leosamsFilmgirlUltra_velvia30Lora.safetensors

144 MB · SafeTensor

Requires membership

<filmgirl-ultra>

🌍欢迎加入QQ群"兔狲·AIGC梦工北厂"，群号 ：780132897 ；"兔狲·AIGC梦工南厂"，群号 ：835297318（入群答案：兔狲）。Telegram群聊“兔狲的SDXL百老汇”，链接： 🚨Recommended parameters for FilmGirl Ultra： Clip skip:1 CFG scale: 9 Direct output image resolution: ~500,000 pixels (640x768) 2024.2.29 Introducing "FilmGirl Ultra",Say goodbye to the AI face of SD1.5 On February 24th last year, I completed the first version of FilmGirl LoRA. This LoRA was my first model to achieve a high download volume and marks the beginning of my dream in AI. Since the launch of SDXL, I have devoted a great deal of effort to improving the and XL models. It has also been 8 months since the FilmGirl series was last updated. In fact, whether it's FilmGirl, or the subsequent or Helloworld XL, I have always been pursuing the ultimate in photorealism. Now a whole year has passed, and to commemorate the first anniversary, I have decided to release a model that elevates the photorealism of SD1.5 to new heights. The LoRA model is no longer sufficient for this mission; the new FilmGirl Ultra is an SD1.5 base model. To completely break away from the homogenization of SD1.5 photorealistic models and the issue of AI faces, FilmGirl Ultra didn't choose basilmix, chilloutmix, or their descendants as the training base model, but instead selected the newly released SPIN-Diffusion by UCLA. is a Self-Play Fine-Tune SD1.5 base model using the winner images of the pickapic_v2 dataset, which outperforms the SD1.5 original base model and SD1.5 DPO base model, and its prompt alignment performance is far superior to heavily fine-tuned and merged base models like Chilloutmix. The training set for FilmGirl Ultra comes from HelloWorld XL. In fact, the first version of HelloWorld XL also used the training set from the last version of FilmGirl LoRA. Throughout this year, I have been meticulously accumulating and selecting this training set, which now totals ~10,000 images. The training process for FilmGirl Ultra utilized multiple labeling methods, including GPT4V natural language captions, GPT4V tag-style captions, and Blip+Clip captions. To ensure that the model is compatible with the commonly used prompts "1girl", "best quality", and "masterpiece", these terms were also appropriately added to some images (but you can still accurately trigger the effect of a little girl with "little girl/child girl"). The reason for using multiple sets of labels is to maximize the likelihood of triggering the desired effect. As part of the FilmGirl tradition, the film style has been given special attention, and you can trigger this style with the prompt "film grain analog photography". This model underwent a total of 7 training phases, with different batch sizes, optimizers, learning rates, and training set ratios used in each phase to achieve the current effect. If anyone is interested in fine-tuning SPIN-Diffusion, I recommend that your total training iterations should exceed 50,000 steps; in fact, I trained for about 100,000 steps with batch sizes ranging from 40 to 64. The photorealistic effect of FilmGirl Ultra exceeded my expectations and is now close to the image quality of SDXL. Below is a comparison of this model with and , the former being the currently highest downloaded base model on Civitai, and the latter being the most photorealistic SD1.5 base model in my opinion for a long time. I pay tribute to these two excellent base models and their creators. close-up couple's portrait,African young woman and man,clear skin face,looking at camera,fashion photography,simple background Negative prompt: watermark,anime,cartoon,open mouth close-up couple's portrait,African little girl and boy,clear skin face,looking at camera,fashion photography,simple background, Negative prompt: watermark,anime,cartoon,open mouth, Thanks to GPT4V captions and the SPIN-Diffusion base model, the model's prompt alignment performance is excellent. Below are some xy plot tests for different concepts. Ethnic test Body shape test Skin color test Age test Animal test However, FilmGirl Ultra doesn't lead in all dimensions. After all, it started from a new point and gave up on the continuous optimization and refinement of the community's 1.5 base models over the past year. Through extensive testing and comparison, I found that this base model has a higher rate of limb errors than the community's mature realistic models. Also, due to a lack of anime-related content in the training set, the output is not good when your prompts involve related tags of ACGN. It is recommended to avoid using words like "digital art", "anime", "cartoon", etc. These two issues are the main current shortcomings of FilmGirl Ultra. FilmGirl Ultra is an annual summary of my first year on my AI journey, a gift to those AI enthusiasts who have supported me. The open-source community has brought me many friends, memories, joy, and knowledge. I also hope to contribute a bit back to the community. I welcome everyone to base your model training or merge it with FilmGirl Ultra. If you find this model helpful in improving your own model, please mention it in the model description. I hope that FilmGirl Ultra and SPIN-Diffusion will become more widely known and used. FilmGirl Ultra will continue to be updated, and I wish everyone happy usage! Hope we can continue to progress with AI, and meet here again this time next year! 去年的2月24日，我完成了第一版FilmGirl LoRA制作。这个LoRA是我的首个高下载量模型，是我的AI之梦的开始。自从SDXL推出后，我将大量精力投入到HelloWorld和AIArt两个XL大模型的改进中。FilmGirl这个系列也已经8个月没有更新了。 其实不管是FilmGirl，还是后来的拍立得LoRA、Helloworld XL，我一直都在追求极致的写实感。如今已整整一年过去，作为一周年纪念，我决定推出一个可以将SD1.5的写实感抬升至新高度的模型，LoRA模型已不足以承载这个使命，新的FilmGirl Ultra是一个SD1.5大模型。 为了彻底摆脱SD1.5写实感大模型的同质化和AI脸问题，FilmGirl Ultra没有选择basilmix、chilloutmix及其子子孙孙们作为训练底模，而是选择了UCLA最新发布的SPIN-Diffusion。SPIN-Diffusion是一个使用 pickapic_v2 数据集胜者图像进行自我对弈微调的SD1.5底模，其表现优于SD1.5原始底模以及DPO底模，同时提示词对齐性能远好于Chilloutmix等经过大量微调与融合的底模。 FilmGirl Ultra的训练集来自HelloWorld XL。实际上HelloWorld XL的第一版所使用的训练集也来自最后一版FilmGirl LoRA。这一年我都在精益求精地积累和筛选该训练集，如今整个训练集数量已达到1万张。FilmGirl U

SHA256

AC8B0E4AA77BE4D8B83DA9BAFE0134A2E36504C9B5263A7030394CFFE4F7003A

ModelTrends.ai Model ID

#2612

Sampler

DPM++ 2M Karras

Steps

25

Guidance

20

Seed

3863324487

Size

512x768

Clip skip

2

Sampler

DPM++ 2M alt Karras

Steps

30

Guidance

7

Seed

3695083888

Size

1024x1536

Clip skip

2

Sampler

DPM++ 2M Karras

Steps

27

Guidance

7

Seed

970000565

Size

512x832

Clip skip

2

Sampler

DPM++ 2M Karras

Steps

27

Guidance

7

Seed

1401714505

Size

512x832

Clip skip

2

Sampler

DPM++ 2M alt Karras

Steps

40

Guidance

15

Seed

902060405

Size

512x768

Clip skip

2

Sampler

DPM++ 2M Karras

Steps

27

Guidance

7

Seed

657490784

Size

512x832

Clip skip

2

Sampler

Heun

Steps

30

Guidance

6

Seed

3891906373

Size

512x768

Clip skip

2

Weight used

0.6