Hexer Minimal Toon Anima V1 vs Illustrious v3.1: base, folders, and license
Contents
Civitai added an Anima V1 to Hexer Minimal Toon | Mexes.
It was published at 2026-06-05 08:47 JST. On the model page it’s still the same “Hexer Minimal Toon,” but the base is different from the existing Illustrious v3.1.
The Illustrious version is used as an SDXL-family checkpoint.
The Anima version is treated as an Anima-family DiT checkpoint.
Same author, same style name, but the folder you drop it in, the loader that reads it, and the LoRA you combine with it are all different.
A checkpoint aiming for the midpoint between toon and anime
On the model page, the author Mexes describes the concept of Hexer Minimal Toon as “the sweet spot that takes the best of both toon and anime.”
Detailed but not too detailed, toon-like body construction but not exaggerated, anime-leaning colors but not a full anime flat-color paint job.
He writes that keeping this balance is the direction of the whole project.
Image generation models can end up with too-dense textures depending on the tuning, or conversely become flat after detail is stripped away.
As the name “Minimal” suggests, Hexer Minimal Toon intentionally dials down excessive rendering at generation time.
It avoids flat coloring while keeping anime-like color choices and a moderate sense of volume.
The author also writes that “Limiting things too much isn’t always the best approach,” and although individual versions may lean more toon or more anime, he keeps the harmony between the two from breaking down.
He even recommends trying older versions if the latest one doesn’t suit you.
Anima V1 is a 4GB-class bf16 full version
The Anima V1 file you can confirm via the Civitai API is hexerMinimalToon_animaV1.safetensors.
Its size is about 4,084,198KB, metadata is SafeTensor, full, bf16.
The AutoV2 hash is 5D111A8DC7.
The existing Illustrious v3.1 is hexerMinimalToon_illustriousV31.safetensors.
This one is about 6,775,432KB, SafeTensor, pruned, fp16.
The AutoV2 hash is 545E88985A.
| Version | Base | File | Size | Precision |
|---|---|---|---|---|
| Anima V1 | Anima | hexerMinimalToon_animaV1.safetensors | ~4.1GB | bf16, full |
| Illustrious v3.1 | Illustrious | hexerMinimalToon_illustriousV31.safetensors | ~6.8GB | fp16, pruned |
The Anima version being lighter doesn’t mean it’s a lightweight build of the SDXL version.
The base model field is set to Anima, so the whole lineage is different.
SDXL-family and Anima-family are different generation engines to begin with
“Different base” doesn’t click on its own, so let me separate what’s inside the two lineages.
Illustrious (SDXL family) is a checkpoint that descends from Stable Diffusion XL.
The main network that handles image denoising is a U-Net (a convolution-based encoder-decoder structure), and the text side conditions on a CLIP-family encoder.
You can reuse assets built for SDXL as-is, including LoRA, ControlNet, and the various samplers.
Anima (Anima family) is, as I’ve covered several times on this blog, a DiT (Diffusion Transformer) based on NVIDIA Cosmos-Predict2-2B-Text2Image.
Instead of U-Net convolutions, DiT splits the image into small patches (tokens) and processes them with Transformer attention.
The text encoder is Qwen3 0.6B (qwen_3_06b_base.safetensors) and the VAE is Qwen Image VAE (qwen_image_vae.safetensors), so the combination of parts is entirely different from SDXL.
Here’s a rough contrast of what changes when you switch to DiT.
| Aspect | Illustrious (SDXL family, U-Net) | Anima (Anima family, DiT) |
|---|---|---|
| Denoising backbone | U-Net (convolution) | Diffusion Transformer (patches + attention) |
| Text encoder | CLIP family | Qwen3 0.6B (separate file) |
| VAE | SDXL-standard VAE family | Qwen Image VAE (separate file) |
| Prompt following | Strongly assumes tag strings (Danbooru tags) | Easier to lean on long-form / natural language via attention |
| High resolution / composition | Pushed up via hires fix etc. | Transformer scales more easily along the resolution axis |
| LoRA compatibility | SDXL/Illustrious LoRA | Anima LoRA only |
| ControlNet | Plenty of compatible assets | No official offering confirmed (as of writing) |
The trends in the table are the general differences between U-Net-family and DiT-family; I haven’t verified that the two Hexer Minimal Toon versions necessarily behave exactly like this.
But at minimum, the file and base model fields make it clear that the combination of loader, text encoder, VAE, and LoRA swaps out by lineage.
Illustrious v3.1 is a DARE merge of Illustrious V2 and RouWei v0.8
The model description says Illustrious v3.1 is a DARE merge using “Illustrious V2 and RouWei v0.8.”
The earlier toon-anime-midpoint direction is realized through a DARE merge in this version.
