,__mature female, dark-skinned female, solo, — копия.jpeg)
Hello my friends.
As you may have noticed, I started training LoRAs for the Anima model. I tested the model on Preview 2 and Preview 3 for about 1.5 months, and now, after the 1.0 release, I can summarize my test results.
I. About model:
First of all - you can read about this model on HF page or civitai page of model.
This is not another SDXL finetune like Pony, Illustrious, or NoobAI. It is a new architecture.
It is also not as heavy as Flux or Qwen, since it has about 2B parameters. But for NSFW / SFW anime, cartoon, and 3D images, it really feels like next gen.
There are many pros, and this model can become for many people what SDXL was for the last two years:
The new 16-channel VAE, compared to only 4 channels in SDXL, the smart text encoder with a 0.6B LLM adapter, compared to SDXL’s 0.3B CLIP-L without an LLM, and flow matching prediction, compared to epsilon prediction in Pony, Illustrious, and NoobAI Eps, or V-prediction with rescale CFG in NoobAI V-pred, all do a very good job.
It is trained on Danbooru / Gelbooru dataset + DeviantArt + LAION-POP, which is mostly SFW “normie model” content, but still high-quality data. So you can use the old booru tag system together with natural language descriptions, because the LLM understands you most of the time.
Details are much better on Anima compared to any SDXL finetune. In txt2img, you can get images that usually required txt2img + highres fix + ADetailer on SDXL.
It knows most artists and characters from Danbooru / Gelbooru up to September 2025. NoobAI V-pred knew content only up to November 2024. As the author said, Gelbooru tags should have priority.
Sizes of checkpoint and loras are smallers (Lora with 64 dim has 200~ mb instead of 400~ on SDXL). No need for quantizing for using in comparison with Flux, Qwen because VRAM comsuption for image gen like SDXL has.
There are also cons:
Sadly, it is not trained on e621, unlike NoobAI Epsilon and NoobAI V-pred. So furry and monster enjoyers, there is work for you to do. Anyway, you can still make furries and monsters on Anima, because booru datasets contain them too, but the datasets are much smaller.
LoRA training is slower, about 1.5–2x, and image generation is also slower, about 2–3x. But it is worth it. Also, you don’t need more VRAM.
Some people say that it is harder to mix checkpoint styles, but I did not have such problems when mixing artists in my tests. Also, I usually prefer style LoRAs anyway.
II. Settings for image generation:
I tested the model primarily with my 3D LoRAs, but also on anime images with the checkpoint’s artist tags and some of my cartoonish LoRAs. I don’t claim this is the absolute truth, but I’ll write down what I use.
First of all - I prefer forge neo webui. The author frequently updates his webui and fixes bugs that you have found and reported in issues.
You can use everything in ComfyUI too but if you use Comfy I think you know what to do :)
Samplers:
1) ER SDE — best universal sampler. It feels like Euler a, but more accurate. Also, it does not kill backgrounds.
2) Euler a — still good for anime, 2.5D, and cartoon images, but you usually don’t want to use it in the highres fix slot. It harms 3D backgrounds too much, especially realistic backgrounds. I don’t recommend it for 3D.
3) Euler — sadly, it does not work well in txt2img. Use it in the highres fix slot if you don’t like ER SDE polishing during the second pass. Can also kill details of 3d and realistic textures like euler a but not so much. Sadly it is nerfed in comparison with his role in SDXL.
4) Any DPM — never. It became garbage after SD1.5, and it is still garbage here. Karras also does not work on Anima.
Schedulers:
1) SGM Uniform — works well, just like it worked on SDXL. You can use it.
2) FlowMatchEulerDiscrete — basically SGM Uniform, but for flow matching models. It uses the model’s shift parameter. Anima has shift 3 by default, but it can be changed. I prefer this scheduler, but the difference is not critical. Mostly useful for txt2img.
3) Beta — another good scheduler. It was controversial on SDXL, and it can still make mistakes during generation, but Beta can be creative. It feels like something between SGM Uniform and KL Optimal. I mostly use it in the highres fix slot, but it can also work for txt2img.
4) KL Optimal — absolute garbage on SDXL, but king of details on Anima. I prefer not to use it in the txt2img slot, but sometimes I use it in the highres fix slot because it works like a free detailer or restart sampler and does not really change the composition. You don’t need it for anime and cartoon images most of the time, but it works very well with some 3D textures and makes them much more detailed. If KL Optimal is too aggressive and creates too much mess, I switch to Beta.
Resolution, CFG, scaling, steps and denoising:
For image gen I use 896x1152, 832x1216, 1152x896, 1024x1024 resolutions.
CFG 3-5. For 3D usually I use 3.5 CFG for gen and 3 for highres fix. CFG 4-5 for anime.
Usually, I use 30–40 steps. The longer your prompt and the more detailed your style, the more steps you need. Sometimes even 50 steps can be useful.
For highres fix, I use 15–20 steps.
Denoising for highres fix: 0.25–0.3 for anime and cartoon images. The effect is not huge because your txt2img image is usually already very detailed by default. For 3D, I usually use 0.5 denoising with 20 highres fix steps if I use Beta or KL Optimal, because these schedulers are detail champions without too much mess.
For upscaling I use foolhardy remacri upscaler and 4xPurePhoto-RealPLSKR. Also HAT-S is good option if you can wait 8-10 more seconds. Usually for 3D pics. DAT 2 2x is also good but it very small - I didn't saw significant difference in quality with HAT-S 2x.
I prefer scaling images with a multiplier of 1.25x - 1.35x, because the further you get from native, the more mutations you'll get. Considering that the images look very good in native, there's no need to use 1.75x - 2x. If you really want high resolution, use extras with a high-quality upscaler (like DAT) or the SD upscale script after second pass.
ADetailer:
Custom resolution: 1024x1024.
Denoising: 0.3.
For eyes: use an individual eye model and “merge” mode in mask settings.
Also, I use pixel padding around 222–256 pixels instead of the recommended 32 pixels.
About NoobAI and Illustrious
I will continue to make "legacy" versions but not sure what I choose - NoobAI V-pred or NAI Epsilon 0.75 (for NAI Epsilon and Illustrious 0.1). Or you need both? Write in comms.
But no more SD1.5 and Pony loras. Sorry but they are too old now.