Introduction

Sticker makers, VTuber riggers and game-UI artists, this one's for you. Anima RGBA landed on Hugging Face on Nov 2 (00:28 UTC) from localai0301. It's an open-weights add-on that makes circlestone-labs' Anima anime/illustration model generate images with a real alpha channel. You don't get a white background that you mask later. The output is a 4-channel RGBA PNG straight out of Save Image.

It ships as two files under Apache-2.0, and it runs in stock ComfyUI with no custom nodes:

  • vae/anima_rgba_vae.safetensors (508 MB): a 4-channel version of the Qwen-Image VAE that Anima uses
  • loras/anima_rgba_lora.safetensors (184 MB): a rank-32 LoRA for the Anima DiT, trained for 25,000 steps on latents from that VAE

How it works

The VAE starts from the Qwen-Image weights, with a new alpha input and output that initially behave like the original. It was then trained to decode straight RGB + alpha. During training, a reference loss kept the new encoder about 6% of a standard deviation away from the original latent space, so Anima barely has to move. The LoRA closes that small gap so the model draws latents that decode into clean cut-outs.

The author's numbers on a holdout set:

  • Alpha error (MAE): 0.0027, about 0.7/255, with most of it on edges
  • PSNR on transparent images composited over white: 34.94 dB
  • PSNR on ordinary opaque images: 34.74 dB vs 34.26 dB for the original Qwen-Image VAE, so it doesn't hurt normal renders
  • Semi-transparent "haze" pixels drop from 12.8% to 0.9% with the LoRA at 1.0. That's with Anima base at 30 steps and CFG 4.5. It also holds up on a fine-tune with a turbo LoRA (21.1% → 3.5%)

Set it up in ComfyUI

Take any working Anima workflow and change three things:

  • Put the VAE in models/vae and select it in Load VAE instead of the Qwen-Image VAE
  • Put the LoRA in models/loras and add it with a normal LoRA loader at strength 1.0
  • Add transparent background (or isolated on a transparent background) to your prompt

cd ComfyUI/models
hf download localai0301/Anima-RGBA vae/anima_rgba_vae.safetensors --local-dir ./_anima_rgba
hf download localai0301/Anima-RGBA loras/anima_rgba_lora.safetensors --local-dir ./_anima_rgba
mv _anima_rgba/vae/anima_rgba_vae.safetensors vae/ && mv _anima_rgba/loras/anima_rgba_lora.safetensors loras/

The author's example settings are 30 steps, CFG 4.5, the er_sde sampler with the simple scheduler, at 1024×1024. Here's one of the example prompts from the model card:

masterpiece, best quality, newest, safe, no humans, transparent background, food, strawberry shortcake, fruit, still life

worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts

If the background stays opaque, raise the trigger weight, for example (transparent background:1.3), and leave the LoRA strength alone. At 1.5 the LoRA starts adding stray semi-transparent smoke around the subject.

Caveats

  • It's built for anime and illustration content on Anima. It's not a general photo matting tool.
  • Generated alpha is mostly binary. Very fine semi-transparency like hair strands, glass and smoke is the weakest area, and see-through clothing usually comes out opaque.
  • Close-ups that fill the frame sometimes keep a white background, especially with other LoRAs stacked.
  • The repo is a few hours old (8 likes at time of writing). Check the base Anima model's own license before commercial use.

If you work in MOSS-Image instead, we covered a similar trick last week with ckmoss RGBA. This is the native option for Anima users.

Original Source

Conclusion

Transparent output has always been the annoying last mile: render, then remove the background, then clean up the fringes. Anima RGBA folds all of that into the VAE, and the numbers say it doesn't cost you anything on normal images. Swap two files, type four words, and your character is ready for the sticker sheet. ✨

—Aurelia ♡