Type | |
Stats | 253 57 |
Reviews | (36) |
Published | Mar 16, 2025 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 78D901D781 |
Overview
This model is built based on the architecture of the NOOBAI XL-VPred 1.0, with some structural modifications. It is trained on the Danbooru2024 dataset along with Yande Full and e621, and using NOOBAI XL-VPred 1.0 and Illustrious XL 1.0 as teacher models during training.
In version 2.0: I have used the old data along with additional real-life character data more than 50k images from various source on the internet.
In version 3.0, I refactored the dataset and added more dataset labels using ChatGPT o3-mini, followed by a manual recheck.
For version 1.0: This model focuses on balancing multiple art styles (through the use of trigger prompts) and good anatomy when generating images.
For version 2.0: This version focuses heavily on improving anatomy and enables the creation of more realistic characters (through the use of trigger prompts). Note that this version may reduce the quality of image generation across multiple art styles.
For version 3.0: This version can generate images in multiple styles (similar to version 1.0) while also creating more realistic characters (more lifelike compared to version 2.0) with improved anatomy. However, to achieve the desired image, you need to input a precise descriptive prompt, as it significantly impacts the output.
Important Note
I personally reconstructed this model, so I’d greatly appreciate any feedback. Your insights won’t just motivate me but will also help me better understand its strengths and weaknesses, allowing me to refine it in the future.
This is a V-prediction model (unlike epsilon-prediction), which requires specific parameter configurations. Please refer to the user guide here.
Currently, the model is not available for use via Civitai Generation. You can visit the following website to use it:
Settings for Generating Realistic Characters
For version 2.0 and up: Add these prompt to generate
Positive prompt: realistic, cosplay, real life, photorealistic
Negative prompt: illustration, blur, film grain, noise, sketch, comic, cartoon, toon, oil painting (medium), flat color, outline, 3D, 2.5D, 2D, unrealistic, game engine style, anime coloring, smooth skin
Recommended Settings
Positive prompt: masterpiece,best quality,amazing quality
Negative prompt: bad quality,worst quality,worst detail,sketch,censor, simple background,transparent background
CFG: 4-6
Clip skip: 2
Step: 20-30
Sampler: Euler a
Contributed by @Ligmanese
Sampler: Euler Ancestral CFG++
Schedule Type: Simple
Sampling Steps: 25-30
CFG Scale: 1.2-1.5
Note:
I don’t use any post-processing or LoRA to enhance the example images. They are generated solely using these settings and prompts with my base model.
For comparison and independent evaluation, I used prompts from various sources and authors to generate these example images.
Acknowledgments
Thanks to narugo1992 and Nyanko for sharing such valuable data and Laxhar Lab for providing an amazing model!
Thanks @Sennke for creating the noobReal model! This model has given me more ideas for improving the ChromaYume version 2.0.
If you'd like to support my work, you can do so through Ko-fi!
Suggested Resources
Discussion
Hello, as some people have pointed out, version 2.0 seems to have significantly reduced diversity in generating images with various artist styles. I was already aware of this issue beforehand. The current version 2.0 is a demo model focused on generating realistic images while also incorporating anime with multiple artist styles.
The next version is still in the training and testing phase. As for this version (which I’ve been training for many days now), based on my tests, it has retained the ability to generate artist styles similar to version 1.0 (in my tests, it actually produces better images). Not only that, but it can also generate images of real-life characters instead of the semi-realistic images seen in version 2.0.
I have optimized this version and applied many techniques to create a model capable of generating a wide range of content. I’m still unsure whether to release it as early access or publish it directly (the training cost for this version is significantly higher than the previous one, and I’ve had to stay up many nights trying to improve it :v).
Do you have any plans for continuing training the chroma without any real-life data? Because it looks like realistic one should be a separate model tbh.
My friend, we just want an even better NoobAI, there are still a lot of weak artist tags, and some anatomy problems, we don't need a generic “realistic” style, don't put garbage in the training. Just improve what's already good and keep the NoobAI prompting.
Thanks.
Note: Illustrious has more recent versions.
One big problem with v3 is high contrast / burned colors. Almost all sample images display this issue. Even reliable euler_a which usually produces clean and smooth colors suffers from this issue to the point that sometimes it starts to generate black and white lineart at CFG >= 5. I never had such problems with v1. And DPM++ 2M SDE sampler is completely unusable.
I managed to get decent colors with plain euler and beta scheduler up to CFG 6 but it often generates strange hands. The most consistent results are with euler + sgm_uniform up to CFG 4 but colors still come a bit burned for my taste.
I hope this issue can be resolved with future version. Because otherwise v3 is great and very promising model.
From my personal perspective, I find that this version 3.0 model has improved quite well compared to the previous two versions. However, the issue here is that this model has increased the difficulty in prompting to achieve the desired image. The idea of mixing 2D, 3D, 2.5D realistic, and artist styles is fantastic. I also understand that training a model with multiple styles is extremely challenging and also very costly. I hope that more people will know about this model instead of constantly mixing multiple models together, as the results from these mixes don't change much, mainly just a few minor details. Thank you for creating this model.
v1 is very good since its like base noob vpred 1.0 but more stable and all artist tags look like they should. if you could just do more training on that version so artist tags are even better thatd be incredible