This is a realistic model based on the Anima model by CircleStone Labs.
It's fine tuned and merged with photorealistic Lora's that i've trained by myself.
No Lora's from other creators were used, only my original ones.
It works best with CFG between 4 to 8 (8 will result in a more "plasticky" look).
I recommend as Sampler res_2s and the Scheduler beta57 for the best results from RES4LYF, but the other usual Samplers should also do their job with 30 Steps.
Consider this model as NSFW, because the original capabilities of Anima are still intact. ;)
If you use my checkpoints for your merges, please give me a credit because working on these checkpoints is really time consuming and a lot of work to do.
Furthermore, please observe the licensing requirements of Circlestone-Labs and Nvidia.
Thank you!
Description
This is a realistic model based on the Anima base V 1.0 model.
It's fine tuned and merged with photorealistic Lora's that i've trained by myself.
No Lora's from other creators were used, only my original ones.
It works best with CFG between 4 to 8 (8 will result in a more "plasticky" look).
I recommend as Sampler res_2s and the Scheduler beta57 for the best results from RES4LYF, but the other usual Samplers should also do their job with 30 Steps.
Diffusion Model only.
Consider this model as NSFW, because the original capabilities of Anima are still intact.
FAQ
Comments (23)
Is it ok if i upload your model in another website?
https://tensorhub.art/models/1003366833060870360
Sure, no problem.
For real, the best realistic Anima!
Thank you.
Holy moly! work very well at 512x512
IMO - best realistic Anima model!
Thx!
@Spaceman72 indeed - the best details depth. Thank you.
wow, this is pony levels of good.
Thank you!
Insane quality. With easy touch of anima realism enhancer and Light Concepts at 2048x2048 resolution it just a checkmate to FLUX and Z-Image.
Thank you.
whats a "touch" for you? like 0.3 ?
@kunde2 more or less. Depends on situation. Actually i utilize different "realism" LoRAs. Not just those one. Sometimes i just can't "trigger realism" with all recommend tags. Especially for v2. So i go easy way and use little help from LoRAs. Like OMG Real on 0.3. This LoRA push hard almost any Anima checkpoint towards to realism. No matter tags used.
@kunde2 anyway, now i am working on creation of own LoRA. I plan to learn it on your model. So i gladly accept any advice from you on that way. I would like to chat with you about that subject, if you don't mind.
@PurePrompt Hey, I'd love to help you out, but i know nothing about creating models. Maybe you wanted to tag @Spaceman72 instead? OMG Real works very well, thank you for the tip!
@kunde2 indeed - about lora training, i wanted to ask Spaceman about...
@PurePrompt I've trained most of the loras with natural promptings instead of tags. But you can also mix your training prompts. I'am using the Anima-Standalone-Trainer (GitHub - gazingstars123/Anima-Standalone-Trainer: Standalone Anima Lora trainer with GUI · GitHub)
@Spaceman72 i'm on Intel Arc B580... no CUDA build. I will try default ComfyUI workflow, with native Triton support, when prepare dataset. Natural prompting, you say? Got it. Florence2 is on the way... or i should utilize somehow Anima Qwen 0.6B?
@PurePrompt No Florence2 is okay. I'm using it, too. But if you want to train NSFW content you should consider to edit the prompts manually afterwards.
@Spaceman72 there is nsfw version of florence, that generate additional nsfw description which can be concatenate to base caption.
@PurePrompt Thanks for the hint. :)
@Spaceman72 anyway - the best captioner is a Llama 3.1. Image matching of prompt is around 95-100%. Almost identical in "Training Prompt" mode. PixelaiLab's "Simple LLM Caption" - is the best for qwen.
https://huggingface.co/John6666/Llama-3.1-8B-Lexi-Uncensored-V2-nf4
https://huggingface.co/aryashah00/survey-finetuned-Llama-3.1-8B-Lexi-Uncensored-V2-nf4
@PurePrompt Thanks, i'll try them. :)