The recommended settings listed are 40 steps, resolution 836x1254, Euler A or DPM++ 2M Karras, CFG 7.
This description is about the Illustrious version.
Anima V1 doesn’t have a version-specific description yet.
There are 5 sample images, all at 1024x1536, but the meta info was empty on the API.
The Anima version switches existing LoRA compatibility
The Anima family has a different loader and different LoRA from the Illustrious and SDXL families.
As I wrote in the article lining up Anima-derivative checkpoints, Anima is an NVIDIA Cosmos-family DiT that uses the Qwen3 0.6B text encoder and Qwen Image VAE.
It’s not something where you drop an SDXL LoRA straight onto Anima.
Conversely, an Anima LoRA will load across Anima derivatives.
But the output changes with each derivative’s tuning.
In the comparison applying the same character LoRA to 6 Anima derivatives, even the same WAI-Anima LoRA varied quite a bit in the reproducibility of a single trigger word and in color.
The Anima version of Hexer Minimal Toon is also treated as a new derivative that goes into that comparison set.
Rather than carrying over the style LoRA and character LoRA you were using with the Illustrious version, it’s a checkpoint where you check compatibility with Anima LoRA.
Which one to pick
When two versions sit on the same model page, it’s hard to decide which one to download.
Since the style name (the midpoint between toon and anime) is shared, the call comes down to your existing assets and workflow.
The Illustrious version fits cases like these.
You already have SDXL/Illustrious LoRA assets and want to make use of them.
You want to control pose and composition with ControlNet.
Illustrious v3.1 lists recommended settings of 40 steps, 836x1254, Euler A or DPM++ 2M Karras, CFG 7, so there’s a baseline for how to generate.
The Anima version fits cases like these.
You’ve already built an Anima-family workflow and want to reuse Anima LoRA.
Since Anima derivatives share the same base, text encoder, and VAE, you can load a character LoRA trained for Anima with the same feel as other derivatives.
The file is light at about 4.1GB and is distributed as a bf16 full version.
If you have assets for both lineages, downloading both and comparing is the fastest path in the end.
But Illustrious and Anima differ in prompt conventions (tag strings, or tags + natural language), so if you feed the same string straight through, you can mistake the difference in convention for a difference in model.
This Hexer Anima V1 itself has both an empty version-specific description and empty sample generation meta, but the base Anima official model card lists recommendations of 30-50 steps, CFG 4-5, and the sampler er_sde, so you can start there for your first attempt.
What to separate and check before using it
If you’re going to try it in ComfyUI, separate the Illustrious and Anima versions by workflow.
Treat Illustrious v3.1 as a normal SDXL-family / Illustrious-family checkpoint.
Handle Anima V1 in an Anima-family workflow. Even the place you put it differs from the Illustrious version.
The Illustrious version is a normal checkpoint, so put it in ComfyUI/models/checkpoints/.
For the Anima version, Anima’s official guidance is to place the backbone, text encoder, and VAE separately: the backbone hexerMinimalToon_animaV1.safetensors goes in ComfyUI/models/diffusion_models/, the text encoder qwen_3_06b_base.safetensors in ComfyUI/models/text_encoders/, and the VAE qwen_image_vae.safetensors in ComfyUI/models/vae/.
The file names are similar too, so once they mix inside a folder it’s hard to track later.
If you drop everything into checkpoints/ with the Illustrious mindset, the Anima version won’t load, so this is the first stumbling point.
The author explains that he posts the preview images as raw output, using none of hires fix, ADetailer, inpaint, or post-processing.
That’s a welcome condition for seeing the model’s bare art style, but since the generation meta on the Anima V1 side is empty, there’s still little material to reproduce the same conditions at hand.
The license reads differently for the Illustrious version and the Anima version
The display on the Civitai API is: use without credit permitted, commercial use of generated images permitted, derivative creation permitted, changing to a different license not permitted, and the license shown on the model page is the Illustrious License.
This display was originally premised on the Illustrious version, and applying it directly to the Anima version is dangerous.
The official license of the upstream Anima (CircleStone) that the Anima version is based on is non-commercial only.
Use of the model itself, derivative models, LoRA, and textual inversion is limited to non-commercial purposes, and commercial use requires a separate license.
On the other hand, it separately specifies, apart from the model itself, that the generated output images may be used commercially.
Furthermore, since Anima is a derivative of NVIDIA Cosmos-Predict2, distribution requires a “Built on NVIDIA Cosmos” attribution notice alongside CircleStone’s license notice.
In other words, you can’t say the Anima version is “free to use without credit.”
Using the output images commercially is mostly fine, but if you commercially use, redistribute, or create derivatives of the model or LoRA, check the upstream Anima license and NVIDIA’s terms directly rather than the Civitai page display.