Sol-H3-VAE, a Fast Decoder for MiniMax-H3, Adds Support in ComfyUI 0.44.0

This is based on the Sol-H3-VAE model card on Hugging Face and the ComfyUI v0.44.0 release notes.

Sol-H3-VAE (model card)

• It is a replacement that distills H3's VAE decoder (36 layers) into a non-causal decoder with 24 layers and 1.62B parameters. It takes the same normalized latents and produces the same frames, and the audio decoder is unchanged. (official)
• 5-second clip decoding time, from the existing H3 VAE to Sol: RTX 5090 at 768p went from 12.5 seconds to 5.0 seconds, GB200 at 720p from 6.16 seconds to 0.34 seconds, and GB200 at 2K from 14.5 seconds to 0.89 seconds. The GB200 values use CUDA graphs, and the 5090 values use eager mode. (official, production team's own measurements)
• Quality: compared with the existing decoder, PSNR is 45.4dB (held-out latents) and 38.2dB on generated 768p clips. It is an approximate replacement, not an identical one. (official)
• Memory: with CUDA graphs, a 5-second clip needs about 24GiB at 720p and about 48GiB at 2K. The whole clip is decoded at once; streaming latents through one at a time is not supported. (official)
• The default fast engine requires Triton, FlashAttention-4, and QuACK. engine="ref" runs on plain PyTorch. On the 5090, the first call takes about 25 seconds for kernel compilation. (official)
• Known bug: On SM120 GPUs (RTX 5090, DGX Spark), FlashAttention-4 4.0.0b19 through b22 can cause memory access errors due to a varlen bug. The guidance is to add a guard or use the ref engine. (official)
• License: The code is Apache-2.0, and the weights follow the MiniMax-H3 license. (official)

ComfyUI v0.44.0 (November 4; the release page gives no year)

• Sol-H3-VAE, the MiniMax-H3 latent upscaler, and H3 context window support were added, and H3 Max partner nodes (Recast, Extend, Reframe, Insert, Lip Sync, Camera Control) were added. (official)
• Video saving gained ProRes output and chroma subsampling options. A node for setting the VAE compute dtype was added, and a conflict that occurred when using that option with the LTX 2.5 VAE decoder was fixed. (official)
• The FlashVSR and image upscale nodes were moved under Comfy Router (under the WaveSpeed partner nodes). (official)

Not verified: independent measurements (all times above are the production team's figures), a visual quality comparison of generated output, and the release date of the model card.