Not loaded
Download
1 variant available
21
184
Model Name: Impasto Anime Style (HE Edition) v0.8-beta (Moss)
⚠️ EXPERIMENTAL BETA BUILD - STILL IN TESTING
Style: Impasto / Soft-focus Painterly
Compatible Base Model: Moss-Image-Base
This is an experimental preliminary training run on the Moss-Image base model, representing a significant shift in our data pipeline.
Instead of a massive dataset, the training set has been intentionally condensed and highly curated down to 1,700 images. Every single image was meticulously processed using Qwen-Image-2.1 to specifically optimize and refine captions for hair, clothing, and skin textures. Additionally, the captioning structure has been redesigned to closely mimic the natural language prompt structure of Krea3, aiming for smoother and more intuitive prompt comprehension.
Current State & Limitations:
* Reference Generation: The original goal of this run was to achieve strong reference-based image generation (Image-to-Image consistency). While this specific capability is not yet fully realized in this early stage, the model remains highly effective at controlling the artistic style in Text-to-Image generation.
* Style Shift: Compared to previous versions, this model introduces a stronger hazy, dreamy, and soft-focus aesthetic (enhanced atmospheric feel). As a trade-off for this softness, the lighting and shadow contrast may appear slightly flatter or less pronounced.
Usage Recommendations:
* LoRA Weight: 0.6 – 1.0 (Recommended starting point: 0.8. Since this is an early training run with a low learning rate, a moderate-to-high weight helps bring out the style without breaking the image.)
Disclaimer: This model was trained on publicly available fan-created content and is intended strictly for personal, non-commercial use.
Training Details:
* Base Model: Moss-Image
* Dataset Size: 1.7k images (Highly curated and compressed)
* Captioning: Qwen-Image-2.1 + Krea3-style prompt structure
* Steps: 1,925 (Preliminary training run)
* Rank: 16
* Learning Rate: 1e-5
* Toolkit: Ostris AI Toolkit
Model Name: Impasto Anime Style (DH Edition) v0.6-beta (Krea2-Raw)
Style: Impasto / Painterly
Compatible Base Model: Krea2-Raw
This version marks a major architectural shift, fully retrained on the Krea2-Raw base model. The dataset has been significantly expanded to 15k images.
To ensure the highest quality and prompt adherence, this version introduces a rigorous captioning pipeline: all images were captioned in JSON format using Qwen3.8-Flash, followed by strict manual spot-checking and review. This results in much cleaner training data and better responsiveness to natural language prompts compared to previous versions.
As always, character identities were not labeled or tagged. This model remains an experimental attempt to generalize the impasto aesthetic, prioritizing artistic style and texture over character fidelity.
Usage Recommendations:
* LoRA Weight: 0.7 – 1.3 (Recommended starting point: 1.0. Krea2 base models can be sensitive to LoRA weights, so start lower and adjust to taste.)
* Prompting: Use natural language descriptions directly.
Disclaimer: This model was trained on publicly available fan-created content and is intended strictly for personal, non-commercial use.
Training Details:
* Base Model: Krea2-Raw
* Dataset Size: 15k images (Expanded multi-game anime fanart)
* Captioning: Qwen3.8-Flash (JSON format) + manual spot-check & review
* Steps: 2,469
* Rank: 16
* Learning Rate: 1e-4
* Toolkit: Ostris AI Toolkit
Model Name: Impasto Anime Style (Anima Edition) v0.5-MQ
Style: Impasto / Painterly
Compatible Base Model: Anima Base (circlestone-labs/Anima)
This LoRA represents a significant evolution from the previous Z-Image-Turbo version, now trained on the Anima Base architecture using firdosi/diffusion-pipe. The dataset has been massively expanded to 11.3k images, curating thick-paint (impasto-style) fan artworks across multiple titles including Genshin Impact, Honkai: Star Rail, Zenless Zone Zero, and Blue Archive.
As with the initial release, the training focus remains strictly on capturing distinctive painterly textures and brushwork rather than character fidelity. Character identities were not labeled or tagged during training; this model is an experimental attempt to generalize the impasto aesthetic across a broader range of anime-style source material.
Usage Recommendations:
* LoRA Weight: 0.3 – 1.2 (Recommended starting point due to the larger dataset and higher rank)
* Prompting: Use natural language descriptions directly. The model responds well to intuitive prompts describing texture (e.g., "thick oil paint," "visible brushstrokes," "impasto style").
Disclaimer: This model was trained on publicly available fan-created content and is intended strictly for personal, non-commercial use.
Training Details:
* Toolkit: firdosi/diffusion-pipe
* Base Model: Anima Base
* Dataset Size: 11.3k images (Multi-game: Genshin, HSR, ZZZ, Blue Archive...)
* Captioning: Qwen3.5 + manual correction]
* Steps: 1,313
* Batch Size: 8
* Rank: 16
* Learning Rate: 3e-5
Model Name: Impasto Anime Style (ZIT Edition) v0.1-HO
Style: Impasto
Compatible Base Model: Z-Image-Turbo
This LoRA was trained on a curated set of ~150 thick-paint (impasto-style) Genshin Impact fan artworks from Pixiv. The training images were selected for their distinctive painterly texture and brushwork. Note that character identities were not labeled or tagged during training—this is primarily an experimental attempt focused on capturing artistic style rather than character fidelity.
Usage Recommendations:
LoRA Weight: 0.8 – 1.0
Prompting: Use natural language descriptions directly. (Note: Training captions were structured Chinese phrases, but the model responds well to intuitive prompts.)
Disclaimer: This model was trained on publicly available fan-created content and is intended strictly for personal, non-commercial use.
Training Details:
Toolkit: Ostris AI Toolkit
Captioning: Qwen3-VL + manual correction
Image resolution: 256–512px
Steps: 1,350
Rank: 4
Learning rate: 5e-4