Illustrious Orchard - A Vivid Anime Blend
Illustrious Orchard is a hybrid SDXL model designed to generate vibrant and detailed anime-style images. It results from a curated fusion process, aiming to synergize the strengths of several popular anime models.
Core Fusion:
This model's UNet is constructed through a two-step weighted merge process:
Primary Blend: (1-b) * HassakuXLIllustrious_v2.1fix + b * AmanatsuIllustrious_v11. This step combines Hassaku's bright, distinct anime style (with the stability fixes and style adjustments of v2.1fix) with Amanatsu's characteristic detailed look and background stability.
Refinement Blend: (1-b) * the result from Step 1 + b * Illustrious-v1.1 UNet. This step further integrates the combined anime UNet with the more refined, natural language-focused characteristics of Illustrious-v1.1.
Key Components:
UNet: Custom merged as described above.
CLIP Text Encoder: Directly inherited from Illustrious-v1.1. This provides the model with Illustrious-v1.1's natural language understanding capabilities and prompt interpretation nuances.
VAE: Directly inherited from Illustrious-v1.1. This significantly influences the final image's color palette, saturation, and overall visual tone, aligning it closely with the Illustrious aesthetic.
Expected Characteristics:
Style: A bright, detailed anime aesthetic primarily influenced by Hassaku and Amanatsu, potentially exhibiting Hassaku v2.1fix's higher style variability and reduced artifacts compared to its predecessor. The Illustrious UNet contribution and VAE may add subtle refinements and influence color/anatomy.
Prompt Understanding: Should benefit from Illustrious-v1.1's CLIP, potentially offering better adherence to complex or nuanced prompts compared to models using different text encoders.
Versatility: Aims to balance the distinct anime styles of Hassaku/Amanatsu with the broader capabilities of Illustrious.
Usage Recommendations:
Inherits recommendations from source models:
Consider negative prompts like worst quality, bad quality, signature, logo, text, speech bubble. Depending on desired output, nose might be useful in negatives if faces appear overly realistic (as noted in Hassaku v2.1fix).
Consider positive prompts like masterpiece.
Franchise tags and metadata tags (e.g., highres) are likely ineffective or pruned.
Recommended SDXL resolutions: 1024x1024, 832x1216, 1216x832, 768x1344, 1344x768, etc.
Highres fix or detailer tools (like Adetailer) may still be beneficial for optimal eye and detail rendering, similar to Amanatsu.
License:
Due to the inclusion of HassakuXL (which itself incorporates other models), this model likely inherits the Fair AI Public License 1.0-SD. Please verify the licenses of all base models used.
Description
This is the inaugural release of the Illustrious Orchard series. It represents a specific two-stage UNet merging strategy aimed at creating a unique anime model blend.
Merge Recipe:
The UNet from HassakuXLIllustrious_v2.1fix (80%) was merged with the UNet from AmanatsuIllustrious_v11 (20%).
The resulting merged UNet (80%) was then subsequently merged with the UNet from Illustrious-v1.1 (20%).
The CLIP Text Encoder and VAE components were taken directly from the Illustrious-v1.1 model without modification.
Goals & Expectations:
The primary goal of this version was to leverage the bright, stable, and potentially varied anime style of Hassaku v2.1fix, enhance it with the fine details and background improvements noted in Amanatsu, and then refine this combination using Illustrious-v1.1's UNet, while fully adopting Illustrious-v1.1's language processing (CLIP) and visual output characteristics (VAE).
Users can expect an anime style that draws heavily from Hassaku and Amanatsu but possesses the distinct color signature and potentially better prompt following associated with Illustrious-v1.1 due to the shared CLIP/VAE. The final balance between the styles is a result of the specific 80/20 merge ratios used in sequence.
Feedback on this initial blend is highly encouraged to guide future versions of the Illustrious Orchard series.