An anime model based on NoobAI XL.
v4.5 Rectified Flow
Many thanks to Bluvoll and CabalResearch.
Without their pioneering work on converting Noob vpred to RF, this model would not exist.
Important
This is a Rectified Flow model, so you must use the ModelSamplingSD3 node (or equivalent).
What's new:
Enhanced lighting along with more diversity
Enhanced colour
Enhanced and more consistent details (especially visible after Hi-res fix)
Reduced blue bias
Overall, if you liked the improvements vpred brought to you over eps, RF further improves upon vpred.
Recommended settings:
Shift 3.0
CFG 3.0 is a good starting point, do experiment
Euler (DPM++ 2M SDE is also an option for Hi-res fix)
Simple
30 steps
Older versions
vpred v4
I decided to revisit Tiro because:
I got a couple of feedback that helped me realise v3.1 is flawed
bluvoll's Noob RF exposed how much potential is left untapped in Noob
After many training experiments, I have a model that I think is worth sharing.
Updates in vpred v4:
Reduced blue bias
Reduced exposure issues
Recommended settings:
CFG 3.0 - 5.0
DPM++ 2M SDE / Euler a
SGM Uniform
30 steps
Alternative settings:
Euler / DPM++ 2M SDE
Exponential
30 steps
Notes:
Reducing exposure issues enabled more variety in lighting, however this also means not all gens will have pleasing lighting, so you might have to roll your seed
For some prompts, model might assume low light/night; add tags like day, add tags like dark, dim lighting, night, chiaroscuro in the negative, add light sources like sunlight, lamp, window light, light bulb, light particles to brighten it up
vpred v3.1
Recommended settings:
CFG 3.0 - 5.0
DPM++ 2M SDE / Euler a
SGM Uniform
30 steps
Updates in vpred v3.1:
Finetuned from base Noob vpred
Mitigates the overexposure issues without significantly impacting the dynamic range and saturation
Much better at flat and sketchy styles
Includes a finetuned VAE with subtle improvements
Important:
If the default style is not to your taste, use artist tags
The aim of this finetune was to specifically try mitigate the issues present in base Noob vpred; hence the model's knowledge remains similar
very awa is a strong tag, use it prudently
Note: If you want apply the changes to your own Noob vpred based model, grab the LoRA extraction from my HuggingFace.
vpred v3.0
Recommended settings:
CFG 3.0 - 5.0
DPM++ 2M SDE
SGM Uniform
30 steps
Updates in vpred v3.0:
Finetuned from base Noob vpred, with no other models merged into the UNet
Much better at flat and sketchy styles
Less overexposure issues
Better at dark compositions - use tags like dark, night
Alternate aesthetic tag: very aesthetic
Important:
If the default style is not to your taste, use artist tags
The aim of this finetune was to specifically try mitigate the issues present in base Noob vpred; hence the model's knowledge remains similar
If you want more compatibility / less bias, grab the base variant from my HuggingFace
vpred v2.0
Recommended settings:
CFG 3.0
DPM++ 2M SDE
SGM Uniform
30 steps
Updates in vpred v2.0:
Better white balance
Less oversaturation issues
Reduced style bias
Note: If you want even less style bias, grab the base variant from my HuggingFace
vpred v1.0
Important: vpred version requires much lower CFG than eps models.
CFG 1.5 is a good starting point for vpred version. CFG 4.0 is a good starting point for eps versions.
Reduce CFG if you find your gens oversaturated.
Your choice of sampler is important for the vpred version. Use ddpm for balanced colours, use stronger samplers like the SDE ones for more saturation.
Follow the same prompting and configuration guidelines as NoobAI XL, especially for vpred version.
Use artist tags if the default style of the model is not to your taste.
I would love to see your gens and creativity!
Inherits the same license as NoobAI XL.
Description
Further trained with larger dataset and improved recipe
FAQ
Comments (6)
With the same prompt, it does produce darker images more easily than other models if you don't explicitly specify that it should be bright or daytime
This is a feature or a flaw depending on your perspective IMO. This is also a bit of a compromise, as fixing the exposure issues properly will require far more compute than I have + the original Noob training dataset.
@latent_space_dreams I think this is a flaw, after adding a prompt about brightness, the whole style of painting changes. For example, when not added, the model's drawing style is more realistic/3D and has more details, and after it is added, it becomes anime-style.
Can you post a few examples so I can understand this better?
@latent_space_dreams https://civitai.com/posts/24117689
Thanks for the example. That is unfotunately an unintended bias. Thanks for bringing this to light; I'll keep this in mind for any future versions of Tiro vpred.
Also, I figured out some mitigations:
1. Add day to positive & add dark, dim lighting to negative
2. Add dark, dim lighting, night, chiaroscuro in negative for more brightness.
3. Add 3d to negative, but this also steers the model towards anime especially at high CFG.
4. Increasing CFG helps with brightness.