<h1 id="qwen-image-vae">Qwen Image VAE</h1><p></p><ul><li><p>Full FP32 Training of Decoder</p></li><li><p>Works in ComfyUI</p></li></ul><p></p><hr /><p></p><h2 id="overview">Overview</h2><p>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.</p><p>The evaluation compares the base and HDR-tuned models using perceptual, structural, distributional, and photometric metrics over identical input data.</p><hr /><h2 id="evaluation-summary">Evaluation Summary</h2><h3 id="perceptual-fidelity-(lpips)">Perceptual Fidelity (LPIPS)</h3><ul><li><p>Base: 0.0177</p></li><li><p>HDR: 0.0786</p></li></ul><p>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.</p><hr /><h3 id="structural-energy-(gradient-magnitude)">Structural Energy (Gradient Magnitude)</h3><ul><li><p>Ground Truth: 404.02 (both models)</p></li><li><p>Base Reconstruction: 313.46</p></li><li><p>HDR Reconstruction: 687.97</p></li></ul><p>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.</p><hr /><h3 id="color-distribution-support">Color Distribution Support</h3><ul><li><p>Ground Truth: 33150.61 (both models)</p></li><li><p>Base Reconstruction: 35004.49</p></li><li><p>HDR Reconstruction: 40133.37</p></li></ul><p>The HDR model produces a substantially expanded color support space, indicating increased chromatic dispersion and reduced quantization collapse.</p><hr /><h3 id="photometric-stability">Photometric Stability</h3><p>Brightness Bias</p><ul><li><p>Base: 0.000351</p></li><li><p>HDR: 0.0000098</p></li></ul><p>Contrast Gain</p><ul><li><p>Base: 0.9984</p></li><li><p>HDR: 0.99999</p></li></ul><p>Both models preserve global photometric consistency, with the HDR variant showing near-perfect affine stability.</p><hr /><h3 id="channel-drift">Channel Drift</h3><ul><li><p>Red Shift:</p><ul><li><p>Base: +0.0116</p></li><li><p>HDR: +0.0104</p></li></ul></li><li><p>Green Shift:</p><ul><li><p>Base: -0.0606</p></li><li><p>HDR: -0.1856</p></li></ul></li><li><p>Blue Shift:</p><ul><li><p>Base: +0.0187</p></li><li><p>HDR: +0.0219</p></li></ul></li></ul><p>The HDR model introduces a significantly stronger negative bias in the green channel, while maintaining comparable red and blue stability.</p><hr /><h2 id="interpretation">Interpretation</h2><p>The base Qwen VAE behaves as a <strong>contractive perceptual projection operator</strong>, prioritizing smooth reconstructions and suppression of high-frequency components.</p><p>The HDR-tuned variant transitions into a <strong>detail-amplifying reconstruction operator</strong>, characterized by:</p><ul><li><p>Increased high-frequency energy</p></li><li><p>Expanded color manifold coverage</p></li><li><p>Higher perceptual divergence under LPIPS</p></li><li><p>Preserved global photometric invariance</p></li></ul><p>This represents a functional shift from a smoothing autoencoder regime toward a high-frequency preserving (HDR-like) reconstruction regime.</p>

Files with SHA256: 5129076078ffa2e2766ce65e857b12b8619e5a739f97c17bd032ce0fb83489a7