Limbicnation/deforum-prompt-lora-dataset

De Forum Cinematic Prompt Dataset

A specialized dataset for fine-tuning language models to generate cinematic video diffusion prompts in the style of "The Deforum Art Film".

Description

This dataset contains instruction-response pairs for training models to generate high-quality video diffusion prompts with:

• —Cinematic language and film terminology
• —De Forum aesthetic (noir, minimalist, art film style)
• —Technical parameters (aspect ratio, guidance scale, seeds)
• —Camera movements and lighting descriptions
• —Negative prompts for quality filtering

Dataset Structure

Fields

• —instruction: User request for prompt generation
• —response: Generated video diffusion prompt with cinematic descriptions
• —style_name: Scene identifier from the storyboard
• —negative_prompt: Terms to exclude from generation
• —tags: Scene categorization tags (noir, cinematic, psychological, etc.)
• —camera_movement: Recommended camera technique
• —technical_params: Dictionary with aspectratio, model, seed, guidancescale, steps
• —scene_context: Additional narrative context
• —text: Full formatted conversation with Qwen3 chat template

Example

{
  "instruction": "Generate a cinematic video prompt for:\nScene: INT. SARAH'S STUDIO - DAY\n...",
  "response": "Cinematic art film scene: INT. SARAH'S STUDIO - DAY...",
  "style_name": "INT. SARAH'S STUDIO - DAY",
  "negative_prompt": "blurry, static, distorted, modern elements...",
  "tags": ["noir", "minimalist", "psychological"],
  "camera_movement": "slow tracking shot following subject",
  "technical_params": {
    "aspect_ratio": "16:9",
    "model": "WanVideo",
    "seed": 4242,
    "guidance_scale": 7.5,
    "steps": 30
  },
  "scene_context": "Sarah wakes up to find herself surrounded by..."
}

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Limbicnation/deforum-prompt-lora-dataset")

# Access training split
train_data = dataset["train"]

# Access a sample
sample = train_data[0]
print(sample["instruction"])
print(sample["response"])

For Training with TRL

from trl import SFTTrainer
from datasets import load_dataset

dataset = load_dataset("Limbicnation/deforum-prompt-lora-dataset", split="train")

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    dataset_text_field="text",  # Pre-formatted with chat template
    ...
)

Data Source

This dataset is derived from "The Deforum Art Film" project storyboard materials, including:

• —Narrative scene descriptions
• —JSON frame sequences with timestamps
• —Visual style descriptors
• —Negative prompt collections

Intended Use

• —Fine-tuning language models for video diffusion prompt generation
• —Training models to understand cinematic language and camera terminology
• —Creating domain-specific LoRA adapters for the De Forum aesthetic

Limitations

• —Specialized for noir/art film aesthetic - may not generalize to all video styles
• —Based on a single narrative work - limited diversity in source material
• —Optimized for Qwen3 chat template - other formats may require adaptation

Citation

@dataset{deforum-prompt-lora-dataset,
  author = {Limbicnation},
  title = {De Forum Cinematic Prompt Dataset},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/Limbicnation/deforum-prompt-lora-dataset}}
}

License

MIT License - See LICENSE file for details.