Sign In

noobReal v3.0

864
11.4k
275
260
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
650
0
Reviews
Published
Nov 1, 2024
Base Model
Illustrious
Training
Steps: 37,999
Epochs: 79
Usage Tips
Clip Skip: 2
Trigger Words
realistic:1.2
cosplay
Hash
AutoV2
E42C8A2BA0
default creator card background decoration
Dark Comic Car
Sennke's Avatar
Sennke

NoobReal is a multi-style model, not limited to realism alone.

  • postive prompts for realistic

    realistic,cosplay,

  • negative prompts for realistic

    illustration,blur,film grain,noise,sketch,comic,cartoon,toon,oil painting \(medium\),flat color,outline,3D,2.5D,2D,unreal,unity,blender,anime,smooth skin,

  • Suggest parameters

    DPM++ 2M

    cfg : 4.5

The model is in anime style by default, because I don't want to destroy the anime concept,just like Lora or Merge model. So when you want to generate a realistic style, please add these negative prompt words. You can also mix these styles.
I'm aimming to use high-quality real photo training data to achieve realistic style, same time to improve anime-style like anatomical accuracy, scene composition, through its realistic concept.

If you like this model, please give me a like and I will continue to upgrade this model❤️


Version 3.0 2025-03-07

Feature

  1. Full finetuning from Epsilon-pred 1.1-Version

  2. Improved Dataset Tagging Structure
    tip: You can still use noobai prompt habits, various style words, artist words, and even mix with realistic styles

  3. Expanded Training Data (now 4,500+ images)

  4. Greater Diversity in compositions, art styles, poses, and character ages

    • V3.0: Fine-tuned the Cyberfix restoration model to minimize its impact on art style and composition. (During long-term testing, I noticed Cyberfix over-suppressed rare concepts, making it hard to generate dynamic poses or bold compositions. This update aims to fix that)

    • if you prefer V2.1's style and stable, there is no need to upgrade. The stability of v3.0 is slightly reduced.

  • 基于Epsilon-pred 1.1-Version进行训练

  • 重新调整了训练集标签结构

  • 训练集张数4500张

  • 更多样的构图、画风、姿势、年龄

    • V3.0重新微调了Cyberfix修复模型,极大减少Cyberfix对模型画风、构图的影响。你如果喜欢V2.1的稳定性,依然可以使用V2.1来出图,可以算作2个版本。(我在长期使用后发现Cyberfix会对个别稀有概念降权严重,导致难以生成夸张的身体姿势,整体构图都趋近保守,所以我进行了微调来缓解该问题)


v2.1

Feature

  • Mainly trained prompt words(Triggers)

    • realistic:daily life photos

    • reversal film:film style

    • cosplay:more fancy outfits

    • illustration:My favorite anime mixed styles ( Not necessary for anime style.Mainly enhanced the diversity of character movements. )

    • scenery: Enhance scene space logic (Not necessary, I use this word to prevent contamination of anime styles)

    • depth_of_field: can add to negative words increase background clarity

    • blurry background: can add to negative words increase background clarity

    • simple background: can add to negative words increase background clarity

  • Recommended negative words for realistic

    • blur,film grain,noise,sketch,comic,cartoon,toon,oil painting \(medium\),flat color,outline,3D,2.5D,2D,unreal,unity,blender,anime,smooth skin,

  • Re-adjusted about 3699 images for

    • Realistic 60% (95% real photos, 5% AI photos。I don’t like the exaggerated body shape of other nsfw Ai realistic models. so I mostly use real-life photos for training.)

    • illustration 35%

    • 3DCG 5%

  • Tag+NLP mixed captions

  • better generalizationtrained

  • better age range

  • better eye quality

  • more better realistic background

  • Hires.fix still recommended. ADetailer are not necessary in v2.1.

  • Used a self-made version of Cyberfix model to repair the limbs (the original Cyberfix will cause changes in painting style and loss of generalization)

    • The limbs are not significantly improved compared to 1.0, but 2.1 can generate a wider variety of body movements.

  • full finetuning from NoobAI-XL v1.0


v1.0

Feature

  • trained on 3000 images and Tag+NLP mixed caption

  • better realistic background,The composition will be some different from noobai1.0

  • Used CyberFix to fix NoobAI-v1.0's anatomy

  • New training method

    • Fixed the noise problem in v0.3

    • Hires.fix and ADetailer are no longer required, although still recommended

    • Supports higher resolutions with less crashing, such as 1024x1536

  • Trigger prompt

    • realistic:daily life photos

    • cosplay:more fancy outfits

  • Negative prompt(if you want to avoid smooth skin or 3D style skin)

    • blur, sketch, comic, cartoon, toon, oil painting \(medium\), flat color, outline, 3D, 2.5D, 2D, unreal, unity, blender, anime ,smooth skin,

  • full finetuning from NoobAI-XL v1.0 (don't use V-Pred version)


v0.3

This model is currently under training.Recommended to enable Hires. fix and ADetailer. If you could use comfyui you will get better results. The case node file is in the gallery post below that you can copy.

With full fine-tuning and tigger words,so it does not destroy the anime styles ability of the NoobAi. Realistic style has been isolated and can also be mixed in other anime styles

V0.3 is a version that I think is at a stage. It is slightly difficult to use but can produce pretty good pictures. So i released this.

trigger prompt: realistic,cosplay,

If you want to mix with the characters that NoobAi originally had, you need to remove the style prompt and increase the trigger weight like: (realistic,cosplay:1.5) recommend weight is 1.0~1.5.

The current V0.3 is trained with 800 images with tag+nlp Mixed caption for test.Background details of the model are better, even without the realistic prompts.

My train target is 2000 images with tag+nlp caption. Just hope NoobAi's anime characters could turn into realistic styles without loss other anime concept.

full finetuning from NoobAI-XL v0.75

License

Same as NoobAI-XL use Fair-AI-Public-License-1.0-SD

https://freedevproject.org/faipl-1.0-sd/