liquidn2vae_IG4

This VAE is based on a FLUX.2-style VAE used in Ideogram 4 workflows, and has been further adjusted for more vivid color reproduction in Ideogram 4-style image generation.

Basic idea

The program trains the VAE using image reconstruction.

The target image is created with PIL image enhancement:

saturation: 1.18
contrast: 1.03
brightness: 1.00

This means the target is not a different image, but the same image with mild saturation and contrast enhancement applied.

The encoder is mostly kept stable.
The main adjustment happens on the decoding side, so the VAE learns to reconstruct images with a more saturated color response.

Loss function

The script uses a combined reconstruction loss.

The total loss is built from three parts:

1. L1 loss

L1 loss compares the decoded image and the target image pixel by pixel.

This is the main reconstruction loss.

Default weight:

L1_WEIGHT = 1.0

2. MSE loss

MSE loss also compares the decoded image and the target image, but it penalizes larger differences more strongly.

Default weight:

MSE_WEIGHT = 0.25

3. Edge loss

The script also calculates a simple edge-preservation loss.

It compares horizontal and vertical pixel differences between the decoded image and the target image.
This is intended to reduce excessive blurring or edge collapse while the VAE learns the stronger color response.

Default weight:

EDGE_WEIGHT = 0.05

The final loss is:

loss = L1_WEIGHT * L1
     + MSE_WEIGHT * MSE
     + EDGE_WEIGHT * EdgeLoss

Optimizer

The script uses AdamW.

Typical settings used for this VAE:

LR = 0.000003
EPOCHS = 1
BATCH_SIZE = 1
IMAGE_SIZE = 1024

The learning rate is intentionally very small.
The purpose is not to heavily retrain the VAE, but to slightly shift the decoder’s color response while keeping the original Ideogram 4 / FLUX.2-style VAE behavior mostly intact.

What this changes

Because the training target is a saturation-enhanced version of the same image, the trained VAE tends to decode images with:

• stronger saturation

• warmer color response

• clearer floral colors

• stronger blue skies

• richer anime-style color output

It does not intentionally change composition, prompt understanding, anatomy, or object placement.
Those are handled by the diffusion model, not the VAE.

In short, liquidn2vae_IG4 is intended as a VAE replacement for Ideogram 4-style ComfyUI workflows when a slightly more vivid and colorful output is desired.

Known limitations

This is an experimental VAE fine-tune.

Known issues:

• Shadows around line art may shift toward pink or magenta.

• Photorealistic images may have unstable saturation.

• Neutral gray areas may become warmer than intended.

• Skin tones may become slightly more colorful.

• Already vivid images may become over-saturated.

• It is better suited for anime-style or illustration-like outputs than strict photorealistic reproduction.