| --- |
| license: apache-2.0 |
| base_model: |
| - lodestones/Chroma1-HD |
| tags: |
| - lora |
| - 3dcg |
| - anime |
| - realistic |
| - liminal |
| - interior |
| - manga |
| - sfw |
| - nsfw |
| --- |
| |
|  |
|
|
|
|
|  |
|
|
| # 3DCG Indoor Character + Manga combo LoRA |
|
|
| Creates **3DCG-style female characters** in indoor environments with strong emphasis on: |
|
|
| - natural elbow / neck / shoulder posing |
| - expressive face close-ups |
| - believable hip & torso movement |
| - pleasant 3DCG skin rendering (works for both SFW and NSFW) |
|
|
| ## Strongest styles / vibes |
|
|
| - Candid 3DCG animation stills (Drills3D / KawaiiDetectiveEnthusiast style) |
| - Modern interior design + depth of field / background blur |
| - Liminal / backroom / empty mall aesthetic (Pinterest mood) |
| - Manga frame composition — Claymore + Tsutomu Nihei (*BLAME!*) influence |
| - Red-carpet / award-event posing (SFW photos of adult industry performers) |
|
|
| ## Helps fix / compensate for Chroma limitations |
|
|
| - Reduces unwanted cartoony / furry 3D look |
| - Improves sci-fi aesthetic compatibility |
| - Better anatomy on arms, necks, elbows, hips in longer shots |
|
|
| ## Training details |
|
|
| The dataset consists of 250 1024x1024 images with 4 columns of a mix of CG , manga , photos and digital art. |
|
|
|  |
|
|
| Shown above: 4 1024x1024 training frames from the dataset. Captions were created using gemini 2.5 flash image captioning. The text captions are provided as a zip file in this repository. |
|
|
| - **Dataset**: 250 selected frames of CG and manga frames |
| → cropped & composed into 4×1 columns in 1024×1024 tiles |
| - **Repeats × Epochs**: 4 repeats × 40 epochs |
| - **Network rank / alpha**: 32 / 32 |
| - **Learning rate**: 0.0035 using 'cosine with repeats' scheduler |
| - **Trainer**: Kohya_ss (no bucketing) |
| - Mostly synthetic data + SFW Pinterest / Flickr reference photos |
| - Keyword is in negative prompt : phrase 'this was used as negative when training the lora' |
| ## Ethics & transparency |
| |
| - Cannot reproduce real people (no real-person likenesses) |
| - All photographic content used = public SFW images only |
| - Training image selection & preparation scripts → included in this repo (Jupyter colab notebooks) |
| |
| If you wish to run this lora model online without any need to download , you can check out the lora on my TA profile: https://tensor.art/u/754389913230900026 |
| |
| Cheers! / Adcom |
| |
| ## Example usage in Chroma |
| |
| ```text |
| Text encoder (suggestion): t5xxl_fp16.safetensors (especially with long descriptive prompts) |
|
|
| Weight range: 0.8 |
|
|
| Example prompt: This image prominently features a solo 1girl, a young female, standing indoors in a room that appears to be a living area. She has long_hair that is brown_hair, styled neatly in a ponytail, and her brown_eyes are directed in a way that suggests she is looking_at_viewer with a gentle expression. She is wearing casual clothing, specifically a light-colored short_sleeves t-shirt, which can be described as a white_shirt, featuring a detailed print_shirt design with a notable 'g'raphic element towards the bottom. On her right wrist, she wears a simple bracelet, which is the only piece of jewelry visible on her. In the background, there is a robust wooden bookshelf filled with numerous books, and a dark brown door is visible on the right side of the frame, firmly establishing the setting indoors. A segment of a couch is also partially visible in the lower left foreground. |
|
|
| Negative prompt: this was used as negative when training the lora. |
