<h1 id="liquidn2vae_ig4">liquidn2vae_IG4</h1><p>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.</p><h2 id="basic-idea">Basic idea</h2><p>The program trains the VAE using image reconstruction.</p><p>The target image is created with PIL image enhancement:</p><pre><code>saturation: 1.18
contrast: 1.03
brightness: 1.00
</code></pre><p>This means the target is not a different image, but the same image with mild saturation and contrast enhancement applied.</p><p>The encoder is mostly kept stable.<br />The main adjustment happens on the decoding side, so the VAE learns to reconstruct images with a more saturated color response.</p><h2 id="loss-function">Loss function</h2><p>The script uses a combined reconstruction loss.</p><p>The total loss is built from three parts:</p><h3 id="1.-l1-loss">1. L1 loss</h3><p>L1 loss compares the decoded image and the target image pixel by pixel.</p><p>This is the main reconstruction loss.</p><p>Default weight:</p><pre><code>L1_WEIGHT = 1.0
</code></pre><h3 id="2.-mse-loss">2. MSE loss</h3><p>MSE loss also compares the decoded image and the target image, but it penalizes larger differences more strongly.</p><p>Default weight:</p><pre><code>MSE_WEIGHT = 0.25
</code></pre><h3 id="3.-edge-loss">3. Edge loss</h3><p>The script also calculates a simple edge-preservation loss.</p><p>It compares horizontal and vertical pixel differences between the decoded image and the target image.<br />This is intended to reduce excessive blurring or edge collapse while the VAE learns the stronger color response.</p><p>Default weight:</p><pre><code>EDGE_WEIGHT = 0.05
</code></pre><p>The final loss is:</p><pre><code>loss = L1_WEIGHT * L1
     + MSE_WEIGHT * MSE
     + EDGE_WEIGHT * EdgeLoss
</code></pre><h2 id="optimizer">Optimizer</h2><p>The script uses AdamW.</p><p>Typical settings used for this VAE:</p><pre><code>LR = 0.000003
EPOCHS = 1
BATCH_SIZE = 1
IMAGE_SIZE = 1024
</code></pre><p>The learning rate is intentionally very small.<br />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.</p><h2 id="what-this-changes">What this changes</h2><p>Because the training target is a saturation-enhanced version of the same image, the trained VAE tends to decode images with:</p><ul><li><p>stronger saturation</p></li><li><p>warmer color response</p></li><li><p>clearer floral colors</p></li><li><p>stronger blue skies</p></li><li><p>richer anime-style color output</p></li></ul><p>It does not intentionally change composition, prompt understanding, anatomy, or object placement.<br />Those are handled by the diffusion model, not the VAE.</p><p>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.</p><h2 id="known-limitations">Known limitations</h2><p>This is an experimental VAE fine-tune.</p><p>Known issues:</p><ul><li><p>Shadows around line art may shift toward pink or magenta.</p></li><li><p>Photorealistic images may have unstable saturation.</p></li><li><p>Neutral gray areas may become warmer than intended.</p></li><li><p>Skin tones may become slightly more colorful.</p></li><li><p>Already vivid images may become over-saturated.</p></li><li><p>It is better suited for anime-style or illustration-like outputs than strict photorealistic reproduction.</p></li></ul><p></p>

liquidn2vae_IG4

<p>We increased the amount of training to make the differences from other VAEs clearer.</p>

Files with SHA256: 535a6733d8e0ad37958af774fbb614ab3b92d85b26bbf1abb41b6160beb82423