V3 (Zapphire-tear)

Enhanced for 2.5D and 2D Mastery

This version is an upgrade from Blue Tear, offering sharper details, increased flexibility, and improved results across 2.5D, 2D, and even 3D images. If you enjoyed the previous iteration, the advancements in this version will stand out immediately.

What's New?

Building on the strong foundation of the earlier model, Crystal-Tear V3 has been updated to deliver better, more reliable outputs, with particular strengths in:

• Wardrobe Flexibility: The model excels at generating a wide variety of outfits, from casual to intricate, adapting seamlessly to the style and character in your prompt.

• Iconic Character Knowledge: With an extensive understanding of characters from anime and video games, it captures their distinct looks without needing additional LORA models. It handles popular series effortlessly, but for less mainstream or older titles, LORA integration can fill the gap perfectly.

• LORA Compatibility: This model works smoothly with LORA models, allowing for even more customization and refinement without losing its core strengths.

Whether you’re working on detailed portraits or complex scenes, the model’s increased flexibility ensures better results with minimal effort. Though it may take a few tries to nail certain complex prompts, the outcome is always visually satisfying.

Key Strengths:

• 2.5D and 2D Focus: Crystal-Tear V3 is built with these styles in mind but also performs impressively in 3D work.

• Portrait and Scene Excellence: The model handles both portraiture and elaborate compositions with ease, making it ideal for a wide range of artistic styles.

• NSFW-Ready: It continues to produce appealing results in NSFW contexts, with clear, consistent outputs.

For more detailed information on the updates and what this version can offer, check out the version info.

Support the Project!

Try Crystal-Tear V3 today and experience its enhanced flexibility and detailed outputs. If you love the results, consider supporting the project by uploading your work, sharing feedback, and leaving reviews. And if you're especially satisfied, a donation of Buzz will help keep the project growing!

---

V2 (Crystal-Tear)

This Version is the upgraded version of Blue Tear, it has better details and is even more stable compared to v1.

Most of the info about this model are already in the apposit section, so here i am gonna go more in dept on what this model can (and cannot) do.

What is this model for?

Since my last lora, i've received a dozen of request of pony version of them, but i had 0 experience on those, so i've never join that train (till now).

This model is the one i plan using as a base for my next works, so it had to have good understanding of the prompt, decent art, and it had to be capable of producing complex images without too many loras.

The v1 was a good step in that direction, but was still lacking, so after many reiteration, i've menaged to achieve my objective.

This model excells at generating portrait, but it can also used for complex illustration (The Gallery will be slowly updated in the next few days, to give you even more references)

i.e

Those kind of images can be easily accomplished with few keywords.

(P.s i've refrained to use upscale, and adetailer for those images, just to show you the "naked" model, if you use those the quality can truth-fully skyrocket).

Model strenghts

This model has a decent amount of knowledge about anime characters, so it can generate their images without the using of loras (as you can see from the gallery).

I find it decent enought, but truth be told, it struggles a bit with side-characters, or with serie that aren't so recent (but worry not - lora's exist for this reasons).

Does it support NSFW?

Welp, yes, it does. And it does it pretty well.

It's flexible, it understand positions and (aslong we don't mess with prompt, like when adding "solo" in a sex scene with 2 people) will produce a pleasing output.

Will this model receive more updates?

Possibly yes, i don't believe perfection exist, so aslong it can be improved i'll be there trying to improve it.

I like the model, how can i help?

Just downloading it and using it is enought. But if you truly want to help me, please review it (possibly spotting eventual problems, so i can work to fix them). Another appreciated thing is to upload your generated images here.

Please refrain to re-upload this model without permissions.

---

V1

Model based on Nova Xl, and created with anime merge.

描述:

Version Specifications for V3

Steps: 20 to 30

CFG Scale: 5 to 7

Resolutions: Same as SDXL (recommended resolutions for optimal output quality)

Highres Fix: Suggested; Ultrasharp Upscaler works well for enhancing details

Prompting Tips:

Pony Score Tags: These tags significantly impact the style, and experimenting with different combinations can yield diverse results.

Score_9, score_8_up, score_7_up: These tend to add more details suited for 2.5D and photorealism.

2D Results: To maintain a 2D style, use "source_anime" in the positive prompt and "gamecg, realistic, photorealistic, 3D" in the negative prompt for optimal separation.

Character Recognition and Styling:

Iconic Characters: This version can easily recognize most iconic female characters from anime and video games.

Male Characters: While the model can produce masculine body types, the range of pre-trained male characters is more limited.

Composition:

Backgrounds: The model tends to generate simple backgrounds by default. You can customize this by specifying "indoors" or "outdoors" if no specific background is desired.

Camera Angles: All common camera angles are recognized and handled accurately.

Clothing Flexibility: The model is highly flexible when it comes to outfit customization. You can modify clothing styles extensively by using appropriate tags.

Known Biases:

Halterneck Bias: If an image includes an unwanted halterneck or similar features, you can correct this by adding "halterneck, collar, detached collar" to the negative prompt.

Stability:

Hands: Hand generation is more stable in this version. If you encounter imperfections, retry using the same seed with "extra" enabled (if using Automatic1111).

Sampler Recommendation:

Euler a is the suggested sampler for achieving consistently high-quality results.

训练词语: