Сайт модели: https://haveall.net/
На сайте можно найти много разных дополнений, инструкций и тысячи готовых промптов.
Чтобы обновления моделей выходили чаще, меня можно поддержать на бусти: https://boosty.to/artxtop
Для связи я в ВК
CFG Scale = 2
SHOWCASE HAVEALL MODEL -> https://haveall.net/ is featured on the model's website. The site shows several hundred examples with images and complete prompts for generation.
Recommendations for generating images with the Haveall model:
CFG Scale: 2
For portraits and realistic images, use CFG Scale: 2. For graphics, you can use CFG Scale: 3
Standard image resolution for generation: 768х768, 832х576, 576х832
You can generate images in other formats: 896х576, 832х512
No Lora and embeddings are required, just simple prompts.
A ComfyUI workflow
Link -> https://civarchive.com/posts/1450922
Includes two parts:
- Simple Generation
- Upscale
Can be generated with the first option, or together with Upscale
Description
CFG Scale = 2
SHOWCASE HAVEALL MODEL -> https://haveall.net/ is featured on the model's website. The site shows several hundred examples with images and complete prompts for generation.
Recommendations for generating images with the Haveall model:
CFG Scale: 2
For portraits and realistic images, use CFG Scale: 2. For graphics, you can use CFG Scale: 3
Standard image resolution for generation: 768х768, 832х576, 576х832
You can generate images in other formats: 896х576, 832х512
No Lora and embeddings are required, just simple prompts.
So, what's new?
The first Haveall model was trained on the SD 1.5 DPO model. In March, UCLA-AGI introduced the SPIN-Diffusion model, which is superior to the SD 1.5 DPO model.
To train Haveall XV2, I used LEOSAM's FilmGirl Ultra, which was trained on the SPIN-Diffusion model. LEOSAM's FilmGirl Ultra was used as a base model on which I trained a dataset of 6k images to my specific settings. The result was the Haveall XV2 model.
FAQ
Comments (15)
Is the difference v1 vs v2 only in their base models?
Yeah, that's basically the difference.
I have done various tests and I can say for sure that SPIN-Diffusion is superior to SD 1.5 DPO in many ways.
V1 was trained on 5k images and V2 on 6k images. Also in some of the old captions I replaced the captions with more detailed ones.
Thus V2 follows the prompt better and understands natural language better.
Is there a chance to get v2 but trained over the previous base?
The thing is, there is still a difference in V1 and V2 training. Version V1 was trained with batch sizes=1. For training I rented a GPU with 24Gb. And it took 20 days to train. The company where I rented the graphics card asked me not to use their graphics cards for such a long period of time anymore :)
For V2 training, I rented an A100 and used batch sizes=8. This reduced the training time to 6 days.
To get the same result as in the V1 version with the new dataset, I need:
- Or 3 16-24gb video cards and then the training will take 7 days on three video cards
- Or one 16-24gb graphics card and then the training will take 20-22 days.
I don't have such a video card at the moment and spent all my money on renting a GPU to traine past models :D
@Haveall noted, thank you!
@Olbanets I made an averaged version, model V1.5, dropped it on huggingface https://huggingface.co/sonangroup/HaveallX/resolve/main/HaveallV1.5.safetensors?download=true
@Haveall I'd love to know some technical details about it! What's the difference?
@Olbanets it's a merge. Through add difference. I tried to combine V1 and V2 to get an average version. As a result, the generation is half of V1 and half of V2.
@Haveall have you tried "train difference" mode in supermerger? it's amazing!
@Olbanets no, I haven't tried it yet. Thanks for the tip!
@Haveall sure, it's so useful! https://github.com/hako-mikan/sd-webui-supermerger/blob/main/calcmode_en.md#train
Agreed, the train difference mode is essential for pretty much anyone who does merges "seriously"
I won't be using other models any time soon, fantastic capabilities.
Dude WTF ! This was the kind of results i was searching from morning. Awesome & unique model man. Great work, respect !
This model is really just a real treasure among thousands of SD 1.5 models, that's exactly what the original SD 1.5 was supposed to be. Just incredible flexibility and versatility, excellent responsiveness, good knowledge of styles and original and very expressive look of the pictures. This model stands out from the rest very much, a real diamond! Amazing job! Thanks!
Дружище, твоя модель действительно великолепна именно своей самобытностью и оригинальностью. Очень круто получилось. Респект и уважуха!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.