Winter Cap With Flaps (Exploration of LoRAs)

Trained on 10 images.

SD:

Anything V5 and Moonbeam are both popular AI checkpoint models used on platforms like PixAI and Stable Diffusion for generating 2D anime-style illustrations. While both excel at high "anime plasticity" and clean character designs, they differ slightly in their default aesthetic, line work, and color shading.

• Core Differences:

Anything V5:

• Focuses on crisp, highly versatile, classic anime and cel-shaded aesthetics.
• Adheres strictly to a wide range of standard prompts, making it a reliable, all-around benchmark model for general anime characters, poses, and detailed outfits.

Moonbeam:

• Tends to lean toward a softer, slightly more atmospheric, and luminous rendering style.
• Often preferred for specific fantasy, glowing, or ethereal lighting effects on characters, blending smoothly with specialized anime LoRA styles.

SDXL:

Haruka V2 is a specialized, in-house fine-tuned anime model created by PixAI, while Illustrious (Illustrious XL) is a broader, foundational base model architecture designed for general high-detail anime generation. In fact, Haruka V2 is built upon the Illustrious ecosystem, meaning they share underlying compatibility while serving different creative goals.

Core Differences:

Haruka V2 (PixAI Fine-tune):

• Aesthetic: Focuses on soft pastel tones, smooth clothing textures, and vibrant hair shine.

• Details: Highly optimized for expressive facial features, detailed eye highlights, and cleaner hand structures.

• Use Case: Best for clean, polished, expressive individual character portraits straight out of the box on PixAI.

Illustrious (Base Model Family):

• Aesthetic: Highly versatile and adaptable across a vast spectrum of different anime art styles.

• Recognition: Built with broad data to natively recognize and render a massive library of popular pre-existing anime characters.

• Use Case: Acts as a general foundation model used by creators to generate diverse styles or to train custom LoRAs.

DiT:

SD and SDXL LoRAs allow weights up to 2.0 because their older UNet architectures require higher scaling to overcome base model traits. While DiT.1 (diffusion transformer) models have much higher prompt adherence and structural sensitivity, meaning weights above 1.2 quickly break anatomy and distort the image.

Architecture Differences

• SD / SDXL (UNet): Built on older convolutional layouts that often require pushing a LoRA past 1.0 to forcefully express specific styles or clothing details.

• DiT.1 (Transformer): Uses a transformer-based backbone that responds aggressively to parameter changes. Higher weights over-correct the internal attention blocks rapidly.

Sensitivity and Limits

• SD/SDXL Range (0.0 to 2.0): Gives leeway to boost weak or undercooked community-trained files, though values past 1.4 often introduce deep-fried artifacts.

• DiT.1 Range (0.0 to 1.2): Capped lower by platform constraints because the transformer network saturates fast; a weight of 1.0 on a DiT LoRA acts like a 1.5+ strength on SDXL.

DiT.1 & DiT.2:

On PixAI, DiT.1 and DiT.2 represent different generations of Diffusion Transformer model architectures and training bases. DiT.1 uses the original Tsubaki base model, while DiT.2 uses the advanced Tsubaki.2 architecture, which yields higher quality, better multi-character composition, and deeper prompt comprehension.

• Core Differences

Base Models: DiT.1 trains against Tsubaki, whereas DiT.2 trains against Tsubaki.2.

Output Quality: DiT.2 produces sharper details, better lighting, and cleaner overall rendering than DiT.1.

Training Time: DiT.1 (Tsubaki) jobs finish in about 70 minutes, while DiT.2 jobs take roughly 2 hours.

Trigger Word Rules:

• DiT.1: Uses concise, short names or tags.

• DiT.2: Requires a detailed feature description of at least 30 characters (reading like a mini inventory of attributes rather than a single name).

DiT.3:

On PixAI, DiT.3 represents the evolution of the diffusion transformer architecture following DiT.1 (the original generation built around Tsubaki and Serin) and DiT.2 (which powered Tsubaki.2 with higher quality and slower training).

Key Differences of DiT.3

• Tsubaki.3 Flagship Model: DiT.3 is represented by models like Tsubaki.3, which improves upon the multi-character composition and spatial relationship handling of DiT.2.
• Superior Spatial and Multi-Character Binding: While older DiT generations could struggle with swapping or bleeding traits across multiple subjects in complex scenes, DiT.3 much more accurately maps clothing, positioning, and descriptions to the correct individual (e.g., keeping specific outfits, hair colors, and actions locked to the left or right character as written in prose).
• Natural Language Prose Understanding: Like earlier DiT models, DiT.3 relies on an LLM-style text encoder rather than tag-based lists (unlike SDXL models such as Haruka v2), but DiT.3 translates complex storytelling and positional clauses ("on the left", "leaning sideways", "book in her lap") with fewer positional mix-ups.

Evolution Summary:

• DiT.1 (Tsubaki, Serin): First-generation DiT architecture; basic natural language comprehension and early multi-subject handling.

• DiT.2 (Tsubaki.2): Second-generation architecture; upgraded image fidelity and dynamic action handling, though prone to occasional feature or clothing swaps in crowded 2+ character layouts.

• DiT.3 (Tsubaki.3): Third-generation architecture; tighter semantic adherence for multi-character placement, more reliable attribute binding (preventing swapped clothes or misplaced features), and refined natural-language scene control.