<h1 id="qwen-image-vae">Qwen Image VAE</h1><ul><li>Full FP32 Training of Decoder</li><li>Works in ComfyUI</li></ul>Feel free to to suggest onsite support, to civitai staff. I don't think they have any agreements like with FLUX<hr /><h2 id="overview">Overview</h2>This model is a fine-tuned variant of the base Qwen Image VAE, modified to emphasize high-frequency detail preservation and expanded color representation, following an HDR-style reconstruction objective.The evaluation compares the base and HDR-tuned models using perceptual, structural, distributional, and photometric metrics over identical input data.<hr /><h2 id="evaluation-summary">Evaluation Summary</h2><h3 id="perceptual-fidelity-(lpips)">Perceptual Fidelity (LPIPS)</h3><ul><li>Base: 0.0177</li><li>HDR: 0.0786</li></ul>The HDR model exhibits a significant increase in perceptual distance, indicating reduced strict identity reconstruction under deep feature similarity metrics and a shift toward detail-enhancing reconstruction behavior.<hr /><h3 id="structural-energy-(gradient-magnitude)">Structural Energy (Gradient Magnitude)</h3><ul><li>Ground Truth: 404.02 (both models)</li><li>Base Reconstruction: 313.46</li><li>HDR Reconstruction: 687.97</li></ul>The base model demonstrates strong low-pass behavior with reduced high-frequency content. In contrast, the HDR model exhibits high-frequency amplification, exceeding the structural energy of the original inputs.<hr /><h3 id="color-distribution-support">Color Distribution Support</h3><ul><li>Ground Truth: 33150.61 (both models)</li><li>Base Reconstruction: 35004.49</li><li>HDR Reconstruction: 40133.37</li></ul>The HDR model produces a substantially expanded color support space, indicating increased chromatic dispersion and reduced quantization collapse.<hr /><h3 id="photometric-stability">Photometric Stability</h3>Brightness Bias<ul><li>Base: 0.000351</li><li>HDR: 0.0000098</li></ul>Contrast Gain<ul><li>Base: 0.9984</li><li>HDR: 0.99999</li></ul>Both models preserve global photometric consistency, with the HDR variant showing near-perfect affine stability.<hr /><h3 id="channel-drift">Channel Drift</h3><ul><li>Red Shift:<ul><li>Base: +0.0116</li><li>HDR: +0.0104</li></ul></li><li>Green Shift:<ul><li>Base: -0.0606</li><li>HDR: -0.1856</li></ul></li><li>Blue Shift:<ul><li>Base: +0.0187</li><li>HDR: +0.0219</li></ul></li></ul>The HDR model introduces a significantly stronger negative bias in the green channel, while maintaining comparable red and blue stability.<hr /><h2 id="interpretation">Interpretation</h2>The base Qwen VAE behaves as a contractive perceptual projection operator, prioritizing smooth reconstructions and suppression of high-frequency components.The HDR-tuned variant transitions into a detail-amplifying reconstruction operator, characterized by:<ul><li>Increased high-frequency energy</li><li>Expanded color manifold coverage</li><li>Higher perceptual divergence under LPIPS</li><li>Preserved global photometric invariance</li></ul>This represents a functional shift from a smoothing autoencoder regime toward a high-frequency preserving (HDR-like) reconstruction regime.

Files with SHA256: d51d07f00746cc50c487144703fdc75973b79b3dca16b8be530fe11833dbbc45