paralaif/Krea2-Manga-Alpha-Image-LoRA

Krea 2 Manga Alpha Image LoRA

Transparent PNGs straight out of Krea 2 Turbo in 8 steps. Prompt as usual, end with , isolated on a transparent background, decode with the two Wan-Alpha VAEs included here, and you get an image with a real alpha channel: no chroma key, no background-removal model, no masking, and no colored fringe on the edges. Only built-in ComfyUI nodes are needed.

It was trained with black-and-white manga in mind (characters, creatures, props, ink effects, sound effects, speech balloons), but it also works in color.

Trailer made only from outputs of this LoRA, composited on colored backgrounds.

Gallery

30 raw RGBA outputs (Krea 2 Turbo, 8 steps, nothing retouched). The page background shows through the transparent areas. Full-resolution files are in examples/.

How it works

Krea 2 uses the Wan 2.1 latent space: its VAE encoder is identical to Wan 2.1's. Wan-Alpha extends that space with alpha. Its encoder fuses the image and its alpha matte into one 16-channel latent, and two decoders read that latent back as color and as alpha. The fused latent is not the plain Krea latent, so this LoRA teaches Krea 2 to generate it directly. No changes to Krea itself are needed at inference time.

What's in this repo

File	Put it in	Notes
krea2_manga_alpha_image_lora.safetensors	models/loras/	the LoRA (rank 32)
vae/wan_alpha_v2_vae_rgb.safetensors	models/vae/	Wan-Alpha v2 color decoder (as a regular Wan 2.1 VAE file)
vae/wan_alpha_v2_vae_alpha.safetensors	models/vae/	Wan-Alpha v2 alpha decoder
comfyui/krea2-manga-alpha-image-t2i.json	drag into ComfyUI	minimal workflow, tested with the files below

Not included (download from the original repos):

File	Put it in	Source
krea2_turbo_fp8_scaled.safetensors	models/diffusion_models/	Comfy-Org/Krea-2
qwen3vl_4b_fp8_scaled.safetensors	models/text_encoders/	Comfy-Org/Krea-2

How to use (ComfyUI)

1. Place the files from the tables above and load comfyui/krea2-manga-alpha-image-t2i.json.
2. Write your prompt and end it with , isolated on a transparent background. The captions followed black-and-white manga illustration of <subject>, isolated on a transparent background.
3. The workflow decodes the latent twice (color and alpha) and joins them with ImageToMask → InvertMask → JoinImageWithAlpha. The saved PNG is RGBA.

Settings used for every image on this page

Model	Krea 2 Turbo fp8
LoRA strength	1.0
Sampler	8 steps, euler / simple, CFG 1
Resolution	832×1216, 1216×832, 1024²

Training

Base	Krea 2 Raw, musubi-tuner (krea2_train_network.py, ConvRot int8 base)
Latents	encoded with the Wan-Alpha v2 encoder (Wan 2.1 encoder + Wan-Alpha merge module) from RGBA PNGs
Data	503 RGBA images, captions black-and-white manga illustration of <subject>, isolated on a transparent background
LoRA	rank 32, alpha 32
Optimizer / LR	AdamW 8-bit, 1e-4, krea2_shift timesteps, 1024 px buckets, batch 1
Steps	about 2500 run, epoch 4 (2012 steps) released; epoch 5 started treating black ink as transparent
Hardware	1× RTX 5060 Ti 16 GB, about 10 s/step

About the data. The training images are synthetic. They were 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 (373 rejected). The artist tags were removed from the captions, but the LoRA may still carry some of those styles.

Limitations

• Dense black masses are the weak spot. A flock of black crows failed on every seed, and solid black areas sometimes turn partly transparent.
• About 1 in 5 prompts needed another seed to get a clean background; the failure is an opaque pale-gray background, which is easy to spot.
• It is tuned for Krea 2 Turbo. On heavier fine-tunes it degrades: on KreaMaxEdit dark clothing turns semi-transparent, and on Kroma the alpha comes out inverted.
• Describe one subject, not a background or a full scene.

License and credits

This LoRA is a Derivative of the Krea Model and is distributed under the Krea 2 Community License Agreement (copy in LICENSE.pdf; attribution in NOTICE). By using it you agree to that license, including its commercial-use revenue threshold and the requirement to use content filters when you deploy it. It is not an official Krea product and is not endorsed by Krea.

• Krea 2 by Krea: Krea 2 Community License.
• Wan-Alpha VAE by htdong et al., htdong/Wan-Alpha-VAE: Apache-2.0. The decoders here are a converted copy (decoder LoRA merged, one file per decoder).
• MossAD-Demo by RicemanT (training-data generator): Apache-2.0.
• musubi-tuner by kohya-ss: Apache-2.0.

The same data also trained the MOSS-Image version: paralaif/MOSS-Manga-Alpha-Image-LoRA.