建了个群方便大家交流：695960980
Hello everyone, it has been a very long time since the last update. I haven't been completely resting; I still love SD and have been working on training the Ronghua model.
This time, I am bringing you version 4.5 of Ronghua. Version 4.5 uses the tool available at https://github.com/jiayev/GPT4V-Image-Captioner for GPT4+WD14+ manual caption. The model has been further trained on the basis of Ronghua 4 preview1(an early internal fine-turn version) ,using an RTX 6000 Ada for about 30 days.
大家好，距离上一次的更新真是过了很长很长的时间，我其实并没有完全停下来休息，我依然热爱SD，一直在做有关容华模型训练的事情。
这次为大家带来的是容华的4.5版本，4.5版本采用 https://github.com/jiayev/GPT4V-Image-Captioner 工具进行GPT4+WD14+人工打标，模型在容华4 preview1（一个早期的内部训练版本）的基础上继续训练，使用RTX 6000 ada训练了大约30天。
推荐CFG: 3.5
Recommend CFG: 3.5
推荐的负面词：
Recommend Negative prompt: (worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art:1.4), (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.2), (blur, blurry, grainy), morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, amateur:1.3), (3D ,3D Game, 3D Game Scene, 3D Character:1.1), (bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities:1.3)
Below are some revised and newly added ancient style concepts (limited by personal ability, not all concepts are guaranteed to be accurately caption)：
以下是部分微调/新增的古风概念（个人能力有限，无法保证全部概念打标正确）：
### Traditional Chinese Musical Instruments
- Dizi (笛) - A traditional Chinese flute.
- Xiao (箫) - A Chinese vertical end-blown flute.
- Pipa (琵琶) - A four-stringed Chinese musical instrument, which comes in various forms like:
- Gudai Pipa (古代琵琶) - Ancient pipa.
- Earth Pipa (土琵琶) - A smaller version of pipa.
- Bent-neck Pipa (曲项琵琶) - Pipa with a curved neck.
- Guqin (古琴) - A plucked seven-string Chinese musical instrument of the zither family.
### Chinese Architectural and Landscape Elements
- Stone Arch Bridge (石拱桥) - A type of bridge featuring a curved arch structure.
- Scholar's Rocks (文人石) - Naturally occurring or shaped rocks which are traditionally appreciated by Chinese scholars.
- Bamboo Raft (竹筏) - A raft made out of bamboo.
### Chinese Clothing and Accessories
- Veiled Hat (帷帽) - A traditional Chinese hat with a veil.
- Fengguan (凤冠) - A type of Chinese phoenix crown worn by noble women.
- Jade Bracelet (玉手镯) - A bracelet made from jade, often worn for both decoration and symbolic protection.
- Zan (簪) - Ornamental hairpins used in traditional Chinese hairstyles.
- Embroidery Shoes (绣花鞋) - Traditional Chinese shoes decorated with embroidery.
### Traditional Chinese Games and Leisure Items
- Weiqi (围棋) - Known in the West as Go, a strategy board game.
- Wooden Swing (木制秋千) - A swing made of wood.
- Rattan Ball (藤球) - A ball made from rattan, used in traditional games.
### Traditional Chinese Writing and Art Tools
- Scroll (卷轴) - A roll of parchment or silk used for painting and writing.
- Bamboo Scroll (竹卷轴) - A scroll made from bamboo.
- Bamboo Book (竹书) - An ancient Chinese book made from bamboo.
- Xuan Paper (宣纸) - A kind of paper originating from Xuanzhou city, used for calligraphy and painting.
- Silk Curtains with Calligraphy (书法丝帘) - Silk curtains adorned with calligraphic art.
### Traditional Chinese Household Items
- Traditional Fan (传统扇子) - Comes in various forms like:
- Round Fan (团扇)
- Folding Fan (折扇)
- Embroidered Fan (绣扇)
- Paper Umbrella (纸伞) - A type of umbrella made from paper.
