Welcome to Taureal Mix!
--
If you love my work you can buy me a coffee : ko-fi.com/tauron you dont have to, but if you do, it helps me stay awake until midnight (which is the only time im able to work on these stuff lol) -- or.. a like or a review wont hurt or takes you more than 1 mins, while it takes me hours, days or even week to make this - it helps the thread alive and afloat.
--
**Update Changelog can be seen at model version details.**
NOTE : Alot of sample images has been removed by CivitAi even though there is no TOS violation.
--
Suitable to those who needs:
All round realistic model with versatile styles.
Low memory cost.
Accurate txt2img results with short/long prompt.
Low model bias = more accurate lycoris/lora representation.
Better integration with my previous / future lora.
Disclaimers:
tested mostly for txt2img, for other functions use at your own discretion.
tested specifically for my previous or future lora. loras from other creators may work or may not work well with my model.
a pruned model is not suitable to be merged. choose the v3.1 or v2.1 instead.
Prerequisite & Resources
These are the resources i use for my standard workflow :
ADetailer
GitHub - Bing-su/adetailer: Auto detecting, masking and inpainting with detection model.
Negative Embeddings (Use 1 only)
Use when needed only, v3.7 pruned without negative ti is observed to produce significantly better result compared to v2.1 / v3.1
Deep Negative V1.x - V1 75T | Stable Diffusion Textual Inversion | Civitai
accurate prompt such as race, skin tone but smoothen the skin texture and has face bias (change result face), also sometimes misinterpret prompt position occassionally.
EasyNegative - EasyNegative | Stable Diffusion Textual Inversion | Civitai
most accurate amongst other negative ti -- prompt such as race, skin tone, face, and texture is very accurate (90%), but sometimes misinterpret prompt position (ex. cum is supposed to be on body, but is on face instead)
most detailed amongst other negative ti -- this embedding (32v or 75v) pushes skin texture and facial detail to the top. however, it changes skin tone, color, and face significantly producing inaccurate result if you are trying to use lyco.
*DO NOT COMBINE NEGATIVE - TI WITH SAME FUNCTIONS IN THE V3.7 MODEL*
Facial Enhancer Embeddings (optional)
yesyeahvh/ulzzang-6500 · Hugging Face
Pure Eros Face - PureErosFace_v1 | Stable Diffusion Textual Inversion | Civitai
Credits to all creators above that helps the model achieve its sample images quality above.
How To prompt this mix
You can follow your own prompting method, this guide only serves as a convention when you are having trouble producing similar results with my lora.
Style —> realistic , realistic details, detailed for realism / anime for anime / masterpiece , best illustration for conceptual art.
Subject —> describe your subject can be anything. 1girl, solo, cute, beautiful, slender, etc. if you use my character lora this should be where you put trigger words of the character
Attire —> describe the outfit of your subject. If you use my attire lora this is where you should put trigger words
Background / Context —> describe the context or the scene
Textual Inversion & LoRa brackets
How to maximize this mix?
Sampler & Steps
SDE Karras Sampler = 16 / 8 Denoise @ 0.32
DDIM Sampler = 16 / 8 Denoise @0.32 (Observed Fast like 2M, but with SDE Quality, DDIM is very good for concise prompt, but is weak at translating complex prompt)
Using negative embeddings is a mustto create sample image quality.Negative Ti is no longer needed to achieve sample image quality
CFG Scale
CFG Scale (creative) 4 ----------- 10 (accurate) -- 12 (Max)
Resolution & Upscaling
Resolution: 512 x 768,
Upscaling = 4x_Ultrasharp lokCX/4x-Ultrasharp · Hugging Face
To get even better results: Use Lightroom to edit your result.
Utilizing ADetailer
Face = Noise 0.3
Hands = Noise 0.25-0.4
Body = Noise 0.3-0.5
The rule of thumb is that there are no constant equation, you need to find one your own that suits what you are aiming, and sometimes you change these value on the fly depending on your goal. its better to know what these parameters do to your results rather than copying values from people's post.
--
As an afterwords, i hope you enjoy using this model, a like or a review definitely helps, or if you find it distasteful, you don't have to use it. there are many models out there that produces similar or superior results.
Description
More model optimization:
Lower dependencies on negative Ti (now only use ng_deepnegative_v1_75t + badhandv4
Better & more accurate low CFG Scale (5 ~ 10)
Wider Dynamic CFG Threshold (10+ needs DT)
Additional Quality bump on skin with fantasy / sci-fi prompts
Better Wet & Water effect (at SDE sampler)
Very Nice Quality with SDE Karras with Low Sampling Steps (16 / 8) with 4x-Ultrasharp can match v3.1 (50 / 25) with ERSGAN_4x+
Observed very high quality output with DDIM with (64/16) steps. DDIM is faster to calculate and produce more creative and detailed result than 2M Karras and is similar to SDE Karras result. However, it can be very tricky and hard to control the result at the moment - Use this to find a good seed.
Hotfix : VAE is baked in
FAQ
Comments (7)
Your prune for 3.7 seems to have gone wrong -- there's junk LoRA data still in there, and the VAE is mangled, so it won't load.
really? let me try and prune again, i made sure to delete the vae when pruning.. what was your problem? when you load the model it shows as error? -- tried and tested again with restarting SD, and i never have problems, how do you find the junk lora data? like how you see it? is there an extension to do that?
So that's exactly it, don't delete the VAE or it isn't a complete checkpoint, and software that actually validates it as such will fail out. If your merging results in a broken VAE and you don't want people using that, you can replace it with the one you recommend and simultaneously safely prune using the Model Toolkit extension for A1111.
@blooest i see, i never noticed because i only tested it directly not with a validation software, thank you for the info. I assume the solution to this is to simply bake the vae in?
Yup. And baking in the VAE you use saves you from having to require one in the description and hoping people read and download it, too. Win win.
Last version is not working for me.
":[InvokeAI]::ERROR --> Error while invoking:
'encoder.conv_in.weight'
I really love the model, the realism, but it has major issues with hands +++
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.