YunusTAS13/InkNoir-Core

InkNoir Core

Dark anime illustration model · SD 1.5 · Diffusers · Safetensors

InkNoir Core, YunusTAS13 tarafından geliştirilen karanlık anime ve illüstrasyon odaklı bir Stable Diffusion 1.5 modelidir.

Modeli indir · Örnekler · Bağımsız testler

---

English

What is InkNoir Core?

InkNoir Core is a dark-anime illustration model based on DreamShaper 8. The InkNoir style adapter was fused into a complete standalone pipeline at LoRA scale 0.4.

This repository includes both a full Diffusers pipeline and a single-file checkpoint. It is not an adapter-only release.

What can it do?

Capability	Status
Text-to-image	Supported
Dark anime illustration	Main focus
Character portraits and full-body art	Supported, hands may fail
Image-to-image	Supported through Diffusers/local UIs
Inpainting	Use an SD 1.5 inpainting pipeline with this style
GIF/video frames	Possible as a separate frame workflow; no native temporal consistency
Pixel-art and sprite sheets	Experimental; exact cell layouts are not guaranteed
Turkish prompts	Basic external translation is recommended

Gallery — independent, unmasked subjects

All gallery prompts, seeds and settings are available in examples/prompts.md. The unrelated-subject benchmark is available in examples/independent/benchmark_report.md.

Files in this repository

File or folder	Description
InkNoir_Core_Full.safetensors	Standalone fused SD 1.5 checkpoint
model_index.json, unet/, text_encoder/, vae/, tokenizer/, scheduler/	Complete Diffusers pipeline
adapters/InkNoir_Core.safetensors	Optional unfused LoRA adapter
run_inknoir.py	Beginner-friendly local launcher
requirements.txt	Python dependencies
examples/	Gallery, prompts, independent tests and comparisons
MODEL_LICENSE_NOTICE.md	Base-model license and attribution notes
PUBLISH_CHECKLIST.md	Release verification checklist

Quick start

Easiest local installation

git lfs install
git clone https://huggingface.co/YunusTAS13/InkNoir-Core
cd InkNoir-Core

python -m venv .venv
source .venv/bin/activate              # Windows: .venv\\Scripts\\activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

python run_inknoir.py \\
  --prompt "dark anime illustration, a lone warrior in a rainy neon city" \\
  --output inknoir_output.png

The launcher automatically selects CUDA, Apple MPS or CPU. The first run downloads or loads the model and may take a few minutes.

Linux with fish

source .venv/bin/activate.fish
python -m pip install -r requirements.txt
python run_inknoir.py --prompt "dark anime illustration, yağmurlu neon şehir"

System requirements

Component	Minimum practical	Recommended
GPU	4–6 GB VRAM at 512x512	NVIDIA CUDA GPU with 8 GB+ VRAM
RAM	8 GB	16 GB
Storage	5 GB free	8–10 GB free on SSD
CPU	Modern 4-core CPU	Modern 6–8-core CPU
OS	Linux, Windows, macOS may work	Linux or Windows
Python	3.10–3.12	3.11 or 3.12

CPU generation is supported but significantly slower. Lower the resolution and steps if memory is limited.

Diffusers usage

import torch
from diffusers import StableDiffusionPipeline

model_path = "YunusTAS13/InkNoir-Core"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

pipe = StableDiffusionPipeline.from_pretrained(
    model_path,
    torch_dtype=dtype,
    safety_checker=None,
    requires_safety_checker=False,
).to(device)
pipe.enable_attention_slicing()

image = pipe(
    "dark anime illustration, a quiet bookstore at night, cinematic lighting",
    negative_prompt="low quality, blurry, deformed, bad anatomy, text, watermark",
    width=512,
    height=512,
    num_inference_steps=28,
    guidance_scale=7.0,
).images[0]
image.save("result.png")

Automatic1111, Forge and ComfyUI

Use InkNoir_Core_Full.safetensors as a normal SD 1.5 checkpoint:

1. Download the file from this repository.
2. Put it in the checkpoint/model folder.
3. Select it as the active checkpoint.
4. Generate with a prompt.

Do not put the full checkpoint in the LoRA folder. The style is already fused into the full checkpoint. If you use the optional adapter instead, load it on DreamShaper 8 with strength 0.3–0.5.

