cujo shape bate com num_classes (tags_arr.shape[0]), com fallback para o output de maior dimensao final. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
booru-helper-mini
ComfyUI custom node pack with three separate tagger nodes:
- WD14 Tagger — SmilingWolf WD14 models (ONNX)
- Camie Tagger —
Camais03/camie-tagger-v2(ONNX) - PixAI Tagger —
pixai-labs/pixai-tagger-v0.9(PyTorch, default) ordeepghs/pixai-tagger-v0.9-onnx(ONNX, no auth required)
Install in ComfyUI/custom_nodes/booru-helper-mini and run:
pip install -r requirements.txt
torch and torchvision are provided by ComfyUI itself; timm is only required for the PixAI PyTorch backend.
Features
- Three independent nodes — pick the model you need
- Real ONNX batching with user-defined
batch_size - LRU session caching (switching between WD14 variants reuses loaded sessions)
- Tuned
ort.SessionOptions+ CUDA provider options for fast inference - Vectorized threshold selection with numpy
- PixAI: FP16 support on CUDA, optional IP (copyright/series) mapping
- Output is a list of strings (one per image in the batch)
PixAI — HuggingFace token
The official PyTorch repo is gated. Set a token via env var before launching ComfyUI:
export HF_TOKEN="hf_..."
Or paste it into the hf_token field of the PixAI node. If you don't have access, switch the node's backend to onnx and it will use deepghs/pixai-tagger-v0.9-onnx instead (no auth needed).
Models directory
Models are stored under ./models/{wd14,camie,pixai}/ inside the node folder. If your ComfyUI has a wd14_tagger folder registered in folder_paths, that one is used as base instead.