anima UWU Base
This stylistic base model is AIO, as it uses a modified text encoder that doesn't make sense to release separately. It's less flexible than the base Anima model in terms of background and environment variety; it tends toward simple, white backgrounds and environments, but that doesn't make it weak.
The model requires some clarification when generating a character. It is not enough to simply write the character's name and where he is from. I would advise making clarifications, for example, writing his hair color and eye color, perhaps characteristic features, so that the model understands what is required of it.
The model uses an expanded, previously tested dataset and many of my developments related to the creation of dynamic and living art.
Workflow for AIO models Anima + lora + upscaler + resizer - YUNI UWU AIO | Anima Workflows | Civitai
Settings
CFG 4-6
Sampler res2s\res_2m\res_3m\ euler\ dpmpp_3m (I tested the model only on them)
sheduler simple\karras (I tested the model only on them)
steps to run -1
steps 10-40 (sometimes you can lower it to 8-9 steps)
Yes, the model works great at small steps without being a turbo or distillate (those who have used my fine-tuning of ilustrios models know what I mean)
768x1280 resolution and inversion (I just didn't try others in my tests)
Description
Background handling has been improved; the model produces images with higher contrast and sharpness, more details, although—depending on the complexity of the prompt—it may generate low-quality images somewhat more frequently than the standard version.
This version could be called 1.5, but that is not entirely accurate; rather, it is a preliminary version of the second iteration, built upon the base model's dataset. It is an experimental model created to explore ANIMA architectures and their training capabilities.
FAQ
Comments (6)
Please remove the videos. They have nothing to do with Anima. You made that with Wan or whatever, it's irrelevant. So frustrating
1) The video cover is directly linked to the model card; it represents how the creator showcases the model.
2) For many people, image-based models serve as raw material for video models; since the focus is primarily on animation, these image models are used to generate the frames or shots for scenes—and this is a perfect example of animating an image produced by such a model.
3) WAN 2—just like WAN 3—could never produce an animation like this; they are severely outdated models.
@Yunmiyun_UwU We all know what image-to-video is. We know how first frame and last frame works. It doesn't excuse an irrelevant workflow with an unknown video workflow or model. That's what MiniMax or Wan or LTX can do. This is posturing and obnoxious.
@agnesthecakeslayer stop bullying creators, you are the only one calling out video preview of models, everyone is fine with it
@Aemeath1 And yet it's literally inaccurate and misleading. Provable by math. Stop defending nonsense behavior for upvotes. Have some empathy for people trying to learn how all this stuff works. Literally a file and workflow that doesn't produce videos!
@agnesthecakeslayer Almost every top creator produces video covers or previews—and that’s not counting hundreds of others—for image models, LoRAs, and workflows alike. To those who—due to a lack of knowledge—cannot identify the model in front of them, fail to review the examples, architecture, and descriptions, and consequently mislead themselves: I advise you to use your own intellect first. Only after you’ve gained knowledge and grasped the basics should you turn to artificial intelligence.
If you are looking for a video model, I recommend installing MiniMax H3. The INT8 version offers the best quality on relatively accessible hardware; if you have at least an RTX 3090, 48GB of RAM (ideally 64GB or more), and about 45GB of SSD space, you’ll get perfectly acceptable speed and quality, even with the default workflow. The key is to avoid using any "turbo" LoRAs. Also, for full INT8 functionality, you need 20GB+ of VRAM on an Ampere-architecture card (or newer). If your hardware is weaker—for example, only 16GB of VRAM (like a 5070 Ti) and 32GB of system RAM—you should install a less precise model with lower quality, such as INT4, GGUF Q5, or NVFP4. If you have less than 16GB of VRAM and less than 32GB of system RAM, I wouldn't recommend attempting video generation at all; it will be extremely slow, and the quality will be poor.