Qwen3.5-4B-Danbooru-Prompt-Generator

Expand a few Danbooru seed tags into a full image-generation prompt. Ships as a merged text-only Qwen3.5 model (BF16 + GGUF) and a ComfyUI node that talks to a model server over HTTP. The node is built for LM Studio: load the model there and the node handles the rest.

Key features

• Semantic negation. Prefix a concept with - (e.g. -nude) and the model avoids that concept and related tags.
• Danbooru tag output. Tags go in as a user chat message, comma-separated tags with underscores come out (looking_at_viewer).
• Rating control. Write rating:general or rating:explicit, exactly like that.
• Chat template. Use the supplied template with thinking/reasoning off. With thinking on, the model writes the tags inside its reasoning block and the final answer comes out empty.

Prompt format

Separate tags with , , use underscores in multi-word tags, prefer tags that exist on Danbooru.

1girl, long_hair, school_uniform, -nude, rating:general

Training data was mostly NSFW, so the model biases toward explicit content. For SFW output, include rating:general and negatives like -nude, then review the generated tags. These instructions do not guarantee SFW output.

Files

Text-only model — it does not take images as input. Base: techwithsergiu/Qwen3.5-text-4B-bnb-4bit, derived from Qwen/Qwen3.5-4B.

ComfyUI

1. Copy comfyui-prompt-generator/ into ComfyUI/custom_nodes/, then install its dependency with ComfyUI's Python and restart:

   python -m pip install -r custom_nodes/comfyui-prompt-generator/requirements.txt
   

   Windows portable, from its root folder:

   .\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\comfyui-prompt-generator\requirements.txt
   

2. Load the GGUF (or model/) in LM Studio, enable the local server, and use the supplied chat template with thinking off. The model files do not have to sit inside ComfyUI; the server can run on the same machine or another one reachable over the network.

3. Set host to the server's API base URL including /v1 (LM Studio default: http://127.0.0.1:1234/v1) and model to the identifier your server exposes.

Required API contract (LM Studio provides it out of the box) — an OpenAI-compatible chat endpoint alone is not enough:

• POST /v1/chat/completions — OpenAI-compatible chat.
• GET /api/v1/models — JSON with a models list; entries expose key/display_name and loaded_instances with instance ids.
• POST /api/v1/models/load — {"model": "<identifier>"}, returns the instance_id used for chat requests.

Node interface

Outputs: generated_prompt (generated tags after rating conversion and cleanup) and formatted_prompt (prefix + prompt + suffix joined with , ). Connect formatted_prompt to your positive prompt encoder.

The node maps rating:general → safe and rating:explicit → nsfw, explicit; other tags pass through. It requests at most 1,024 output tokens and drops the final fragment if the server reports a length cutoff. Rating conversion only rewrites output text — it does not filter content.

Transformers inference

python -m pip install "transformers>=5.5.0" accelerate safetensors

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("./model")
model = AutoModelForCausalLM.from_pretrained(
    "./model", dtype=torch.bfloat16, device_map="auto"
).eval()

messages = [{"role": "user", "content": "1girl, red_hair, -nude, rating:general"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=1.0)

print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip())

This returns the raw response, including any rating: tag; the node's rating conversion and prefix/suffix are separate post-processing. BF16 needs roughly 8.4 GB of weights plus KV cache — use the GGUF build for lower memory.

Troubleshooting

• Node missing: install requests into the Python env that runs ComfyUI, restart, check the startup log.
• Connection refused: host must end in /v1; check address, port and network reachability (note 127.0.0.1 is local to where ComfyUI runs).
• 404 or model-load error: the weight path is not the model identifier; match the server's model-management API.
• Empty response: disable thinking in the server's chat-template settings. With thinking on, the tags end up inside the reasoning block and the final answer is empty.
• Slow first request: the node allows up to 300 s for model loading and generation; reduce context or GPU offload on memory errors.

License

Apache-2.0. See the Apache License 2.0 and the upstream Qwen license.