■This is an experimental ground for Anima LoRAs.
The LoRAs I create serve more of a role in stylistic tuning rather than reinforcing single concepts.There is no inherent superiority among the LoRAs; they simply serve different purposes. You can find the description for each LoRA in its respective tab.
■If you notice pixel patterns in your generated images, try the VAE below. It removes the pixelation typical of the qwen_vae and makes the images much sharper.
https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image
■If the VAE looks a bit too saturated for your taste, try using this VAE merge node to blend it with the original qwen_vae to dial it in. A 0.5 merge should be just right.
https://civarchive.com/models/2487550/vae-merger-for-comfyui-anima
■I would like to share the possibilities of anima with everyone.
My wish is for many people to discover basemodels with potential and to see their possibilities unfold even further. I would be happy if I can help make that happen.
■The anima architecture has been heavily modified and retrained from scratch—almost like a radical overhaul—but the base architecture is Cosmos-Predict2-2B. To put it simply, you could compare it to a lightweight version of Flux 1.
I think it’s a great architecture capable of generating high-quality images despite its small size.
I have a very good impression of Anima when it comes to color vibrancy, lighting, and prompt adherence.
Its knowledge regarding NSFW concepts, characters, and styles is at least on par with SDXL. The overall feel of using it is also quite similar to SDXL.
Furthermore, it is lightweight and easy to train. I believe large-scale training is also a viable option.
Since we are provided with a base model that already contains a vast amount of knowledge, we rarely need to teach the model completely unknown concepts from scratch.
Basically, our work mostly comes down to adjusting styles or reinforcing minor concepts and characters. You can achieve great results without having to do any heavy lifting.
The developer has also ensured transparency by sharing the training settings, which means you are less likely to stumble over black-box issues.
Because of this, the burden of training is low, making it an option well worth nurturing together as a community.
■The LoRA has a stronger influence when used with shorter prompts.
Since longer prompts often carry their own inherent style, the LoRA will primarily take on a supporting role to refine and unify the overall image.
■https://github.com/gazingstars123/Anima-Standalone-Trainer
This is a great tool that allows for easy training, even on Windows. It has everything you need for LoRA training.
Even if you run into any issues, you should be able to get your LoRA training up and running by troubleshooting with ChatGPT or Gemini.
If you find the tool useful, you might want to consider making a donation to him. It allows the developer to focus more on development and deliver even better tools, which ultimately benefits you in the long run.
■If you have any questions, please feel free to ask!
日本語での質問も大丈夫ですのでご気軽にお声がけください～
If you enjoy my work and feel like making a donation, I would be incredibly grateful. Even the smallest token of support goes a long way in helping me lighten the load.
Whether I receive donations or not, I will keep doing what I love and pursuing my interests just as I always have. However, simply knowing that this kind of support exists brings me immense joy and serves as a huge driving force for me.
Description
This is a style LoRA focused on the aesthetics of cool, retro anime.
It was trained on a dataset of 16,000 carefully selected images from what can be considered masterpieces of traditional cel animation. You can find the list of the included works below.
■Recommended Settings:
A lower LoRA weight, around 0.5 to 0.8, is generally recommended and easier to work with. This allows you to achieve more stable generations while still enjoying the excellent color palette and overall atmosphere.
■Style Characteristics:
The overall style leans heavily into the defining elements of cel animation. It features stunning analog-style background art, a calm yet rich color sense that feels truly artistic, and classic cel animation effects like bloom and lens flares. Instead of overly vibrant, digital-looking colors, it tends toward deep, earthy tones that give a more profound and nostalgic impression.
■Versatility:
Because the dataset includes a wide range of genres—such as Sci-Fi, traditional Japanese style, and fantasy—this LoRA is highly flexible and can adapt to almost any genre of image you want to generate.
■Face Styles & Adjustments:
Character face styles are not strictly fixed, so you may notice the vibe shifting between the various works included in the dataset. Generally, however, they lean slightly towards a realistic touch, often giving the impression of young women with eyes that aren't overly large. Since the dataset contains just as many male images as female ones, male characters will follow the same stylistic tendencies when generated.
If you prefer to avoid overly realistic faces, or if you don't want the face style fluctuating with every generation, simply lowering the LoRA weight to allow the base model to have a stronger influence will usually strike the perfect balance.
Alternatively, it works even better when combined with your ideal style LoRAs.
Think of my LoRA as the base seasoning. The dish is finally completed by adding your own spices.
Since highly specialized style LoRAs often rely on smaller datasets and lack diversity, my LoRA helps compensate for that weakness.
