NewBie-image Deployment and Zero-Threshold Usage Tutorial (ENGLISH)
输入“/”快速插入内容
NewBie-image Deployment and Zero-Threshold Usage Tutorial (ENGLISH)
NewBie-image Deployment and Zero-Threshold Usage Tutorial (ENGLISH)
用户5734用户5734
用户7387用户73872025年12月7日修改
❗
Pre-reading reminder: This document is suitable for quickly understanding the deployment and use of NewBie-image; it is suitable for beginners with no prior knowledge to read.
When using AI image generation to publish images on the Internet, you should comply with local laws and regulations.
This article is machine-translated from NewBie-image部署以及使用零门槛教程, so there may be grammatical or phrasing errors. I recommend reading it alongside the original Chinese page for reference, or using an AI assistant to clarify any ambiguities in this translation.
☕
NewBieAI Lab Written by:年糕特工队
😀
Welcome to join the AI drawing Discord community:https://discord.com/invite/bDJjy7rBGm(Discord channel of NewBieAI Lab)
1. Introduction
AI painting has been in the public eye for almost three years now, from the initial leak of NovelAI 1.0 to the alternative for anime PONY V6, as well as the NOOBXL and ILL series of models that symbolize absolute popularity; the open-source community has already demonstrated a fairly mature side. Today, we're introducing the brand-new model NewBie-image developed by NewBieAI-Lab; it is trained on an anime dataset, labeled using natural language + TAG; based on the Next-DiT neural network, it incorporates Jina and Gemma3 as text encoders, which provide extremely strong natural language understanding and tag recognition capabilities, as well as image performance completely superior to SDXL. Although the model has only undergone less than 20% of the training expected in its design, it still demonstrates sufficient strength; this represents another significant step forward for the anime-style text-to-image community.
If you are willing to support NewBieAI Lab in achieving its subsequent training goals, you can contact our relevant members.
2. Introduction
NewBie-image is trained on a high-quality anime dataset, using the annotation method of natural language + TAG. By the time of its open source, it had completed 10 full rounds of training without any distillation or post-training aesthetic fine-tuning. The main use case of NewBie-image is anime and second-generation image generation. Even without using LoRA, it can excellently complete most image generation tasks.
Thanks to the use of natural language annotation, NewBie-image can understand fairly complex prompts and maintain a complete composition when generating images, while the TAG attributes are also preserved. Even when only danbooru's TAG labels are used, the model can still produce correct results.
It is worth mentioning that the model is in a very early version. Due to insufficient training and funding, there are still a large number of concepts in the model that have not been fitted, waiting for subsequent plans to address.
3. Deployment
For beginners, deploying NewBie-image may not be as simple as imagined. It is not like the previous "download-import-use" method of SDXL. However, it doesn't matter. The author will provide the most detailed teaching.
Before starting the deployment, let's still go through the basic configuration requirements as usual: The requirements of NewBie-image are basically similar to those of the SDXL series of models, but the generation speed is a bit slower. However, the author cannot provide a real "bare minimum" configuration, although the model should be able to run under fairly extreme configurations. But it is still recommended that you should have at least 8GB of video memory and a core stronger than RTX4060, as well as at least 16GB of Random Access Memory. The author used a RTX4070TIS 16GB graphics card and 32GB of Random Access Memory during testing. For reference, when generating a 1024*1024 image, the time taken for 28 steps was 40 seconds.
Meanwhile, since the installation environment is extremely cumbersome and prone to problems, this guide only covers setup for NVIDIA graphics cards.
🌐
During deployment, you need to access sites such as Python, GitHub, and Hugging Face extensively. The author recommends that you use a proxy or any environment that can connect to the above sites for deployment.
And the author highly recommends that you use AI to assist you in installing the relevant content. This document has public download access; you only need to download it as a PDF file, then hand it over to AI, and you can ask it about the specific installation details.
❗
Due to the rushed launch and the experimental nature of the model, NewBieAI Lab is working hard to simplify the installation process and steps, as well as ensure compatibility with existing ComfyUI tools. However, this will take some time (approximately 2–3 days). NewBieAI Lab will promptly announce updates in this tutorial and the communication channel once compatibility with existing tools is achieved, and will also thoroughly revise this tutorial accordingly. For now, you may use the more complex installation method provided below for early access.
Overall, you need to deploy/download the following:Python (the author recommends 3.11.X here)
cuda + PyTorch (the author recommends CUDA 12.8 + PyTorch 2.7)
NewBie-image weight file
gemma3
FLUX.1 VAE
jinaCLIP
triton-windows (if it's Linux, install the corresponding Linux version)
ComfyUI-KJNodes-main
Overall, The NewBie-image loading node depends onflash-attentionandtritoncomponents, and you need to ensure the successful installation of these two nodes.
If you don't understand what the above content is for, or don't know where certain files should be placed, please follow the author's instructions step by step.
Environmental Management
Generally speaking, directly installing a new environment may conflict with the existing environment on the computer. Therefore, we use a conda virtual environment here to solve this problem.
Download conda: You can click here to download a Miniconda. If you have previously installed conda, you can skip this section.
After downloading, let's double-click to run it.