Recommended settings

Setting	Starting value	Note
Resolution	512x512	Safe SD 1.5 starting point
Steps	24–32	More steps are not always better
CFG	6.5–7.5	Very high CFG may add artifacts
Full checkpoint	No extra LoRA	Style is already fused
Separate LoRA	0.3–0.4	Start low and increase carefully
Sampler	DPM++ 2M Karras or Euler a	Good starting choices

Start with a neutral negative prompt:

low quality, blurry, deformed, bad anatomy, text, watermark

Image-to-image and inpainting

For image-to-image, use strength=0.25–0.40 when preserving the reference is important, and 0.45–0.65 for stronger changes. For precise object replacement, use a dedicated SD 1.5 inpainting pipeline.

from PIL import Image
from diffusers import StableDiffusionImg2ImgPipeline

img2img = StableDiffusionImg2ImgPipeline.from_pipe(pipe)
source = Image.open("reference.png").convert("RGB").resize((512, 512))

result = img2img(
    prompt="dark anime style, preserve the character silhouette, colder blue lighting",
    image=source,
    strength=0.35,
    guidance_scale=7.0,
    num_inference_steps=28,
).images[0]
result.save("img2img_result.png")

Turkish prompts

The SD 1.5 text encoder is more reliable with English. Turkish prompts can be translated before generation:

Türkçe: karanlık anime tarzında, yağmurlu neon şehirde yalnız bir savaşçı
İngilizce: dark anime style, a lone warrior in a rainy neon city

The model itself does not contain a general Turkish translation model. See examples/prompts.md for bilingual examples.

Animation, GIF and sprite sheets

InkNoir Core is a still-image model. It can produce individual frames or keyframes, but it does not guarantee temporal consistency by itself. For animation:

• keep a fixed seed and character description;
• generate frames with a separate assembly/stabilization workflow;
• use a dedicated video model when motion consistency is required;
• generate sprite frames separately when exact frame count and cell placement matter.

Independent evaluation

The independent test set intentionally includes unrelated subjects: an unmasked woman, golden retriever, mountain lake, red sports car, bookstore, still life, dragon and robot workshop. It is included to show both strengths and failure cases instead of only the training-like masked character.

• Independent benchmark report
• Prompt and strength control comparison
• All prompts and seeds

Known limitations

• This is an experimental derivative model, not a perfect general-purpose image model.
• Hands, text, exact identity and complex object insertion may fail.
• High adapter strength can produce face artifacts.
• The full checkpoint does not provide native temporal consistency.
• Turkish works best after an external translation step.
• Pixel-art and exact sprite-sheet layouts are experimental.

License and attribution

The base DreamShaper 8 model is listed by its author under creativeml-openrail-m. This fused derivative must retain the base model’s usage restrictions and attribution requirements. Read MODEL_LICENSE_NOTICE.md before redistribution.

The publisher must verify that all curated training images are cleared for redistribution. Users are responsible for their outputs and use of the model.

---

Türkçe

InkNoir Core nedir?

InkNoir Core, DreamShaper 8 tabanlı, karanlık anime ve illüstrasyon üretimine odaklanan Stable Diffusion 1.5 modelidir. InkNoir stil adaptörü 0.4 gücüyle tam modele birleştirilmiştir.

Bu repo yalnızca küçük bir LoRA dosyasından oluşmaz. Yaklaşık 2 GB’lık tam checkpoint, eksiksiz Diffusers pipeline’ı, çalıştırıcıyı, örnekleri ve testleri içerir.

Hızlı kullanım

git clone https://huggingface.co/YunusTAS13/InkNoir-Core
cd InkNoir-Core
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\\Scripts\\activate
python -m pip install -r requirements.txt
python run_inknoir.py --prompt "karanlık anime tarzında yağmurlu neon şehir"

Başlangıç için 512x512, 24–32 adım ve CFG 6.5–7.5 kullanın. Türkçe promptları İngilizceye çevirerek göndermek daha tutarlı sonuç verir.

İçerik

• InkNoir_Core_Full.safetensors: tek dosyalık tam model;
• model_index.json, unet, text_encoder, vae, tokenizer, scheduler: eksiksiz Diffusers pipeline’ı;
• adapters/InkNoir_Core.safetensors: isteğe bağlı ayrı LoRA;
• run_inknoir.py: kolay yerel çalıştırıcı;
• examples/: görseller, promptlar ve bağımsız testler.

Görseli değiştirmeden görme özelliği

InkNoir Core kendi başına görüntü-anlama modeli değildir. Hugging Face’te bir VLM/image-text-to-text modeli eklenerek fotoğrafı değiştirmeden analiz ettirmek mümkündür. Örneğin kullanıcı fotoğraf yükleyip “Bu görselde ne var?” diye sorabilir; model yalnızca metin cevabı üretir. Bu özellik image-to-image işleminden farklıdır.

Sınırlamalar

Model deneysel bir stil modelidir. El, yazı, kesin karakter kimliği, karmaşık nesne ekleme ve animasyon tutarlılığı garanti edilmez. GIF/video için ayrı kare üretim ve birleştirme süreci gerekir.

Lisans

DreamShaper 8’in lisans ve kullanım koşulları geçerlidir. Ayrıntılar için MODEL_LICENSE_NOTICE.md ve DreamShaper 8 model kartına bakın.

---

InkNoir Core · YunusTAS13

Hugging Face · GitHub