KielForge-fast (SDXL LoRA)

KielForge-fast is a fine-tuned Stable Diffusion XL (SDXL) LoRA designed to generate and modify high-fidelity, intricately detailed futuristic portraits, advanced sci-fi characters, and concept art via both Text-to-Image and Image-to-Image workflows with remarkable realism and sharp textures.

---

🎨 Model Details

• Developer/Creator: KielTech
• Base Architecture: StabilityAI SDXL Base 1.0
• Model Type: Text-to-Image & Image-to-Image / Fine-tuned SDXL LoRA & GGUF variants
• Available Formats: GGUF (kielforge-fast_q4_0.gguf, kielforge-fast_f16.gguf) & 16-bit LoRA weights (pytorch_lora_weights.safetensors)
• Quantization Level: Q4_0 (4-bit round-to-nearest quantization for optimal balance between VRAM footprint and generation quality) alongside full 16-bit precision options
• Language: English

---

⚙️ Processing & Conversion Pipeline

The weights for this model were trained, merged, and processed inside a Kaggle development environment. The custom fine-tuned LoRA was optimized and converted into both 16-bit and GGUF standards to ensure maximum compatibility, flexibility, and efficient execution on consumer hardware running local inference stacks as well as cloud notebooks.

---

🚀 Recommended Usage & Parameters

• Resolution: 1024 × 1024 pixels (Native SDXL resolution)
• Sampling Steps: 25 – 35 steps
• Sampler: DPM++ 2M Karras, DPM++ SDE Karras, or Euler a
• CFG Scale (Guidance): 5.0 – 8.0
• Img2Img Strength: 0.45 – 0.6 (Sweet spot for modifying details while preserving composition)

---

💻 How to Use

1. Text-to-Image (StableDiffusionXLPipeline)

import torch
from diffusers import StableDiffusionXLPipeline

base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
pipe = StableDiffusionXLPipeline.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True
).to("cuda")

pipe.load_lora_weights(
    "kiel2/KielForge-fast", 
    weight_name="pytorch_lora_weights.safetensors"
)
pipe.enable_attention_slicing()

prompt = "A stunning portrait of a futuristic warrior, highly detailed armor, masterwork"
negative_prompt = "blurry, distorted, low quality, low resolution"

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    guidance_scale=7.5
).images[0]

image.save("generated_image.png")

2. Image-to-Image (StableDiffusionXLImg2ImgPipeline)

import torch
import gc
from diffusers import StableDiffusionXLImg2ImgPipeline
from PIL import Image

torch.cuda.empty_cache()
gc.collect()

init_image = Image.open("test_output.png").convert("RGB")
init_image = init_image.resize((1024, 1024))

base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True
)
pipe.enable_model_cpu_offload()

pipe.load_lora_weights(
    "kiel2/KielForge-fast", 
    weight_name="pytorch_lora_weights.safetensors"
)

prompt = "A stunning portrait of a futuristic warrior with glowing neon cybernetic implants on her face, highly detailed armor, masterwork"
negative_prompt = "blurry, distorted, low quality, low resolution"

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    image=init_image,
    strength=0.5, 
    num_inference_steps=30,
    guidance_scale=7.5
).images[0]

image.save("generated_image_amended.png")

3. Local Inference UIs (ComfyUI / WebUI Forge) Download your preferred weight variant (kielforge-fast_q4_0.gguf, kielforge-fast_f16.gguf, or the LoRA files) directly from this repository.

Place the file into your local inference UI's appropriate directory (e.g., ComfyUI/models/unet/ for GGUF files or ComfyUI/models/loras/ for the LoRA adapter weights).

Load the model through your text-to-image or image-to-image workflow to generate or transform your images!

⚠️ Limitations & Bias When using the 4-bit quantized Q4_0 version, users may occasionally notice minor quantization artifacts compared to the full 16-bit float variants, traded off for significantly faster generation speeds and a dramatically lower memory footprint.

This model inherits the general capabilities, constraints, and safety profiles of the underlying SDXL base architecture.