SlopPrompt

Slop Prompt v5

A fine-tuned Qwen3.5 4B model for enhancing AI image-generation prompts. It was trained on a multi-format anime-art caption dataset and supports 11 conversion directions between Danbooru tags, short phrases, long descriptive paragraphs, and structured JSON.

Model variants

Base model: MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2

Dataset credit

Training data derived from Anime Art Multicaptions v5.0 by Minthy
https://huggingface.co/datasets/Minthy/Anime-Art-Multicaptions-v5.0

All examples were anonymized: character, artist, source, and copyright names were removed from inputs and targets.

What the model can do

The model was trained on 11 prompt-conversion modes. Each mode has its own system-prompt instructions. Use the exact system prompt for the direction you want.

Training system prompts

Use these as the system message for each mode.

Base system prefix (prepended to every mode):
You are a prompt enhancer for AI image generation. Do not include real names of characters, artists, sources, or copyrights in your output. Use generic descriptions instead.

What to expect

• Strong on the 11 trained directions. It follows the system prompt, produces fluent English, and keeps outputs free of real names.
• Cross-language input is understood. You can write the user message in French, Spanish, etc., and it will usually answer in English as instructed.
• Hallucination mode invents plausible missing details from an incomplete tag list.
• JSON outputs are generally valid and include keys such as character, background, texts, and atmosphere.
• Output length: By default it writes one or two paragraphs (~150–250 words) and then stops. For longer outputs, force a minimum token count (see parameters below).
• Reverse tag extraction (description -> tags) was not trained. If you ask for that, it will usually return short sentences instead of Danbooru tags.

Vision support

The base model is a multimodal Qwen3.5 checkpoint, and the published Hugging Face folder still contains the original vision encoder weights. You can load it with:

from transformers import AutoModelForImageTextToText, AutoTokenizer

model = AutoModelForImageTextToText.from_pretrained(
    "SlopPrompt-v5",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("SlopPrompt-v5", trust_remote_code=True)

The vision backbone is unchanged from the base model. You can pass images through it, but the model was not specifically fine-tuned for vision-to-prompt tasks.

The GGUF releases are text-only.

Recommended generation parameters

These settings work well for most modes.

max_new_tokens = 512
temperature = 0.7
top_p = 0.9
repetition_penalty = 1.05
no_repeat_ngram_size = 5

For deterministic output, set do_sample=False.

For longer output, set:

min_new_tokens = 400
max_new_tokens = 1536
repetition_penalty = 1.01
no_repeat_ngram_size = 2

The model will then produce ~300–400 words before stopping.

Usage with Hugging Face Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "SlopPrompt-v5"
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

system = (
    "You are a prompt enhancer for AI image generation. "
    "Do not include real names of characters, artists, sources, or copyrights in your output. "
    "Use generic descriptions instead.\n\n"
    "You are given a partial, incomplete list of Danbooru tags. "
    "Reconstruct and enhance the missing details, then write a detailed, descriptive paragraph."
)
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": "1girl, loli, at park, night, sitting on bench, dress"},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,  # important for clean output
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.05,
    no_repeat_ngram_size=5,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Usage with llama.cpp (GGUF)

The GGUF was verified with a current llama.cpp build. Run it in single-turn mode with reasoning disabled.

PROMPT='<|im_start|>system
You are a prompt enhancer for AI image generation. Do not include real names of characters, artists, sources, or copyrights in your output. Use generic descriptions instead.

You are given a partial, incomplete list of Danbooru tags. Reconstruct and enhance the missing details, then write a detailed, descriptive paragraph.
<|im_end|>
<|im_start|>user
1girl, loli, at park, night, sitting on bench, dress
<|im_end|>
<|im_start|>assistant
'

llama-cli \
  -m SlopPrompt-v5-Q8_0.gguf \
  -p "$PROMPT" \
  -n 512 \
  -t 16 \
  -ngl 99 \
  --temp 0.7 \
  --top-p 0.9 \
  --repeat-penalty 1.05 \
  --no-display-prompt \
  --no-conversation \
  --single-turn \
  --reasoning off

On an RTX-class GPU expect ~200+ tokens/s for Q8_0 and ~150 tokens/s for BF16.

File layout

publish/
├── README.md
├── SlopPrompt-v5/                 # HF Transformers BF16 safetensors (text + vision)
│   ├── config.json
│   ├── generation_config.json
│   ├── chat_template.jinja
│   ├── tokenizer_config.json
│   ├── tokenizer.json
│   └── model.safetensors
├── SlopPrompt-v5-BF16.gguf        # llama.cpp BF16
└── SlopPrompt-v5-Q8_0.gguf        # llama.cpp Q8_0

Limitations

• Trained on English anime-art captions. Other languages may work as input but output is expected to be English.
• Does not reliably perform reverse extraction such as long -> tags because that mode was not in the training set.
• May hallucinate extra details in reconstruction/hallucination modes; use a lower temperature or greedy decoding for more conservative outputs.

Disclaimer

This model generates text based on patterns learned from training data. The authors provide no guarantees about accuracy, safety, appropriateness, or fitness for any particular purpose. Outputs may be unexpected, inaccurate, or inconsistent with the provided system prompt. Review and filter outputs before using them in production or for downstream image generation.

Credits & license

• Base model: MuXodious/Qwen3.5-4B-SOMPOA-heresy-v2
• Dataset: Minthy/Anime-Art-Multicaptions-v5.0
• Released under the terms of the base model and dataset licenses.