- Pottery Jug (陶壶) - A jug made from pottery.
- Ceramic Wine Vessels (陶瓷酒器) - Vessels used for serving and drinking wine.
- Embroidery Hoop (绣花箍) - A tool used to hold fabric taut while embroidering.
- Incense Burner (香炉) - A container for burning incense.
### Other
- Imperial Memorial (奏折) - A formal report or statement written to the emperor.
Additionally, below are some high-frequency tags:
另外以下是部分高频tags：
文化服饰 - Cultural attire
hanfu, gufeng, cultural attire, cultural costume, historical costume, hanfu clothing, embroidered silk robe, embroidered garments, ruqun, tang dynasty style, embroidered robe, ancient chinese attire, traditional makeup, traditional chinese hairpin, phoenix crown, traditional hanfu attire, phoenix coronet, chinese traditional clothing, jewelry, embroidered garment
场景设置 - Scene setting
Indoor setting, outdoor setting, bedroom setting, ancient Chinese architecture
拍摄和视觉效果 - Filming and visual effects
Soft lighting, vibrant colors, natural light, warm color palette, pastel colors, high-quality, soft focus background, blurred background, bokeh effect, clear image quality, warm lighting, cinematic composition, ambient lighting, sharp focus, depth of field, shallow depth of field, dramatic lighting
人物特征和表情 - Character features and expressions
Serene expression, youthful appearance, Asian ethnicity, black hair, side profile, makeup, eye contact, long hair, playful expression, looking at camera, young adult, contemplative, poised posture, elegant posture, seated pose
幻想和角色扮演 - Fantasy and role-playing
Fantasy, character portrayal, animal ears, fantasy theme, anime-inspired, cosplay photography, costume play, cat ears, tail
历史和电影 - History and movies
Historical drama, xianxia, wuxia, period film, film still
NFSW：
cleavage, sexy, suggestive pose, thighhighs, high heels, seductive expression, sensual, large breasts, nipples, panties, sheer fabric, see-through, fishnet stockings, boudoir photography, nude
=======================
Big thanks to these friends for the buzz tips support! (Listed in no particular order, up to 2023/10/23): 9812chenditemin561、aibernardbittler714、Anbachi2021、chirenshuomeng、Dklove、dkorbat、firisu、jeffca2525105、jiutaogong、mmcCaterpillar、mr7th、nakaaachin、Radeux、rogerwang、Sini5555、spacemonkeyx01473、TracQuoc、wordsmaster、xibik10354751、yama4343、yi_yu、yuleirem123286、zhixuan08299144
***RongHua 1.0 and 1.1 Not include NSFW content, maybe next version...***
【容华】是一个服装、道具、化妆，都经过特化的国风模型，得益于SDXL庞大的参数量，它兼容多种画风，你可以通过不同的提示词组合画出各种风格化的图像。
Currently, [Ronghua] has not merged any other models, and the model is based on SDXL Base 1.0, which has been trained for more than 150+ hours. All materials are from the Internet.
Ronghua still has many problems, welcome everyone to use it and give me suggestions.
This is the first time I have done ft for a large model.Ronghua’s birth was thanks to the opinions and support of the group friends, especially Neo_cute, who gave me direction and confidence, as well as Redpriest9527, aistha, tonyhs, MengX, Faispirit, LYuyang, Maper_leaf, Kelly_V, kwResearch and many other great works that have improved Ronghua’s artistic level.
There is also the creator of the paradise, Merjic, and all the brothers and sisters, who make me happy and fulfilled every day.
