Hikari_noob_v-pred_0.6 Open-Source Anime Image Generation Model - Based on Fine-tuning Fusion for Stable Quality Improvement

Hikari Noob V Pred 0.6

Developed by RedRayz

An anime-style text-to-image model fine-tuned from NoobAI-XL (ν-prediction) and integrated with SPO LoRA, improving stability and image quality

Image Generation EnglishOpen Source License:Other #Anime Image Optimization #High Quality with Low Steps #ν-prediction Sampling

Downloads 37

Release Time : 11/14/2024

Model Overview

A Stable Diffusion model optimized for anime-style image generation, supporting multiple samplers with good results achievable in just a few steps

Model Features

Efficient Generation

Good results achievable in just 10 steps, recommended 12+ steps

Multi-Sampler Support

Supports various samplers beyond Euler

Detail Optimization

Enhanced detail rendering and refined contours, reducing overexposed skin tones

Style Flexibility

Weak base style, can reshape style through prompts or LoRA

Model Capabilities

Anime-style image generation

Landscape image generation

Character image generation

Use Cases

Anime Creation

Anime Character Design

Generate anime character images in various styles

High-quality character rendering with rich details

Anime Scene Creation

Generate anime-style backgrounds and scenes

Optimized landscape quality with richer details

🚀 Hikari Noob v-pred 0.6

This is a fine - tuned model based on NoobAI - XL(ν - prediction) and merged with SPO LoRA, aiming to generate high - quality text - to - image results.

🚀 Quick Start

Requirements

AUTOMATIC1111 WebUI on dev branch
Latest version of ComfyUI
Latest version of Forge or reForge

Instruction for AUTOMATIC1111

Switch branch to dev (Run this command in the root directory of the webui: git checkout -b dev origin/dev or use Github Desktop).
Use the model as usual!

✨ Features

Improved stability and quality.
Works with samplers other than Euler.
Good results with only 10 steps (12 steps or more recommended).
Fixed a problem in which the quality of output was significantly degraded when the number of tokens exceeded 76.
The base style is not strong and can be restyled by prompts or LoRAs.
This model does not include any base model other than NoobAI (ν - prediction 0.5), so it has the equivalent knowledge. You can generate characters that have appeared by August 2024.

📚 Documentation

About v0.6

v0.5 was too flat, so we improved the quality of detail rendering and made outline thinner.
Improved the quality of the landscapes just a little bit.
Reduced skin color overshoot.

Prompt Guidelines

Almost same as the base model.

To improve the quality of background, add simple background, transparent background to Negative Prompt.

Recommended Prompt

Positive: None (Works good without masterpiece, best quality). Negative: worst quality, low quality, bad quality, lowres, jpeg artifacts, unfinished, photoshop (medium), abstract or empty.

Recommended Settings

Steps: 8 - 24
Sampler: DPM++ 2M(dpmpp_2m)
Scheduler: Simple
Guidance Scale: 2 - 7

Hires.fix

Hires upscaler: 4x - UltraSharp or Latent(nearest - exact)
Denoising strength: 0.4 - 0.5(0.65 - 0.7 for latent)

Merge recipe(Weighted sum)

Stage 1: Finetune Hikari Noob v - pred 0.5 and merge(see below) *A,B: Hikari Noob v - pred 0.5 based custom checkpoint
v0.5(NoSPO) * 0.75 + A * 0.25 = tmp1
tmp0 * 0.75 + B * 0.25 = tmp2
tmp2 + SPO LoRA * 1 + sdxl - flat * - 0.25 + sdxl - boldline * - 1 = tmp3
Adjust tmp3(0.2,0.2,0.2,0.1,0,0,0,0) = Result

Training scripts

sd - scripts

📄 License

This model is licensed under Fair AI Public License 1.0 - SD.

If you make modify this model, you must share both your changes and the original license.

You are prohibited from monetizing any close - sourced fine - tuned / merged model, which disallows the public from accessing the model's source code / weights and its usages.

Do not reupload this model.

image/jpeg