In version 2.0 I added a node to get image metadata and another node for skimmed cfg to improve image quality even with increased cfg. If the images look strange, it's because I'm using my wildcards that have Illustrious loras, haha.
En la version 2.0 añadí un nodo para que guarde la metadata de las imagenes y otro nodo de skimmed cfg para que salgan mejor aunque se suba el cfg. Si se ven raras las imagenes es porque estoy usando mis wildcards que tienen loras de Illustrious juaaa
Hi!
The user @artificialotaku asked me about Anima, and I got curious, so I asked Grok. He told me it's a model made by Comfy that understands whole sentences instead of just prompts, so I decided to try it. I liked it, and it was easier to use than I thought. So I've created a very simple workflow for making images with Anima.
The necessary models are these:
This one goes in models/text_encoders. Make a folder called Qwen3-0.6B-Base and download all the files there.
https://huggingface.co/Qwen/Qwen3-0.6B-Base/tree/main
I'm using this model because I have no idea what I'm doing, and this one says it's the official one. I'll look around to see if there are any really cool ones for making anime, although this one does a decent job considering I asked it to make two different characters together without any Loras or anything. This goes in models/diffusion_models
https://civarchive.com/models/2458426/anima-official
This is the VAE file; it goes in models/vae: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
I recommend skipping the detailer if you want to do some testing; it takes a lot of time.
¡Hola!
El usuario @artificialotaku me preguntó por Anima, y me picó la curiosidad y le pregunté a Grok, como me dijo que es un modelo que hicieron los de Comfy que entiende frases enteras en lugar de solo prompts, decidí probarlo, me ha gustado y era más fácil de usar de lo que pensé. Así que he hecho un workflow muy simple para hacer imagenes con Anima.
Los modelos necesarios son estos: Este va en los models/text_encoders, haz una carpeta llamada Qwen3-0.6B-Base y bájate todos los archivos ahí.
https://huggingface.co/Qwen/Qwen3-0.6B-Base/tree/main
Uso este modelo porque no tengo ni idea y este pone que es el oficial, ya miraré por ahí si hay alguno superguay para hacer anime, aunque este lo hace bien para haberle pedido que me hiciera dos personajes diferentes juntos sin Loras ni nada. Este se coloca en models/difussion_models
https://civarchive.com/models/2458426/anima-official
Este es el vae, se coloca en models/vae: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
Aconsejo "omitir" el detailer si queréis hacer pruebas, se tira mucho rato con eso.
Description
FAQ
Comments (11)
Look who arrived at Anima! Welcome! And yeah, maybe I forgot to mention that Anima knows A LOT of characters by default, so I'm glad you managed to get your Albert in there. Thanks for the WF.. about Anima checkpoints, check out the reviews on https://civitai.com/models/2544636/wai-anima ... seems like a favorite!
Yes! Plus, it's easy to make Loras for Anima thanks to the workflow they recently published! https://civitai.red/models/2502969/anima-lora-trainer-for-comfyui I'll have to look to see if there's another workflow like this for Illustrious!
@saehara151 woah, cool! Wait, so, we can train loras straight inside comfyui now?
@ArtificialOtaku yeah! It's easy, although it takes some time, hehe
@saehara151 amazing, would you say it's faster than training with Ai Toolkit? I haven't made any Loras yet, so, i wonder what the best choice would be
@ArtificialOtaku I've only ever done LoRas using Civitai before, so I don't know how long AIToolkit takes. I have an RTX 5090, and training a LoRa with 22 images took me an hour and a half. For a LoRa with 26 images, I told it to only use 16GB of VRAM, and it took two hours.
@saehara151 nice, I have a 5090 as well... gonna have to check that anime lora trainer then... did you follow any tutorials for that?
@ArtificialOtaku No, I just used my Civitai dataset, which already had the images and captions done.
When I got an error, I asked Grok what to do, and that's it, haha.
You just have to create a folder called train_data_dir and inside it put another folder with the name of the Lora preceded by the number of times you want it to train. I used this: 10_eto_oc_with_wesker
It's also better not to set the VRAM to the maximum your RTX has because mine kept crashing and wouldn't do anything, so I set it to 16GB of VRAM.
@saehara151 got it, thanks for the info--- this actually might be easier on AI Toolkit, I'll definitely look into it.
Thank you so much for the amazing work! The anime images are incredibly creative and feature minimal adjustments to achieve 80% realism.
Ah, thank you so much, I'm glad it was helpful!