
1. Introduction
This is a LoRA for photorealistic images of still life, especially of porcelain that has a pattern called "qing hua ci" (青花瓷 qīng huā cī). Training resources are taken from photographs under Commons published on the web, and I would like to serve a LoRA made with 65 images.
This LoRA was trained on my computer (Ryzen 5 2600 + 16 GB RAM + 500 GB SSD + NVIDIA GeForce RTX 2060 6 GB) with SD-Trainer by Akegarasu (https://github.com/Akegarasu/lora-scripts) on sd-scripts by kohya-ss (https://github.com/kohya-ss/sd-scripts/tree/main), and well known configurations of training as follows (I don't change most default parameters):
-
DreamBooth method without regularization images
-
pretrained model: ULTRA v3 (for SD 1.5) https://civitai.com/models/228525?modelVersionId=470417
-
resolution = 512 x 512
-
epochs = 20
-
optimizer type = "Lion"
-
steps = 1,960
2. Usage
To be a simple way in your prompt, please add LoRA decralation with some weight value just as "<lora:BT_Qing_Hua_Ci_v1:1>" and put the trigger keyword "qing hua ci" in some place of your prompt. In most cases, a keyword of style will be better to write a later position in your prompt.
A weight value can vary a result, you know. When you set "0," no effect will be shown, and when you set "1" or more value, the result may be worse by the well-known "overfitting," and most images are likely resembled with the original images used in my training. So, you will have to set the value in your own situtation (what other LoRA you use, and where do you put a keyword in your prompt, and son on.) If you set a weight over 1, the overfitting may be happen. But most results are good even by a weight of 1.0 as the following image.
A value for CLIP SKIP will be available not only 2 but also 1.
Anyway, here's a sample image.
Prompt: (still life, simple background, blue background, studio lighting, flashlight), Chinese imperial dish with flowering prunus pattern, that is called ((qing hua ci)), Chinese porcelain with blue underglaze decoration, Ming dynasty blue and white porcelain of qing hua ci, porcelain dish decorated with cobalt blue underglaze, Celadon-glazed porcelain dish with underglaze blue decoration, Jingdezhen porcelain dish with blue and white design, Exquisite porcelain dish with a serene blue hue, Timeless porcelain with intricate blue patterns of qing hua ci, a delicate, circular porcelain plate with a beautiful qing hua ci floral pattern, a large, white porcelain dish plate adorned with intricate blue floral designs of qing hua ci. <lora:BT_Qing_Hua_Ci_v1:0.7> BREAK
no humans, still life, unparalleled masterpiece, finely detail, incredibly absurdres, perfect artwork, highres, intricate, UHD, HDR, 32K, 16K, promotional poster, natural lighting, professional photograph (in style of Tina Modotti:1.5), vibrant details, abstract beauty, studio quality, UHD, professionally color graded, (ethereal and emotive photography, professional color grading, rich deep colors, studio quality, color-accurate, spectrum-precise, perceptually accurate, color calibrated, tone-mapped), (professional photograph with Fujifilm X-T5, XF 56mm F1.2 R WR, Fujifilm Superia 800) <lora:photo_aesthetic_slider:0.7> <lora:hyperdetailer_v080:0.3>
Negative prompt: (multiple angles, multiple views, collage, multiple images, frame, border, split frame, split screen, humans, girl, boy:1.5), easynegative, badquality, verybadimagenegative_v1.3
Steps: 35, Sampler: DPM++ 3M SDE, Schedule type: Polyexponential, CFG scale: 7.5, Seed: 3601092209, Size: 1024x576, Model hash: 2289c564ac, Model: ultra_v3, VAE hash: 63aeecb90f, VAE: ClearVAE_V2.3.safetensors, Lora hashes: "BT_Qing_Hua_Ci_v1: 551bc3935f37, photo_aesthetic_slider: b16a0b5c2256, hyperdetailer_v080: 05dfb0659165", TI hashes: "easynegative: c74b4e810b03, verybadimagenegative_v1.3: d70463f87042", Version: v1.10.1
3. Weight
[Notice] It is hard to add images on the lower part of web page on this editor application, so unfortunately I will not use explanatory images below.
Results would be varied by a weight value in LoRA declaration, depending on other LoRA you use, the position of trigger word ("qing hua ci") in your prompt, base model, or others. Generally speaking, a weight under 0.3 will be ineffective and a weight over 0.7 can make an overfitting.
And if you use this LoRA for other purpose such as a pattern of clothes, you will not get a better result. This LoRA will make an effect just like a blue lens filter.
4. the Last word
Have fun, and thanks to users, creators of scripts, applications, models, and at last but not least, to crews at Civitai.
描述:
The initial version is here.
训练词语: qing hua ci
名称: BT_Qing_Hua_Ci_v1.safetensors
大小 (KB): 32351
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success