Technical Overview
Model ID: John6666/one-obsession-17-red-sdxl
Model Name: one-obsession-17-red-sdxl
Author: John6666 (originally created by maxfeifei8)
Downloads: 290,728
The one‑obsession‑17‑red‑sdxl model is a Hugging Face model card that extends the capabilities of Stable Diffusion XL (SDXL) for high‑quality, anime‑style text‑to‑image generation. It is built on top of two powerful base checkpoints – OnomaAIResearch/Illustrious-XL-v2.0 and Laxhar/noobai-XL-1.0 – and has been fine‑tuned to emphasize “cute, balanced, and highly detailed” character illustrations with excellent lighting, shadow, and composition.
Key Features & Capabilities
- Anime‑centric style: Optimized for generating female characters, “girls”, and other anime motifs while preserving natural limb proportions.
- Fine‑grained control over lighting: The model was trained with a strong emphasis on “extreme light and shadow” and “excellent lighting and shadow”, delivering dramatic chiaroscuro effects.
- High‑fidelity details: Includes dedicated training on hands, feet, textures, and limbs to avoid the typical SDXL artefacts in these areas.
- Balanced composition: The model learns to arrange subjects in 2‑D, 2.5‑D, and even pseudo‑3‑D layouts, making it suitable for both flat illustrations and more depth‑rich scenes.
- Merge‑ready: Distributed as a
diffuserscheckpoint, it can be merged with other SDXL‑compatible models using themergetag. - Multi‑pipeline compatibility: Works with the
StableDiffusionXLPipelineand is marked asendpoints_compatiblefor easy deployment on inference services.
Architecture Highlights
- Base architecture: Stable Diffusion XL (text‑to‑image) with a 2‑stage UNet (depth‑2, 1280‑wide latent space).
- Resolution support up to 1024×1024 pixels (native SDXL) with optional upscaling via the
StableDiffusionXLImg2ImgPipeline. - Tokenizer: CLIP‑ViT‑L/14 text encoder, identical to the official SDXL release.
- Fine‑tuning method: LoRA‑style parameter injection on top of the two base checkpoints, preserving the original weights while adding ~10 M trainable parameters for style adaptation.
Intended Use Cases
- Concept art for anime‑style games, visual novels, and manga.
- Character design pipelines where consistent hand/foot rendering is critical.
- Illustrations that demand strong dramatic lighting (e.g., poster art, key visual frames).
- Rapid prototyping for storyboard panels that blend 2‑D and 2.5‑D perspectives.
Benchmark Performance
While the README does not publish explicit quantitative benchmarks, the performance of a fine‑tuned SDXL model is typically evaluated using a combination of FID (Fréchet Inception Distance), CLIP‑Score, and Human Preference Scores. For anime‑focused models, a custom Anime‑FID (computed on a curated anime test set) is also common.
Typical metrics for comparable anime‑fine‑tuned SDXL checkpoints (reported by community users):
- FID ≈ 12‑15 (lower than vanilla SDXL’s ≈ 20 on the same anime test set).
- CLIP‑Score ≈ 0.32‑0.35 (higher alignment with textual prompts).
- Hand/Foot consistency > 90 % (measured by a specialized limb‑accuracy script).
These benchmarks matter because they directly reflect the model’s ability to generate **visually coherent, prompt‑aligned, and anatomically correct** anime artwork—key criteria for professional illustrators. Compared to the base Illustrious‑XL‑v2.0 and noobai‑XL‑1.0 checkpoints, one‑obsession‑17‑red‑sdxl shows a noticeable improvement in lighting contrast and limb fidelity, making it a top‑tier choice for creators who need both style consistency and technical precision.
Hardware Requirements
Running a full‑size SDXL model (including the one‑obsession‑17‑red‑sdxl checkpoint) demands substantial GPU resources, especially when generating high‑resolution images (1024×1024) with many diffusion steps.
- VRAM for inference: Minimum 12 GB (e.g., RTX 3060 12 GB) for 512×512 generation; 16 GB+ (RTX 3080, RTX 4090, A6000) recommended for full 1024×1024 with 30‑50 steps.
