This Minimax H3 VAE was trained with the following goals:<ol><li>Broader 2D illustration / anime expressiveness: improved reconstruction across a wide range of 2D illustration and animation styles.</li><li>Removal of grid artifacts: eliminates the 16×16 compression seam grid and the 256×256 tiling grid.</li><li>Sharper edges: removes blurring and smudging along sharp edges.</li><li>Color fixes for anime styles: removes the issue where skin tones turn flushed red (as if drunk) and colors bleed or diverge.</li><li>Stable flat shading: reduces subtle color jitter in flat-shaded areas.</li><li>Improved single-frame stability: better stability under the T=1 (single-frame) condition.</li></ol>Training Methodology<ul><li>Base model: fine-tuned from the official fp16 VAE</li><li>Dataset: 15,000 video clips (1/16 – 4 megapixels), 15,000 images (1/16 – 16 megapixels)</li><li>Data composition: ~60% 2D illustrations in various art styles, ~20% 3D-rendered clips, ~20% live-action footage</li><li>Training steps: 240,000 total steps (30,000 steps × batch size 8)</li><li>Loss: fine-tuned with multiple aesthetic evaluation criteria applied as separately weighted loss terms</li></ul>Training Results<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/59ee3052-99f3-4f45-930e-ba0dbab28adc/original=true/59ee3052-99f3-4f45-930e-ba0dbab28adc.jpeg" />The image on the far left is the original source. The subsequent images show the difference (amplified 8x) between the original source and the decoded outputs from the following configurations: Original VAE (T&gt;1), Seamless VAE (T&gt;1), Original VAE (T=1), and Seamless VAE (T=1).<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d9fc9ea3-efd4-443d-bd13-e39dde18c041/original=true/d9fc9ea3-efd4-443d-bd13-e39dde18c041.jpeg" />Shown are the results of an encoding-decoding roundtrip in fp16 at T=1; the left side displays the output from the original VAE, while the right side shows the output from the Seamless VAE.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ea33095a-55ca-4ade-89f1-fb566506c503/original=true/ea33095a-55ca-4ade-89f1-fb566506c503.jpeg" />Looking at the enlarged image, you can see that the messy artifacts—typical of quantization loss—have been cleared up, and the skin tone has shifted from a reddish hue back to that of the original image.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0a3e4ffc-209a-4ff7-ac37-11e0117274fa/original=true/0a3e4ffc-209a-4ff7-ac37-11e0117274fa.jpeg" />This image is the result of decoding somewhat noisy latent data generated by a DiT, rather than data produced by a clean encoder. As you can see, while many of the seams have been eliminated, they have not been removed entirely.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/7e2b5800-3da9-4858-8926-1c6002c523fb/original=true/7e2b5800-3da9-4858-8926-1c6002c523fb.jpeg" />Shown are the results of a roundtrip using int8convrot at T&gt;1; the left side displays the output from the original VAE, while the right side shows the output from the Seamless VAE.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a3be7afe-b771-43d4-8c65-f5b9c89f9290/original=true/a3be7afe-b771-43d4-8c65-f5b9c89f9290.jpeg" />In the zoomed-in view, you can see that the strange, mottled spots appearing in areas with flat shading have been removed.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2bc4fe55-fc56-4963-b420-3e3bcc84f278/original=true/2bc4fe55-fc56-4963-b420-3e3bcc84f278.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/599e5e97-4929-40e6-a6de-54090ac45d7c/original=true/599e5e97-4929-40e6-a6de-54090ac45d7c.jpeg" />This image is the result of a round-trip process applied to a film grain source overlaid with pure noise. While the sharp, high-frequency characteristics of the noise have been intensified, the randomness of the noise has been lost; its properties have shifted such that the same noise pattern repeats every 16x16 pixels.LicenseFollowed by <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE">minimax-h3-community-license-agreement</a>.

Files with SHA256: 422f14074b150f8aeaf68861abd3de35533f52d2116f7cedd66878bd7053f078