Minimax H3 Latent Upscaler has been released in BF16 format

Developers presented Minimax-H3-Latent-Upscaler-BF16-MAXQUALITY — an optimized version of the image upscaling model. The release is aimed at eliminating visual artifacts that occur during standard processing by using the method of conservative mixed quantization.

The model was quantized from the original FP32 format to BF16. For the correct operation of the tool in the ComfyUI environment, users need to set the precision=bf16 and device=cuda settings. It is also recommended to use the tile_size=256 and temporal_overlap parameters in the range of 2 to 4 to achieve maximum generation quality.

During testing, the model received a rating of 12 stars based on the results of a two-hour run. The use of BF16 reduces the requirements for video memory while maintaining the accuracy of calculations, which is critical for high-resolution upscaling tasks. This approach minimizes the loss of details characteristic of aggressive weight compression.

The tool is designed for integration into existing image generation pipelines. The update solves the problem of quality degradation when scaling, providing more stable operation on consumer NVIDIA graphics processors.