Make sure to leave a bid on this Checkpoint if you want it available on the site: https://civarchive.com/auctions
Every bit of buzz counts!
=============================================
SmoothMix Ultimate!
It combines all other SmoothMix Checkpoints into one!
It can create high quality anime or semi-realistic images, pushing the realism just enough to still be an stylized realism.
Use the triggers prompts "smoothmixanime" and "smoothmixrealism" to achieve the style you want.
You should be able to more easily mirror an anime stylized image to a more realistic version of it by just changing a single prompt! Give it a try! ^^
===============================================
NEWEST VERSION: Anima!
Finally the Anima version is done! ^^
Just one of the prompts "smoothmixanime", "anime style", "smoothmixrealism", "realistic style" should be enough!
Not using these prompts will give you a similiar result to Illustrious v5.
Make sure to download the necessary files - VAE ad Text Encoder - to use it.
All images were made on Stable Diffusion Forge - Neo.
Anima recommended settings:
Step 20~25 / CFG: 2~3
Smapling (ER SDE/Euler a/DPM++ 2M)
Scheduler: Karras*/Simple/Beta (Schedules makes a big difference on the results)
Scheduler Karras is not recommended for Realistic images/style and can result in weird images!
Resolution: 832x1216 or above.
Adetailer
Hires. fix: Remacri
All images used anima_baseV10_txt.safetensors and qwen_image_vae.safetensors!
===================================================
Illustrious V.2!
Adjusted the base SDXL VAE with this useful tool VAE-BlessUp and adjustments with Supermerger! No need for VAE anymore! The colors should feel a little better both on realism and anime styles! ^^
Illustrious v.2 recommended settings:
Steps: 20 ~ 30 - (Euler a/DPM++ 2M) - CFG: 3 ~ 5
Resolution: 832x1216 or above.
VAE: No need anymore! ^^
Adetailer
Hires. fix: Remacri
Use the set of Smooth Embeddings! The Positive and Negative Embedding will help creating a high quality image!
Use Smooth Detailer Booster for better effects and details.
Use Smooth Lighting Enhancer for better lighting effects!
ZImage Turbo V.1!
Base model is ZImage Turbo.
ZImage Turbo is a heavily censored base model, so its native NSFW support is limited. While SmoothMix ZImage Turbo can produce some nice images with nudity, the coverage and quality remains modest. For higher‑quality or more complex NSFW results, you’ll need to use LoRAs specifically trained for that type of content.
All images on the Showcase used the new Smooth Workflow v.4!
ZImage Turbo v.1 recommended settings will vary depending on the combo of Sampler + Scheduler you are using. Make sure to check out all images I posted for several examples - here are some recomendations:
Sampler (Euler A) + Scheduler (Simple/SGM Uniform/Beta)
Steps: 8 ~ 10/12 (Go over 10 only for very complex images - AKA very detailed/long prompts)
CFG: 1.1 ~ 1.6 (More CFG adds more contrast - Great for night/sunsets or dark indoors images)
Sampler (DPM++ SDE) + Scheduler (Simple/SGM Uniform/Beta)
Steps: 8 ~ 10/15 (Go over 10 only for very complex images - AKA very long prompts)
CFG: 1.1 ~ 1.5
Sampler (DPM++ SDE/Euler A) + Scheduler (DDIM Uniform)
Steps: 8 ~ 10 (Keep in this range or the Scheduler will have no effect or completely mess up you image)
CFG: 1.1 ~ 1.5
ModelSamplingAuraFlow: 0.7
CLIP/TextEncoder: ZImageTurbo_textEncoder
VAE: ZimageTurbo_vae
ILLUSTRIOUS V.1!
Base model is IllustriousXL v2.0.
VAE is NOT embedded on this model - make sure to select your favorite!
Illustrious v.1 recommended settings:
Steps: 30 ~ 40 - (Euler a/DPM++ 2M) - CFG: 3 ~4
Resolution: 832x1216 or above. (Showcase image are 1024x1496 +hires)
VAE: sdxlVAE, Luna, Neptunia, Better Colors
Adetailer
Hires. fix: Remacri
Use the set of Smooth Embeddings! The Positive and Negative Embedding will help creating a high quality image!
Use Smooth Detailer Booster for better effects and details.
Use Smooth Lighting Enhancer for better lighting effects!
Now go and have fun!
Description
FAQ
Comments (5)
Just a bit information; "ZImage Turbo is a heavily censored base model" is false. It's just not heavily trained in nsfw concepts, but it's not "censored".
Kinda like how SDXL was back when it was first released; not censored, just lacking data.
Anyone who considers ZiT to be "heavily censored" is new and doesn't know what they are talking about.
yeah, censored at the training stage when the material for training was selected... one of the main ways models are censored.
So how much illustrious/noob tagged knowledge should this have? if i wanted to prompt heavily pre-trained illustrious and noob characters will it work just fine? Is it flexible between natural language and tagged prompts?
The Z-Image Turbo is a distilled model. She understands concrete nouns better and abstract emotional descriptions worse.
This is good, it allows you to do complex things, but it's also bad, it requires more detailed instructions.
The peculiarity is that any text neural network can easily convert text into a prompt, Just give the command "rewrite the text without using abstractions and metaphors" (or "write the text using keywords without using abstractions and metaphors") and you can use an almost ready-made prompt. All that remains is to add quality tags, but it will work well without them.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.