RickC_SuCaiV1_Flux_lora

Hi everyone, I'm RickC.

I'm thrilled to share my very first trained LoRA – a result of my initial foray into model training, built upon the Flux.1 D fp8 base model!

I've named it "SuCai" (素彩), derived from its core characteristics: "Su" (素) represents the sketch/line art style often applied to the background, while "Cai" (彩) signifies the vibrant colors preserved in the main subject. My goal with this LoRA is to forge a unique visual experience, seeking a fascinating balance and blend between flatness and volume, 2D and 3D, the stylized and the implied real.

What can the "SuCai" LoRA do?

This LoRA isn't just for generating appealing webtoon-style characters (both male and female); it excels at creating a unique line-and-color combination effect (colored subject + line art background). Whether you're depicting landscapes, flora, or even interpreting beloved IP characters (e.g., Asuka, Hatsune Miku, Kamen Rider, Saitama, Goku, 2B – try using their names directly in the prompt, you might get some fun surprises!), it consistently delivers the distinct 'SuCai' visual flair.

Core Functionality & Features:

Its most prominent feature (especially noticeable at higher weights) is its ability to render the main subject with vivid colors and detail, while simultaneously transforming the background into a richly detailed, structurally defined monochrome line art or sketch style. This strong contrast creates a distinctive 2.5D visual effect and an engaging sense of "dimensional fusion."

How to Use the "SuCai" LoRA:

• Trigger Word:

  • Include sc_comic in your prompt to activate the LoRA.

• Weight:

  • Works effectively within a weight range of 0.5 to 1.0, showcasing varying style intensities.

  • Higher weights generally result in a more pronounced line art background and clearer color separation for the subject.

  • I recommend starting around 0.8 to experience the signature [Colored Subject + Line Art Background] 'SuCai' effect, then fine-tune based on your needs.

• Prompt Structure - Important Tip!:

  • (1) Define the [Colored Subject] Clearly:

    • Describe your main subject in detail. Use prompts to emphasize the parts you want colored, for example: ... (woman wearing a (vibrant red dress:1.1) and (colorful makeup:1.1)), colored skin ... or ... (gleaming (blue car body:1.2)) .... This helps guide the LoRA's color focus.

  • (2) Describe a Complex [Line Art Background]:

    • The more complex and detailed the background structure (e.g., detailed cityscape background, ornate library interior), the better the visual effect when converted to line art.

  • (3) Use [Quality Tags]:

    • Add tags like masterpiece, best quality, high detail, illustration to improve overall quality.

  • (4) Optional [Style Tags]:

    • Reinforce the concept with terms like lineart style, dimensional blend, 2.5D effect, sketch background, SuCai style, chroma line style.

  • (5) [Avoid] Conflicting Terms:

    • Please do not use terms that conflict with the target style, such as realistic, photograph, photorealism, 3D render.

Generation Recommendations (Based on limited testing):

• Steps: Recommend setting between 20 - 25 steps.

• Recommended [Sampler + Scheduler] Combinations:

  • Based on my tests, these combinations performed well. Feel free to experiment:

    1. Euler + ddim_uniform

    2. Euler + normal or simple

    3. DDIM + sgm_uniform

    4. Deis + beta

  • You're also welcome to try other popular combinations like DPM++ 2M Karras, etc.

Some Honest Words (Please Read): Limitations & Future

This is my first attempt at training a LoRA, learned entirely through self-exploration, so there are bound to be imperfections. Thank you for your understanding.

• Hand & Feet Issues: Although I included ~15% training images specifically targeting hands and feet, the current version still has a chance of generating distorted hands/feet. This is a key area I plan to study and improve upon!

• Limited Generalization: Due to training on the cloud with its inherent limitations, the dataset size used for this version was modest. Consequently, the LoRA's generalization ability might not be very strong, potentially underperforming on less common concepts or styles.

• V1 Preview Release: This is primarily a V1 preview released for sharing and gathering feedback. For now, it's mainly intended for online generation platforms. I will continue learning and gaining experience, aiming to fix existing issues or potentially retrain a more robust version in the future.

Final Words:

Despite its current limitations, I sincerely hope this exploratory "SuCai" LoRA brings some novel fun and unique visual sparks to your creative process!

🌟 I'm eagerly looking forward to seeing your amazing image shares! The more, the better! 🌟

Your creations are the best way to test and showcase the LoRA! Please unleash your creativity and share the stunning images you generate using "SuCai". Your shared images and any feedback (whether praise or constructive criticism) are incredibly valuable to me and will be my greatest motivation to keep learning and improving!