Cross-scene VAE comparison

Compared with qwen_image_vae, liquid3844 consistently produced slightly stronger luminance contrast, saturation, color separation, and edge energy across five controlled test scenes.

Average changes across the five tests:

• Luminance contrast: +1.24%

• Mean saturation: +2.53%

• Colorfulness: +3.01%

• Mean Sobel edge strength: +2.03%

• Sobel RMS: +3.19%

Structural differences remained small, with SSIM ranging from 0.9755 to 0.9957.

liquid3844 also shows a measurable tendency to push tonal extremes further. In the high-key test, pure-white pixels increased from 4.766% to 15.418%. In the dark test, near-black pixels increased from 24.804% to 29.917%.

This suggests that the fine-tuned VAE increases perceived contrast and edge/color separation, but may also increase highlight clipping and shadow compression in extreme scenes.

One realistic skin-detail test showed lower Laplacian variance (−13.03%) despite higher Sobel edge energy, suggesting that very fine high-frequency texture may be smoothed while broader edges become stronger.

Designed for Anima 3.8B, compatible beyond it

liquid3844 was primarily developed for Anima 3.8B, but it is not limited to that model. Since the encoder was kept frozen and only the decoder tail was fine-tuned, the original latent space was preserved.

It can also be used with Anima 2B, Anima 2.9B, Anima 3.8B, and KREA2 workflows that use the same compatible VAE latent format.

The tuning focuses on the final image decode rather than changing the latent representation, with emphasis on:
edge clarity, contrast, saturation, color separation, microtexture, and darker tonal definition.

For best results, compare liquid3844 directly with qwen_image_vae using the same latent, prompt, seed, and sampler settings.