Bl4ckSpaces/SpaceDiffusion-XL-V-PRED-2.0

By accessing this model, you agree to use it only for legal, safe, and non-NSFW content creation

🌌 SpaceDiffusion-XL V-PRED 2.0 (v1 Final)

Welcome to the apex of anime and illustration generation. SpaceDiffusion-XL V-PRED 2.0 is a highly advanced, V-Prediction-based SDXL model designed to push the boundaries of anime aesthetics, anatomical precision, and complex prompt adherence.

This model is not a simple mix; it is the culmination of advanced merging architecture, relentless DoRA fine-tuning, and highly curated datasets.

---

🧬 Model Lineage & Architecture

SpaceDiffusion-XL V-PRED 2.0 is born from a rigorous and complex developmental pipeline:

1. The Foundation (AetherV-XL): The core base of this model is AetherV-XL, which was crafted using advanced merging techniques (Singular Value Decomposition / SVD and Semantic Routing) to perfectly balance the structural understanding of NoobAI-XL (V-Pred) and the breathtaking aesthetic versatility of Illustrious-XL (V-Pred).
2. The Forging (DoRA Fine-tuning): AetherV-XL was not left as a mere merge. It underwent intense, multi-day fine-tuning using DoRA (Weight-Decomposed Low-Rank Adaptation) at Rank 64 across dual-GPU clusters to aggressively enhance detail rendering, lighting logic, and character consistency.
3. The Final Fusion: The resulting DoRA weights were permanently fused into the base model at full FP32 precision before being precisely quantized down to FP16 for optimal inference performance without losing a single bit of quality.

📚 Dataset Knowledge Base

• Primary Knowledge: Comprehensive coverage of anime, illustration, and pop culture subjects up to late 2025.
• Modern Injection: Includes a highly concentrated, specialized micro-injection of early 2026 data (via our final DoRA finetuning phase using multi-resolution smart bucketing).

---

⚠️ MANDATORY SETTINGS (READ BEFORE USING!)

Because this model uses the V-Prediction schedule, standard SDXL settings will result in deep-fried or completely broken images. You MUST follow these exact parameters:

• Prediction Type: v_prediction (CRITICAL!)
• CLIP Skip: 2 (CRITICAL for anime styles. Do not use Clip Skip 1).
• Prompting Style: Danbooru Tagging (e.g., 1girl, solo, looking at viewer, blue hair...).
• Recommended Sampler: Euler a, DPM++ 2M Karras, or DPM++ 3M SDE Karras.
• Steps: 25 - 40
• CFG Scale: 5.0 - 7.5
• Resolution: Highly optimized for 1024x1024, 832x1216, 1024x1536, and up to 1.7x mega-resolutions (1280x1280, 1088x1600, 1600x1088).

---

🛠️ How to Use (Tutorial)

1. ComfyUI (Recommended)

To use this model in ComfyUI, you must tell the sampler to use V-Prediction.

1. Load the SpaceDiffusion_XL_V_PRED_2_0_v1_FINAL.safetensors using the standard Load Checkpoint node.
2. Add a ModelSamplingDiscrete node.
3. Connect the model output from the Checkpoint node to the model input of the ModelSamplingDiscrete node.
4. Set the sampling setting on the ModelSamplingDiscrete node to v_prediction.
5. Connect the model output to your KSampler.
6. Ensure your CLIP Text Encoders are set to CLIP Skip -2.

2. A1111 / Forge WebUI

1. Place the model in your models/Stable-diffusion folder.
2. Go to Settings > Stable Diffusion.
3. Look for the setting related to SDXL and V-Prediction. Ensure your WebUI is updated to a version that auto-detects V-Prediction from safetensors metadata.
4. Set CLIP skip to 2 in the UI settings.

3. Diffusers (Python)

import torch
from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler

model_id = "Bl4ckSpaces/SpaceDiffusion-XL-V-PRED-2.0"

# Explicitly define the V-Prediction scheduler
scheduler = EulerDiscreteScheduler.from_pretrained(
    model_id, 
    subfolder="scheduler", 
    prediction_type="v_prediction"
)

pipe = StableDiffusionXLPipeline.from_single_file(
    "SpaceDiffusion_XL_V_PRED_2_0_v1_FINAL.safetensors",
    scheduler=scheduler,
    torch_dtype=torch.float16
).to("cuda")

# Don't forget Clip Skip 2 is natively handled if you use the correct prompt weighting or Compel!
prompt = "masterpiece, best quality, ultra-detailed, 1girl, solo, glowing eyes, cyberpunk city"
image = pipe(prompt, num_inference_steps=30, guidance_scale=6.5).images[0]
image.save("output.png")