■About the Files & Backups
To stabilize the style, this main LoRA is a merge of several different training steps.
I am also sharing the unmerged backup LoRAs that were used as ingredients for this merge.
■List of Works: (The trigger tag does not exist.)
Cowboy_Bebop_movie&TV
ghost_in_the_shell
Rurouni_Kenshin_OVA
Memories
patlabor_movie_1.2.3
Vampire_Hunter_D
The_Vision_of_Escaflowne_movie
X (CLAMP 1996) movie
GUNDAM_The 08th MS Team_OVA
GUNDAM_0083: Stardust Memory_OVA
GUNDAM_W Endless Waltz_movie
■A Personal Note on the Anime
I truly believe the titles in this list represent the absolute pinnacle of traditional cel animation. The artwork in every single one is on par with fine art, and seeing it brought to life with such stunning animation is a truly rich visual experience. They are all fantastic masterpieces absolutely worth your time, so I highly recommend checking them out if you ever get the chance.
FAQ
Comments (11)
were to find Standalone_TIPO
StandaloneWildcardProcessor EnhancedSaveImage?
I actually made these nodes to reduce my own workload for auto-generating prompts. You don't really need them if you normally just type your prompts manually.
I was planning to upload a ComfyUI portable version with these already integrated at some point, but I haven't gotten around to it yet. If you're still interested and want to give them a try, I've uploaded the custom nodes to the link below.
https://huggingface.co/hjhfgfxj/comfyui_node_lab/blob/main/comfyui_node_v001.zip
Just a heads-up, standalone_tipo is I'm not entirely confident it will work flawlessly on other people's setups, so maybe only try it if you really want to. I believe it should work if you activate your ComfyUI virtual environment (venv) and install the following required libraries:
"pip install tipo-kgen transformers llama-cpp-python"
Custom Node Descriptions:
enhanced_save_image.py: A save node that properly records metadata (like prompts) for uploading to Civitai. I created this because metadata wasn't being saved correctly when using Wildcards.
standalone_wildcard.py: This just extracts the ImpactWildcardProcessor from the ComfyUI Impact Pack. I made it because installing the entire Impact Pack just to use Wildcards felt like a waste.
tipo_v02.py: The original Tipo extension is quite old and hasn't been working properly lately, so I improved it.
None of these are completely new nodes built from scratch; they are essentially just modifications of existing nodes tweaked to fit my workflow better. If you're already using something similar, there's no need to switch.
Let me know if you have any questions or run into any issues after installing them!
this is what i am getting [WARNING] Traceback (most recent call last):
File "C:\A.I\Packages\ComfyUI\nodes.py", line 2246, in load_custom_node
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 936, in exec_module
File "<frozen importlib._bootstrap_external>", line 1073, in get_code
File "<frozen importlib._bootstrap_external>", line 1130, in get_data
FileNotFoundError: [Errno 2] No such file or directory: 'C:\\A.I\\Packages\\ComfyUI\\custom_nodes\\ComfyUI-Standalone_TIPO\\__init__.py'
but it just stems from that i can replicate the quality of the pics you make lol ,but i try.
@heatwoodzachary763
Placing tipo_v02.py directly inside the custom_nodes folder might fix the issue. Please do the same for the other nodes—just place them directly in that folder.
It seems like you might have created a subfolder and put the Python files in there, which is likely causing the error.
@heatwoodzachary763
Also, if you want to get closer to my generation settings, I can share a few tips.
I have two main sets of settings: "v01," which I used when generating the sample images, and "v02," which is a recently revised version.
One thing to note is that I always use a Turbo LoRA for inference now. I didn't use one before, but generating a massive amount of sample images for LoRAs takes quite a toll on a slower GPU, so I started using it.
Both v01 and v02 use the Turbo LoRA linked below.
■anima turbo lora v0.2
https://civitai.red/models/2560840/anima-turbo-lora
■rdbt_anima_b1_v0.39.b
https://civitai.red/models/2364703/rdbt-or-anima-lora?modelVersionId=3017013
■rdbt_v1.0_anima_b1_8-step
https://civitai.red/models/2364703/rdbt-or-anima-lora?modelVersionId=3101358
Here are my inference settings and some tips, just in case they are helpful for your own generations.
■v01 Inference Settings
workflow image
Merged anima turbo lora v0.2 and rdbt_anima_b1_v0.39.b at a 0.5 ratio.