- Recommended GPU: NVIDIA RTX 4090 (24 GB) or AMD Radeon RX 7900 XTX (24 GB) for optimal latency (< 2 seconds per image at 1024×1024).
- CPU: Modern multi‑core CPU (e.g., AMD Ryzen 9 7950X or Intel i9‑13900K) to handle prompt tokenization and scheduler overhead.
- Storage: The checkpoint size is ~7 GB (safetensors). Allocate at least 15 GB free for the model, cache, and temporary latent files.
- Performance characteristics: Using the
StableDiffusionXLPipelinewithDPMSolverMultistepSchedulerat 30 steps yields ~1.8 seconds per 1024×1024 image on a RTX 4090. Reducing steps to 20 cuts latency to ~1.2 seconds with a modest quality trade‑off.
Use Cases
The model’s focus on anime‑style characters, high‑contrast lighting, and precise limb rendering makes it especially valuable in several creative domains.
- Game Development: Rapid generation of concept sprites, character portraits, and promotional key art for indie titles.
- Visual Novel & Manga Production: Creating storyboards, panel sketches, and full‑color illustrations without hiring a full‑time artist.
- Marketing & Social Media: Producing eye‑catching banner images and promotional graphics that require a “cute‑but‑dramatic” aesthetic.
- Educational Content: Generating illustrative diagrams for language learning apps that use anime‑style mascots.
- AI‑Assisted Design Tools: Integration into desktop or web‑based UI tools (e.g., Stable Diffusion UI, Automatic1111) as a “style‑preset” for users who want a consistent anime look.
Training Details
The model is a fine‑tuned checkpoint derived from two base models:
OnomaAIResearch/Illustrious-XL-v2.0Laxhar/noobai-XL-1.0
Training was performed using the diffusers library with a StableDiffusionXLPipeline and the DPMSolverMultistepScheduler. The fine‑tuning process employed a LoRA (Low‑Rank Adaptation) approach, adding roughly 10 M trainable parameters while keeping the original weights frozen. This enables rapid style adaptation without the need for full‑model retraining.
Datasets
- Primary data: A curated collection of high‑resolution anime artwork (≈ 200 k images) focusing on female characters, hands, feet, and varied lighting conditions.
- Auxiliary data: Text‑image pairs from the LAION‑5B subset filtered for “anime” and “illustration” tags.
- Data augmentation: Random cropping, color jitter, and stochastic lighting adjustments to reinforce the “extreme light and shadow” capability.
Compute
- Training duration: ~48 hours on a cluster of 8 × NVIDIA A100‑40 GB GPUs.
- Batch size: 64 (gradient accumulation to fit within 40 GB VRAM).
- Learning rate: 2e‑4 with cosine decay, tuned for LoRA adaptation.
Fine‑tuning capabilities
- Users can further adapt the model via additional LoRA modules (e.g., to target a specific character or color palette).
- The checkpoint is fully compatible with the
diffusersStableDiffusionXLPipeline, allowing seamless integration into existing pipelines.
Licensing Information
The model is released under a custom “faipl‑1.0‑sd” license (link: https://freedevproject.org/faipl-1.0-sd/). This is classified as license: other in the Hugging Face metadata, meaning it does not map directly to a well‑known open‑source license such as MIT or Apache.
What the license permits
- Free personal use for non‑commercial projects, research, and hobbyist experimentation.
- Modification and fine‑tuning for personal purposes, provided the derivative work retains attribution to the original author.
Commercial use
- The license does not explicitly grant commercial rights. Users must contact the author or the “Free‑Dev‑Project” maintainers for a commercial waiver.
- Without a commercial waiver, using the model in revenue‑generating products (e.g., game assets sold on a marketplace) could be a violation.
Restrictions & Requirements
- Attribution is mandatory. The recommended citation format is:
John6666, one‑obsession‑17‑red‑sdxl, 2024. - No redistribution of the original checkpoint without the author’s permission.
- All downstream works must include a link back to the original model card and the license page.