Transparent PNGs straight out of Krea 2 Turbo in 8 steps. End your prompt with isolated on a transparent background, decode with the two Wan-Alpha VAEs, and you get an image with a real alpha channel: no chroma key, no background removal, no masking, and no colored fringe on the edges. Only built-in ComfyUI nodes.
Trained with black-and-white manga in mind (characters, creatures, props, ink effects, sound effects, speech balloons), but it also works in color.
Everything you need (ComfyUI): https://huggingface.co/paralaif/Krea2-Manga-Alpha-Image-LoRA
1. This LoRA → models/loras/
2. vae/wan_alpha_v2_vae_rgb.safetensors and vae/wan_alpha_v2_vae_alpha.safetensors (Hugging Face repo, converted from htdong/Wan-Alpha-VAE, Apache-2.0) → models/vae/
3. Minimal workflow: the attached zip (also on Hugging Face)
4. Krea 2 Turbo + Qwen3-VL-4B text encoder: Comfy-Org/Krea-2 on Hugging Face
How it works: Krea 2 shares the Wan 2.1 latent space. Wan-Alpha extends that space with alpha (one fused latent, two decoders), and this LoRA teaches Krea 2 to generate that fused latent directly.
Tips
- Always end the prompt with isolated on a transparent background, and describe one subject, not a place.
- About 1 in 5 prompts needs another seed; the failure is an obvious opaque pale-gray background.
- Dense black masses (a flock of crows) are the weak spot.
- Tuned for Krea 2 Turbo; on heavy fine-tunes like Kroma the alpha breaks.
Training: Krea 2 Raw, musubi-tuner, latents encoded with the Wan-Alpha v2 encoder, 503 RGBA images, rank 32, AdamW 8-bit 1e-4, krea2_shift, epoch 4 (2012 steps), one RTX 5060 Ti 16 GB. The training images are synthetic: generated with MossAD-Demo using its built-in artist-style tags, on a flat indigo background, cut out with a chroma key and reviewed by hand. The artist tags were removed from the captions, but the LoRA may still carry some of those styles.
License: Krea 2 Community License Agreement (included in the zip; https://krea.ai/krea-2-licensing). Krea 2 is licensed under the Krea 2 Community License Agreement. This LoRA modifies the Krea Model; it is not an official Krea product and is not endorsed by Krea. Commercial use is only allowed within the license's revenue threshold, and deployments must use content filters. Credits: Krea 2 (Krea), Wan-Alpha VAE (htdong et al., Apache-2.0), MossAD-Demo (RicemanT, Apache-2.0), musubi-tuner (kohya-ss, Apache-2.0).
Also available for MOSS-Image: https://huggingface.co/paralaif/MOSS-Manga-Alpha-Image-LoRA