ThalisAI/graveyard_ghoulz

Graveyard Ghoulz — Qwen-Image Fine-Tune

• Base Model: Qwen-Image
• Source Dataset: Images generated using Illustrious-series checkpoints
• Training Method: Full fine-tune with dual-prompt dataset (ToriiGate natural-language + WD14 tags)
• Latest Snapshot: graveyard_ghoulz_musubi_fine_v6_epoch000005_fp8.safetensors

Overview

Graveyard Ghoulz is a fine-tuned derivative of Qwen-Image, specializing in bold, character-driven illustration. The model inherits Qwen-Image’s multimodal fidelity and has been retrained on a curated dataset of images rendered with Illustrious checkpoints, blending anime and toon stylization with the cleaner, more articulated composition of modern CGI workflows.

The resulting style emphasizes strong outlines, flat-color precision, and crisp edge definition, suitable for both stylized pin-ups and expressive cartoon scenes. Faces, hair, and figure proportions are rendered with high consistency, while linework remains sharp even at large resolutions.

⚗️ Experimental Notice

This model represents an early experiment in fine-tuning Qwen-Image on a single, locally-hosted GPU. It serves as a qualitative benchmark for small-scale adaptation and as a feasibility test for limited-hardware training. Expect artifacts, variable quality in finer details, and inconsistent hands or anatomy in some generations.

Qualitative Evaluation

Criterion	Rating	Notes
Style	✅ Good	Strong line art, consistent toon shading
Hands	⚠️ OK	Occasional distortion and extra fingers
Genitals	⚠️ OK	Reasonable structure, sometimes lacking detail
Nipples	✅ Good	Detailed, stylistically coherent, well-shaped

Content Profile

• Primary content: NSFW / erotic illustration
• Typical tone: Softcore to explicit, with emphasis on sensual anatomy, teasing expressions, and minimal censorship
• Range: Includes many SFW and softcore scenes for general illustration; hardcore content is present but less frequent
• Ethical note: All material was created synthetically and adheres to platform safety rules

Intended Use

• Character and pin-up illustration
• NSFW and aesthetic anime render workflows
• Dataset generation for downstream fine-tunes or LoRA adapters

Prompting Style

The dataset was duplicated and captioned in two modes:

• ToriiGate natural-language prompts for expressive, sentence-based conditioning
• WD14 tag prompts for tag-based control and style matching

This dual encoding improves compatibility with both freeform and tag-driven prompting styles, making the model versatile across captioning pipelines and control UIs.

How to Prompt Effectively

Prompts can be as short as a few words (e.g., 1girl, solo), but image quality and composition depth improve significantly with length. Short prompts often result in flatter lighting, less engaging poses, and a higher chance of artifacts or missing details.

WD14 tag prompts work well when formatted traditionally, such as: 1girl, breasts, nipples, solo, green eyes, long hair, white background, simple background, blush, upper body, large breasts, closed mouth, looking at viewer, collarbone, shirt, bare shoulders, hair between eyes, white shirt, off shoulder, white hair

These concise, comma-separated tags yield images that align closely with the dataset’s structure. The 1girl tag is present in most captions and helps anchor compositions to the model’s most consistent archetype.

ToriiGate natural-language prompts, by contrast, describe scenes in complete sentences and carry a richer sense of narrative context. They allow Qwen-Image’s text encoder to interpret subtle details such as mood, fabric texture, or emotional tone. A typical example might begin:

a close-up portrait of a young woman with long, flowing white hair and striking green eyes. Her hair is styled in a way that frames her face, with some strands falling between her eyes. She has a slight blush on her cheeks, giving her a coy and alluring expression as she looks directly at the viewer. The woman's outfit is quite revealing, with her shirt open to expose her large, perky breasts. Her nipples are clearly visible through the thin fabric of her shirt. The shirt itself is white and appears to be made of a soft, silky material. The way the shirt is open suggests a sense of vulnerability and sensuality. The woman's pose is intimate and inviting, with her body slightly tilted towards the viewer. Her breasts are the focal point of the image, drawing the eye with their size and shape. The way the shirt is open also reveals a hint of her cleavage, adding to the overall erotic atmosphere of the image. The background of the image is a simple white, putting all the focus on the woman and her alluring appearance. The simplicity of the background also serves to highlight the woman's features and the details of her body. Overall, the image has a sensual and provocative atmosphere, with the woman's pose, expression, and partially exposed breasts creating a sense of intimacy and desire. The soft, silky material of her shirt adds to the erotic feel of the image, while the woman's direct gaze and slight blush suggest a sense of confidence and allure.

While the WD14 style provides reliability and control for structured datasets, the ToriiGate style yields more expressive and cinematic results. For the best balance of quality and detail, use medium or long prompts, with either WD14 tags or ToriiGate prompts. You can also mix them, beginning with WD14 tags to define key traits, then append a short natural-language phrase to describe pose, mood, or lighting.

Example Prompts

(to be filled in)