===================================
第一次做大模型的ft，摸爬滚打跌跌撞撞，容华的诞生得亏群友意见和支持，特别鸣谢Neo_cute，给了我方向和信心，还有Redpriest9527、aistha、tonyhs、 MengX、Faispirit、LYuyang、Maper_leaf、Kelly_V、kwResearch 等等很多大佬的作品提升了容华的艺术水平。
还有乐园的创建者麦橘大佬，和里面的所有哥哥姐姐们，弟弟妹妹们，让我每天都很快乐充实。
Description
FAQ
Comments (56)
请问作者用了多少张图片训练呀
新的模型光和质感都很漂亮，期待更多更好的国风模型
另外很想了解训练大模型需要什么素材？有没办法可以做到提示词区分朝代风格，提示词可以对背景进行更丰富的描述，还有人的表情，我试过很多大模型都做不到伤心哭泣的表情
@admyp291 SDXL的大模型训练各种类型画风的素材都可以的。要区分朝代风格，目前SDXL的训练可能不太行（text encoder原因），我认为目前以lora去实现是更好的方式。伤心哭泣的表情我去试试，的确这类素材太少见。
还有一些控制景别的提示词的用处不大，比如cowboy shot, medium shot, close_up, full body, 或者一些控制动作的词
@admyp291 不起作用的原因，我想应该是因为“cowboy shot, medium shot, close_up, full body”这些tags其实属于danbooru，SD1.5从SD1.4继承了这些tags，而SDXL并没有。
好喜欢你这个模型，可惜显卡跑XL太痛苦，大佬能整个SD1.5这种风格模型么
做出来的图都是噪点是怎么回事？
我的也是 请问您解决了吗
可以发一下你们的出图参数
@fok3827 按照这个模型的示例图参数来的，也全是噪点
是不是vae没关？我之前用1.5版本的vae出的图都是模糊不清的
@a17371250000181 你提醒我了，SDXL的vae和SD1.5的vae是不通用的，请大家把外挂VAE设置成NONE，模型已经内置了VAE，不用外挂
关掉外挂VAE就好了
麦橘的乐园有链接吗？怎么进入呀？谢谢想交流一下
Very cool
可以教炼模型吗？想跟你学
测试了这个模型的NSFW内容的生成能力，比较可行的思路是这样的：1.在提示词上只描述人物、动作、场景，不能写任何服装相关的词。2.在负面提示词中填写衣服、裤子、内衣内裤等单词。3.如果还是没起效果，需要在正向提示词上边加强化提示词，比如：赤裸，大、中、小乳房等。
厉害
模型加载不了是为什么呀
显存太小？
@liuxy 显存是16g的
有没有可能，下载的文件损坏？
估计是SD的版本问题，我的也加载不了，把启动器更新至1.5.1试试
同样问题，最新1.6webui，
模型有問題，運行不了，我下了几遍了
Currently allocated : 3.44 GiB
Requested : 3.12 MiB
Device limit : 4.00 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB
你得说具体原因啊……光说一句模型加载不了别人怎么帮你
同样加载不了，奇怪，明明列表里有，怎么就没法选择呢。
支持国产！支持国风！
诚邀各位大佬一起攻破《洛神赋》里边的画面：她的形影，翩然若惊飞的鸿雁，婉约若游动的蛟龙。容光焕发如秋日下的菊花，体态丰茂如春风中的青松。她时隐时现像轻云笼月，浮动飘忽似风吹落雪。远而望之，明洁如朝霞中升起的旭日;近而视之，鲜丽如绿波间绽开的新荷。她体态适中，高矮合度，肩窄如削，腰细如束，秀美的颈项露出白皙的皮肤。既不施脂，也不敷粉，发髻高耸如云，长眉弯曲细长，红唇鲜润，牙齿洁白，一双善于顾盼的闪亮的眼睛，两个面颧下甜甜的酒窝。她姿态优雅妩媚，举止温文娴静，情态柔美和顺，语辞得体可人。洛神服饰奇艳绝世，风骨体貌与图上画的一样。她身披明丽的罗衣，带着精美的佩玉。头戴金银翡翠首饰，缀以周身闪亮的明珠。她脚着饰有花纹的远游鞋，拖着薄雾般的裙裾，隐隐散发出幽兰的清香，在山边徘徊倘佯。忽然又飘然轻举，且行且戏，左面倚着彩旄，右面有桂旗庇荫，在河滩上伸出素手，采撷水流边的黑色芝草。
