Model Description
Mugen is a continuation of our SDXL to Flux 2 VAE conversion, renamed to signify a substantial divergence from the original NoobAI models.
It has been trained for 7 additional epochs, totaling under 8000$ for a full latent space conversion, while preserving and improving upon model anime knowledge.
In particular, we have paid attention to characters in this iteration, and developed in-house approach for benchmarking their performance, about which you can read below.
Overall, model performs particularly well with textures and patterns that were previously simply impossible due to SDXL VAE. We prioritized keeping our training as standard-friendly as possible, so local community can easily train on it like on a new Base Model, which it practically is.
We provide 4 models:
Mugen: A base model.
Mugen - Aesthetic: Slightly tuned on a limited dataset model for better quality output.
Mugen - Aesthetic - Anzhc/Selph: Further tune on opinionated dataset selection.
Developed by: Cabal Research (Bluvoll, Anzhc)
Funded by: Community
License: fair-ai-public-license-1.0-sd
Resumed from: NoobAI Flux2 VAE v0.3
Character Knowledge Benchmark
This benchmark measures character similarity across 1815 characters in this iteration of it. For convenience, we've gathered few major categories: gachas and vtubers.
We utilize reference(non-generated) set of images, and measure character features against ai-generated data - this is the similarity score. Our custom in-house model for character discrimination trained on ~1.2kk images is used. Results are sorted indiscriminately, by score, treat it as general character knowledge index, not as any specific characters in particular. Same point on graphs might, or might not be corresponding to the same character.
Due to compute constraint, we selected only single model to compare against - not yet released latest version of Chenkin model, which currently is the most trained SDXL-based anime model.
Future benchmark iterations might include different arches, more models and more characters.
Bias and Limitations
General data biases from Danbooru might apply.
Flux 2 VAE seem to have brown bias overall, which can be alleviated by adding sepia or brown theme to negative.
Model Output Examples
You can download most of those images from Here for reference.
Recommendations
Characters
While in benchmark we test characters purely with their own trigger with no helper tags, it is advised to utilize series/game for better adherence. Characters that might appear not working initially could start working with appearance tags.
Inference
Comfy
We will provide a Node, and hope it will be adapted natively in main repo eventually:
https://github.com/Anzhc/SDXL-Flux2VAE-ComfyUI-Node
Just install it, and it will patch the model config, no node changes required.
SwarmUI also requires only the node to be installed.
Same as your normal inference, but with addition of SD3 sampling node, as this model is Flow-based.
Recommended Parameters:
Sampler: Euler A, Euler, DPM++ SDE, etc.
Steps: 20-28
CFG: 4-7
Shift: 8-12
Schedule: Normal/Simple/SGM Uniform
Positive Quality Tags: masterpiece, best quality
Negative Tags: worst quality, normal quality, bad anatomy, sepia
Alternative Extended Negative: (worst quality:1.1), normal quality, (bad anatomy:1.1), (blurry:1.1), watermark, sepia, (adversarial noise:1.1), jpeg artifacts
(Some of our testers pointed out that they prefer longer negative)
A1111 WebUI
Recommended WebUI: ReForge - has native support for Flow models, and we've PR'd our native support for Flux2vae-based SDXL modification.
How to use in ReForge:
Support for RF in ReForge is being implemented through a built-in extension:
IMPORTANT
Set your preview method to this, if you use it.
Flux2VAE does not currently have an appropriate high quality preview method, please use Approx Cheap option, which would allow you to see simple PCA projection(ReForge).
Recommended Parameters:
Sampler: Euler A Comfy RF, Euler A2, Euler, DPM++ SDE Comfy, etc. ALL VARIANTS MUST BE RF OR COMFY, IF AVAILABLE. In ComfyUI routing is automatic, but not in the case of WebUI.
Steps: 20-28
CFG: 4-7(or 7-15, if it appears to be weak/bugged)
Shift: 8-12
Schedule: Normal/Simple/SGM Uniform
Positive Quality Tags: masterpiece, best quality
Negative Tags: worst quality, normal quality, bad anatomy, sepia
Alternative Extended Negative: (worst quality:1.1), normal quality, (bad anatomy:1.1), (blurry:1.1), watermark, sepia, (adversarial noise:1.1), jpeg artifacts
(Some of our testers pointed out that they prefer longer negative)
ADETAILER FIX FOR RF: By default, Adetailer discards Advanced Model Sampling extension, which breaks RF. You need to add AMS to this part of settings:
Add: advanced_model_sampling_script,advanced_model_sampling_script_backported to there.
If that does not work, go into adetailer extension, find args.py, open it, replace builtinscripts like this:
Here is a copypaste for easy copy:
_builtin_script = (
"advanced_model_sampling_script",
"advanced_model_sampling_script_backported",
"hypertile_script",
"soft_inpainting",
)
Or use my fork of Adetailer - https://github.com/Anzhc/aadetailer-reforge
LoRA Training
You can directly reference config with all parameters: Download
Hardware
Model was trained on cloud 8xH100 node.