Steps: 12
CFG: 1
Sampler: Gradient Estimation
Notes on v01:
Gradient Estimation is essentially a faster-converging version of Euler, and it produces quite sharp results. I sometimes use ResMultistep among the Euler-based samplers as well. The difference is subtle, but I adopted Gradient Estimation here because it converges faster and yields sharper images.
I think that setting was actually quite good for achieving that flat, 2D anime-style look. However, it felt a bit tricky to use since the results were pretty hit-or-miss.
Also, don't worry too much about the fact that I specifically merged the LoRAs. I just do it because it feels more intuitive for my workflow. Simply applying both LoRAs sequentially at a 0.5 weight each would likely produce very similar results.
■v02 Inference Settings (My current setup)
workflow image
rdbt_v1.0_anima_b1_8-step only.
(Alternatively, you can merge anima turbo lora v0.2 and rdbt_v1.0_anima_b1_8-step at a 0.5-0.7 ratio.)
Steps: 12
CFG: 1
Sampler: Euler
Notes on v02:
This setup solves some of the issues I was struggling with in v01. My main problems were that the contrast was too high (making it hard to bring out details) and the generations were prone to breaking (visual artifacts).
The excessive contrast was caused by the combination of Gradient Estimation and anima turbo lora v0.2. Because Gradient Estimation converges quickly, it naturally produces stronger contrast. I switched back to the standard Euler sampler, which gave me a much better balance of contrast and fine details. But ultimately, this comes down to personal preference—feel free to use whichever sampler you like!
I also decided to stick with just rdbt_v1.0_anima_b1_8-step for the turbo LoRA. It's very stable and pairs well with style LoRAs.
anima turbo lora v0.2 has a very strong anime style, but it tends to lose background details easily. (In addition, some breakdown occurs.)On the other hand, rdbt_anima_b1_v0.39.b has a great atmosphere, but it breaks frequently, especially when I apply my own LoRAs on top of it.
By the way, Anima recently released a new official turbo model, but I'm not a big fan of it. The anime style itself is great, but it feels too flat and lacks background details for my taste.
Since rdbt_v1.0_anima_b1_8-step has a bit of a 2.5D vibe, if you don't like that look, you can mix in a little bit of anima turbo lora v0.2 to balance it out. The texture reproduction is high, so depending on the situation, it can actually add more details.
Honestly, though, when applying a style LoRA, I personally prefer starting with a slightly realistic base because it retains better lighting and finer details. That's why I currently just use rdbt_v1.0_anima_b1_8-step on its own.
I hope these inference notes are helpful to you!
As for the scheduler, I primarily use simple or sgm_uniform.
simple is basically stable across almost any model.
sgm_uniform is almost identical to simple, but I personally feel that it’s just a bit more stable.
i got it set up whats on the photo and the loras...so much thanks.
@heatwoodzachary763 you're welcome!
oh i got your auto-generating prompts work flow working at last, and i made a work flow base on what you showed me i hope to make some knock outs in time.
@heatwoodzachary763
I'm glad to hear it worked out!
Regarding "tipo," there is a lot of useful information on the official GitHub repository, so it might be worth checking out:
https://github.com/KohakuBlueleaf/z-tipo-extension
When adding meta tags for character names, series titles, art styles, or anime screenshots, you might need to use specific formatting like this:
<|characters|>, <|copyrights|>, <|meta|> etc.
Also, if a character name includes parentheses, like (fate), the system might not recognize it properly. Escaping it like \(fate\) will usually solve this issue.
As for my Retro LoRA, I highly recommend merging it with other LoRAs. Personally, I feel that my Retro LoRA offers a lot of diversity and generates wonderfully artistic backgrounds, but I'm not a huge fan of its specific character style. I consider this LoRA to act more like a base model—providing a "base seasoning" or foundation.
if a dataset is too large and contains too many diverse types of images, it actually becomes harder to target and consistently generate images with that specific "good vibe" you're aiming for.
Because of this, you'll likely get much better results if you merge it at a 0.5 ratio with a more specialized style LoRA, like "Vampire Hunter D" or "GM Thunderbolt". Combining it with other creators' amazing style LoRAs is also a great idea.
Typically, LoRAs that consistently produce good results are trained on smaller datasets, which often limits their versatility.
By combining them with my Retro LoRA, they can complement each other and bring out the best of both worlds.
https://civitai.red/models/2718533/hdr-vae-anima-qwen-image
Also, using this VAE might solve the detail issues in anima or krea2. I used to apply sharpness in my own workflow, but when using this VAE, you probably won't need to set it to such a strong value.
Either way, thank you so much for trying out so different things!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.