找不到联系模型作者的办法，出此下策，就直接在这里留言了：因为我是重度的古文诗词爱好者，所以对能呈现国风的东西都很热忱。如果我们从学术上说，所谓的国风出现“美学”观念，从务实到务虚的过渡中，最重要的分水岭是战国时期出现了庄周，就是那个做梦，梦到自己是一只飘逸自在的蝴蝶，醒来时弄不清楚到底是那只蝴蝶梦到了自己是庄周，还是庄周梦到了自己是蝴蝶的庄子。庄子这人的学问本质上对实际的事务是毫无作用的，比如功名利禄，升职加薪、迎娶白富美、走上人生巅峰…这些常见的动物本能是没有任何关系的。庄子之学本质就是最没用的无用之学，他说有一条不知道长度有多少千里的大鱼（没有破万），大得离谱，我们知道日本最南端到最北端也才3500公里，那这条鱼也许最大比日本国土面积还要大近2倍！这条大鱼在某一天突然发生变异，变成了一只鸟，这鸟也大得离谱，翅膀上的羽毛铺天盖地，像躺在天边的巨云！这牛皮吹得，让人心里爽啊！最振奋人心的是这鸟飞起来的时候啊，双翅拍打水面，能激起三千里的浪涛（差不多半个日本南北端的直线距离，这不是浪涛，是海啸了），这鸟儿环绕着旋风可以只冲上九万里的高空（大约比超音速飞机和高速军用飞机能飞的最高高度还要告数倍）。。。李白诗歌里的“扶摇直上九万里”就是直接套用这里来的。试想一下，但凡头脑清醒的人，那个会这么想象，这么对别人讲话。但是，就是因为这种跨度极大，突破思想安全边际的想象力，才能造就顶峰的文学之美。而对于美，现代脑科学的研究也表明，这东西与计算“利害得失”的大脑区域不在同一区域，是处于完全不同的脑区，你在特别郁闷特别痛苦的时候，只要亲身看到壮美的山河,或其他感觉很美的景观，你的郁闷和痛苦，就会突然消失，就是这个原因。说到这里，我们回到今天的主题，或者我想对主角的接下来训练模型选材上的一些心愿：可以在训练的时候增加一些传统文化中频繁出现的各种花鸟鱼虫，像牡丹、杜鹃、梅花、兰花、萤火虫等等，如有可能，可以参考屈原的楚辞里出现的各种美人香草集，而在先秦时期常见的战车、弓箭、弩、矛、戈、枪、剑、盾、皮甲、铠甲等元素在有余力的时候也请加入，多训练一些男性服装，这个非常有价值的模型，太会画女人，而不太会画男人。
可以加作者的qq群：695960980
@chirenshuomeng 非常感谢
@Showevr 不客气
加载不出来模型 有没有大佬知道啥原因
size mismatch for model.diffusion_model.output_blocks.8.0.skip_connection.bias: copying a param with shape torch.Size([320]) from checkpoint, the shape in current model is torch.Size([640]).
我也是这个问题，问了GPT它说这是训练模型和当前模型的参数不匹配，要改成匹配的，但是我又不知道该怎么改
@a13126871219847 sd1.5的vae、lora、embeddings和SDXL都不能兼容，注意检查自己是否有混用的情况
要用支持sdxl1.0的幻世4.4
@fok3827 大佬您这个回答太专业了我看不太懂[捂脸]，我就是个只会照着大佬视频部署本地环境的菜鸟，模型列表里确实有两个vae前缀的模型，是要把它们从模型文件夹删除再加载荣华吗？
这个模型搭配openpose会造成明显的画面劣化，有人知道是什么原因吗
目前sdxl的模型搭配controlnet都会引起画质下降，但是如果用fooocus会好很多
@chirenshuomeng 用comfyui的听完默默地留下了眼泪
阿，倒可以一次绘制完，然后低强度重绘一次，细节好多了
@mesmer 我也是comfyui