TQD Score Estimate (TJ)

Label your dataset's strengths instead of deleting its weaknesses

TQD Score Estimate (TJ) scores each image in a batch with a local vision LLM - giving it a structure_score (composition/pose) and a detail_score (face/detail) - and writes out the exact files a Krea2 TQD LoRA training run needs. TQD stands for Timestep-aware Quality Decoupling, a training approach that doesn't throw away imperfect images but instead exposes each one to the diffusion timesteps where it's strong. This node is the dataset-prep half of that: it's the thing that turns a pile of images into scored, captioned training data.

Straight up: this is an experimental node, marked as still in testing by the author, and it exists to feed one specific training pipeline (the author's musubi-tuner-gui tj-custom branch). If you're not training Krea2 LoRAs with TQD, you don't need this node. If you are, it replaces a miserable manual step - hand-scoring thousands of images - with a local model call.

How it works

The scoring runs through a local Vision-LLM via one of two model_backend options:

• GGUF / llama.cpp - needs a GGUF vision model plus an mmproj_file projector (same setup as the pack's other LLM nodes; Qwen2.5-VL GGUF is the recommended family). chat_handler auto-detects.
• ComfyUI TextGenerate - the alternative backend using ComfyUI's text-generate nodes (text_encoder_name like gemma4_e4b_it_fp8_scaled.safetensors, with a clip_loader_type picker).

Key behaviors worth knowing from the source:

• Images are scored one at a time (deliberately never batched) to protect VRAM.
• It skips images already present in the scores file and upserts re-scored images rather than appending duplicates.
• temperature defaults to 0 - deterministic scoring, which is what you want for dataset labeling.
• save_results writes the image, a .txt caption sidecar (generate_caption), and tqd_scores.jsonl into output_dir. Each JSONL line has exactly image_file, structure_score, detail_score.

The outputs that matter

Six of them, all list-type except preview_text: structure_score, detail_score, reasoning, caption, jsonl_line, and preview_text (a human-readable summary shown in the node). Practically, you'll use the scores/JSONL outputs if you want to wire the results somewhere, but the real product is the files written to output_dir.

Install it

This node is where the pack's LLM dependencies bite. Install the pack first:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE.git

Or via ComfyUI Manager → search TJ_NODE. Then for the GGUF backend you need llama-cpp-python - the default pip install is CPU-only and slow, and may not load modern vision models. GPU setup means building from source (the pack ships build_llama_gemma4.bat and LLAMA_GPU_SETUP.md covers a CMAKE_ARGS="-DGGML_CUDA=on" build). Models go in ComfyUI/models/text_encoders/ - main GGUF + mmproj both required for vision.

Where people get burned

The obvious one: a vision-handler error on the GGUF backend, which almost always means llama-cpp-python is too old or CPU-only - reinstall with GPU support. Second, the scores are only as good as the model, and the defaults are tuned for the author's uncensored Qwen3.5 4B - swap in a stronger vision model (Qwen2.5-VL 7B is the README's recommended) if your scoring feels mushy. And since this is explicitly experimental, keep the auto-verification it does (image count vs score count mismatch checks) turned on - a broken scores file at training time is a much worse failure than a warning here.