Software
Custom fork of SD-Scripts(maintained by Bluvoll)
Acknowledgements
Sponsors
To a special supporter who singlehandidly sponsored whole run and preferred to stay anonymous
Testers
ComradeAnanas
Daruda
Drac
itterative
kagame
Remix
Ryusho
edf
Epic
Ly
Panchovix
Rakosz
Sab
Silvelter
Talan
Void
Why ping
Support
If you wish to support our continuous effort of making waifus 0.2% better, you can do it here:
https://ko-fi.com/bluvoll (Blu, donate here to support training)
https://ko-fi.com/anzhc (Anzhc, non-training, just survival)
BTC: 37fLcfxX5ewhJXnb3T9Qzu9jiSLjVtoUJX
ETH: 0xfdF54655796bf2F5bf75192AeB562F8656c1C39E
Send DM to Blu if you want to donate on another network.
Description
FAQ
Comments (32)
It's a bit inconvenient that it needs this special Empty Latent node but the model is really great! It actually has the latest character knowledge cutoff of any Anime model at the moment :)
(for reference: https://civitai.com/articles/14008/character-knowledge-cutoff-comparison)
If you mean the node that comes with the custom nodepack doesn't need to be explicitly used, it should work without actually adding the special empty latent node to a workflow
@DraconicDragon You still have to install the custom node to be able to use the model. And when I use the model with the normal Empty Latent node, the resolution of the resulting image is much lower than expected.
@lizardon1024 You can bother comfyanon about it if you want it to be supported in comfy by default.
@lizardon1024 Oh yeah, installing the nodepack is indeed a slight inconvenience
I wasn't really aware of the issue with the resulting image being lower/half res since it works for me - I asked 2 other people and for one it worked fine with other latent nodes like me, and for the other they seem to be experiencing the same issue you are
If you installed from comfyui manager, did you install with "latest" option? It could be that since that will install 1.0.0 which is a bit older than the most recent commit(s) which includes some fixes. If you switch to "nightly" I believe it should work
I pinged Anzhc in their discord about this to update the nodepack version on the manager so that people installing with the default "latest" version selection get the fixes
@DraconicDragon A great, yeah switching to Nightly solved this issue for me as well. Thank you! :)
LLM TE, Rectified Flow, 16-channel VAE...
Everything was good and effective.
Now the SDXL UNet is sweating right now. It's the last piece of the puzzle for the Ship of Theseus.
Positive quality tags have a stronger impact than negative ones, which, as I recall, is worse because it's better to specify 100 things you don't want to see than to add abstract tags like "masterpiece or best." The model has very good aesthetics, but terrible quality, especially in terms of clarity. The best in terms of character understanding, but still makes them worse than competitors. It's easier to find the right character with "lore" (I tried to make trigger_(zenless_zone_zero) specifying 15 tags plus the name, I succeeded, but I wasn't particularly happy with the result). We're waiting for new versions; as I recall, the model is clearly better than "anima." P.S. It's clearly too early to leave Illustris.
Negative has a strong impact.
How is it better than Anima?
From my testing it's not really that different from general noobai/illustrious
@NekoYaSan It is not better than anima, but comparing both is problematic since each have issues, and this is more of a project showing that this can be done on very limited budget.
Specially true when we applied for the Comfy Grant and we were ignored, most likely in favor of anima, take this as a 'cool' toy that might or might not have a future.
@bluvoll I admire your efforts, and I hope my question wasn't taken as criticism.
The very fact that guys can get together like this and make their own model is very cool.
It's just that for me, as a regular user, the difference is not so noticeable.
It seems to me that they preferred the "Anima" because of the text encoder that gives the "brains" to the model, which looks more promising. At least that's how I understand it.
@NekoYaSan no worries, I understand, since I'm fully aware of this model's issues
@bluvoll By the way your model also works in webui forge neo.
Need to add rectified in the file name e.g "mugenSDXLWithFlux2s_05_rectified".
And use Lumina preset for shift to appear.
need community support and civit ai needs to add another category for this model.
given how well it gens i honestly prefer this over current chenkin and noob + their other finetunes, didnt know these details were even possible.
Cool model. Better patterns and details, but needs more training
Does this work in Reforge?
@DiffusedSignal yes, just make sure its up to date
It's miles behind other models right now. Quality is not there. Hoping it's eventual full release will be great!
This is always the problem with model tech advancement... Almost universally every upgrade is far behind current generations because it hasn't been finetuned to hell and back. It takes people taking these foundation models and building upon them before we end up truly ahead.
Great model, had the pleasure to early test it, seems like the best possible future for sdxl based tech
seems promising, thanks
very cool project, I am looking forward to further training progress!
till this day I am using SDXL as a final polish step. there is something in the unet which the newer dit models simply cannot achieve it terms of clarity and reduction of noise, even considering the bugged original sdxl vae.
let's see what unet can do with the flux 2 vae!
Can't run with SwarmUI, custom node installed.
You should set the model prediction type to sd3 in the models tab.
cool...
It's a nice model, and actually something innovative, been following this huge effort from the very start. But it seems like at this point, it forgot some characters, just an observation
What's the cutoff date for this dataset?
@Luigiman98398 feb 2026
@bluvoll Does it include artist tags?
@Luigiman98398 yes
I don't know if you are still active but I really like the concept! I am new to the stable diffusion scene and I'm trying to learn how these models work. I would love to use this model as a base to practice working with. I want to learn the benefits of Rectified Flow and V-Pred and how I can use this model to bolster ones I already like.
Also, looking at the results with the FLUX 2 VAE is insane, I really love FLUX but my RTX 4050 can't run it without massive offloading to my RAM and it's very slow tbh. So having the VAE that FLUX uses on a base model that I can run decently fast is really appreciated.
I wish you well with your work.