Anima-RGBA Adds Native Transparent Backgrounds to Anime AI Without Extra Tools
A LoRA plus modified VAE for the Anima model that generates clean RGBA images with a real alpha channel, no background removal needed.
- Anima-RGBA adds native transparent background generation to the Anima anime diffusion model.
- Ships as a 508 MB RGBA VAE and 184 MB LoRA, Apache-2.0 licensed on Hugging Face.
- Drops into existing ComfyUI Anima workflows with no custom nodes, just a VAE swap, LoRA load, and trigger phrase.
- Hits 0.0027 alpha MAE and 34.94 dB PSNR, with cleaner cut-outs than rembg-style post-processing.
- VAE initialized so it behaves identically to the original before training; latent drift only 6% of a standard deviation.
- Weak on fine semi-transparency like hair strands, glass, and smoke; alpha is mostly binary.
Anima-RGBA files add native transparent-background generation to Anima, an anime illustration model derived from Qwen-Image. Its decoder writes a fourth alpha channel containing per-pixel opacity, allowing ComfyUI to save an RGBA PNG directly. Conventional workflows generate against a flat color and then estimate a mask with rembg, BiRefNet, or SAM, adding a second pass that can leave halos or erase fine details.
The release contains a 508 MB modified variational autoencoder, or VAE, and a 184 MB LoRA adapter for the diffusion transformer. The VAE encodes and decodes four-channel images; the LoRA guides Anima toward latent representations that the new decoder can separate into color and opacity. Both files are available under Apache-2.0, though developers should review the upstream Anima license separately.
Three changes, no custom nodes
A standard ComfyUI workflow for Anima needs three changes:
- Replace the Qwen-Image VAE with
anima_rgba_vae.safetensorsin the Load VAE node. - Load
anima_rgba_lora.safetensorsthrough a standard LoRA loader at strength1.0. - Add
transparent backgroundorisolated on a transparent backgroundto the prompt.
The modified VAE Decode node emits straight RGB plus alpha, and Save Image writes the result as an RGBA PNG. When a background remains opaque, the author recommends increasing the prompt weight on the transparency phrase while keeping the LoRA at 1.0. Tests at LoRA strength 1.5 produced stray semi-transparent pixels around subjects.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
