Prompt Studio (TJ)

Prompt Studio (TJ) is the all-in-one of the TJ LLM family: it's both a Prompt Enhancer (rewrite your text into the right model dialect) and an Image to Prompt (turn a picture into text), with the mode picked automatically from what you have connected. Connect an image and it captions/analyzes it; leave the image socket empty and it behaves exactly like the enhancer. One node, two jobs, no mode-switching by hand.

How it works

The mode widget gives you Auto (default), Image to Prompt, or Prompt Enhancer. In Auto, the presence of the optional image input decides: image connected → vision mode, otherwise text rewrite. When it's doing vision, the vision_task dropdown sets what you want out of the picture:

• Caption (plain description) - just describe it.
• Caption + Format - describe it, then rewrite the caption through the model_format rules below.
• SD/Booru Tags - extract tag-style tokens, great for feeding back into tag-trained models.
• Pose & Anatomy Focus - analysis aimed at composition/pose detail.
• Content Quality Check - a review-style pass (feeds the pack's quality-gate workflows).
• Custom Instruction - your own ask in custom_instruction.

Everything else mirrors the Prompt Enhancer: purpose (Image/Video/Edit/Custom), model_format (Krea2, Flux natural language, booru tags, LTX prose, and friends), aesthetic style injection, raw_prompt plus raw_prompt_input for a wired-in prompt, extra_instructions and system_prompt_override. The LLM knobs are the same too - model_backend (GGUF/llama.cpp or ComfyUI TextGenerate), gguf_model, mmproj_file (needed for the vision mode, it's the multimodal projector), n_gpu_layers, n_ctx, max_tokens, temperature, top_p, repeat_penalty, seed, and lock_in.

Outputs: prompt (the result) and thought_process (reasoning trace). Wireless is built in via get_name_prompt / get_name_image and set_name, so a remote prompt or image can be received and the result published without wires.

Installing it

Standard pack install:

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

Then the LLM setup: llama-cpp-python (GPU build via build_llama_gemma4.bat / LLAMA_GPU_SETUP.md for speed and newest-model support), and GGUF files in ComfyUI/models/text_encoders/. For vision you need both the main GGUF and its mmproj projector file - the pack recommends Qwen2.5-VL 3B/7B (bartowski's quantized versions) or MiniCPM-V if you're tight on VRAM.

The honest take

Prompt Studio is the node to reach for if you do a lot of img2img and image-conditioned work - one node that captions a reference and rewrites prompts for the same model saves you from juggling two separate LLM setups. The gotcha is the usual one for this pack's LLM nodes: it's only as good as your GGUF setup, and a CPU-only llama-cpp build will make every vision call slow enough to annoy you. Get the GPU build right once and it's the rare single node that actually replaces two.