This Minimax H3 VAE was trained with the following goals:
Broader 2D illustration / anime expressiveness: improved reconstruction across a wide range of 2D illustration and animation styles.
Removal of grid artifacts: eliminates the 16×16 compression seam grid and the 256×256 tiling grid.
Sharper edges: removes blurring and smudging along sharp edges.
Color fixes for anime styles: removes the issue where skin tones turn flushed red (as if drunk) and colors bleed or diverge.
Stable flat shading: reduces subtle color jitter in flat-shaded areas.
Improved single-frame stability: better stability under the T=1 (single-frame) condition.
Training Methodology
Base model: fine-tuned from the official fp16 VAE
Dataset: 15,000 video clips (1/16 – 4 megapixels), 15,000 images (1/16 – 16 megapixels)
Data composition: ~60% 2D illustrations in various art styles, ~20% 3D-rendered clips, ~20% live-action footage
Training steps: 240,000 total steps (30,000 steps × batch size 8)
Loss: fine-tuned with multiple aesthetic evaluation criteria applied as separately weighted loss terms
Training Results
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>1), Seamless VAE (T>1), Original VAE (T=1), and Seamless VAE (T=1).
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.
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.
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.
Shown are the results of a roundtrip using int8convrot at T>1; the left side displays the output from the original VAE, while the right side shows the output from the Seamless VAE.
In the zoomed-in view, you can see that the strange, mottled spots appearing in areas with flat shading have been removed.
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.
License
Followed by minimax-h3-community-license-agreement.