File size: 3,705 Bytes
ecfe241 4de2925 d9f4bb3 4de2925 d581ccc d9f4bb3 792e0d6 34d7da5 3f32f56 d9f4bb3 f30a71a 5c6f55f f30a71a 2958ff6 d9f4bb3 d37777d d9f4bb3 f2ffa18 d9f4bb3 3e58cfc a89de79 f2ffa18 d9f4bb3 7370ccb d9f4bb3 7370ccb d9f4bb3 7370ccb d9f4bb3 4de2925 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | ---
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. |