Anima-RGBA

Anima-RGBA

Adds transparent-background (RGBA) image output to the Anima model through a 4-channel VAE and a matching LoRA, for use in ComfyUI. For artists who need cutouts with a clean alpha channel.

Made by localai0301 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Anima-RGBA lets the Anima image model produce pictures with a real alpha channel, so the background is truly transparent. It has two parts: an RGBA VAE and an RGBA LoRA.

The VAE is a four-channel variant of the Qwen-Image VAE behind Anima. Where the original gives opaque RGB, it decodes latents into pictures that carry a true alpha channel. The LoRA teaches Anima to make latents that this VAE decodes cleanly.

Who it suits: Artists and builders who make cutout images in ComfyUI and want transparent backgrounds without separate background removal.

Industry | Who uses it | How they use it | Example
Game development | Artists who make sprites and props | Generate objects on a transparent background | A character cutout ready to place in a scene
Design | Graphic designers | Produce isolated elements as RGBA PNG files | A sticker-like illustration with clean edges
Web and apps | Interface designers | Create images that sit on any page color | An icon-style picture over a colored banner
Research | Image model researchers | Study alpha-capable VAE training | Compare alpha error and reconstruction quality

Skip it if you do not use Anima, since the VAE and LoRA are built for that model.

Where to use it:

• ComfyUI, with an existing Anima workflow and no custom nodes
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Transparent illustrations | Artists | An RGBA PNG with a clean alpha channel
Isolated objects | Designers | A subject with no background to remove
Inspecting alpha | Researchers | A separate alpha channel view of each result

How to set it up:

• ComfyUI with a working Anima workflow
• Python and pip to run the download command
• The two files from this repository, about 508 MB for the VAE and 184 MB for the LoRA

1. Download the repository files.
2. Put the VAE file in the models/vae folder and select it in the Load VAE node instead of the Qwen-Image VAE.
3. Copy the LoRA file into models/loras, then load it with a standard LoRA loader node set to strength 1.0.
4. Add the trigger phrase transparent background to your prompt.

pip install -U huggingface_hub
hf download localai0301/Anima-RGBA

File | Size | Notes
anima_rgba_vae.safetensors | 508 MB | Full 4-channel VAE with encoder and decoder, in the key format ComfyUI expects for the Qwen-Image VAE
anima_rgba_lora.safetensors | 184 MB | LoRA for the Anima DiT, saved at step 25,000

How to use it: Write your prompt as usual and include one of the trigger phrases. When the result still has a solid background, give the trigger phrase more weight and leave the LoRA strength alone.

transparent background
isolated on a transparent background
(transparent background:1.3)

After that, the VAE Decode node gives you four channels, and Save Image stores the file as an RGBA PNG. Raising the trigger weight barely changes the picture. Pushing the LoRA to 1.5 leaves stray, partly see-through specks near the subject.

Setting | Value in the card examples
LoRA strength | 1.0
Steps | 30
CFG | 4.5
Sampler and scheduler | er_sde and simple
Size | 1024 by 1024

Example: You add the trigger phrase to a prompt such as "transparent background, strawberry shortcake" and run the edited workflow. Save Image writes an RGBA PNG of the cake with nothing behind it.

For newcomers and experts: If you are new to ComfyUI, take a workflow that already runs Anima and change only the VAE, the LoRA and the prompt. Open the saved PNG over a checkerboard to see the transparency.

If you know the field, the new encoder takes RGB composited over white plus alpha, and the decoder returns straight RGB plus alpha. A reference loss ties its latents to the original encoder, which leaves the LoRA a small alpha difference to learn.

Prompt with a transparent background trigger
Anima with the RGBA LoRA makes latents
RGBA VAE decodes to four channels
RGBA PNG saved

Good to know:

• Without the LoRA the VAE still makes an alpha channel, but it looks hazy
• The card tested 200 transparent and 50 opaque holdout images at 512 px, none of them used in training
• The alpha mean absolute error was 0.0027 on the transparent images
• The new latents drift about 6 percent of a standard deviation from the original ones

Title: PSNR on opaque images, higher is better
Original Qwen-Image VAE: 34.26 dB
Anima RGBA VAE: 34.74 dB

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the prompts you give your workflow
• Not for: use without Anima, since both files are made for it

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Downloads in the last 30 days: 3,509
• Likes: 59
• License: apache-2.0
• Based on: circlestone-labs/Anima
• Tags: diffusers, lora, rgba, transparency, alpha, anima, qwen-image, comfyui, text-to-image

Links

• Model: https://huggingface.co/localai0301/Anima-RGBA
• Files: https://huggingface.co/localai0301/Anima-RGBA/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download localai0301/Anima-RGBA

Then follow the usage in its model card. Nothing in this kit runs by itself.