Introduction

Two of the best speed tricks for MiniMax H3 just moved from custom workflows into ComfyUI core. On October 28, the ComfyUI team merged native support for Sol-H3-VAE, NVIDIA's distilled fast video decoder for H3, and for the community MiniMax H3 latent upscaler, complete with a new MiniMax H3 Latent Upscale With Model node.

Both are open weights on Hugging Face, both run locally, and you no longer need the Sol-H3 Studio bundle or extra node packs to use them. Update ComfyUI to the latest master (they landed after the v0.43.0 release), drop the files in the right folders, and wire them into any H3 graph you already have.

What shipped

  • Sol-H3-VAE support (PR #17411): the stock Load VAE node now recognizes Efficient-Large-Model/Sol-H3-VAE. It loads as released (decoder only) or merged with the H3 encoder into a single VAE file.
  • H3 latent upscaler support (PR #17412): Load Latent Upscale Model now detects minimax_h3_latent_upscaler_3d_conv_v1 from LBH-123-AI/Minimax_h3_latent_Upscaler.
  • New node: MiniMax H3 Latent Upscale With Model resizes the video part of an H3 latent to a target width and height (defaults to 2560×1440). Audio passes through untouched.
  • Both PRs were written by kijai, the developer behind many of the community's favorite video wrappers, and merged by comfyanonymous.

The decode speedup

Sol-H3-VAE swaps H3's 36-layer decoder for a 24-layer, 1.62B-parameter decoder distilled from it. Instead of chopping frames into 256px tiles and overlapping time chunks, it decodes the whole clip at once with windowed attention.

Here's what the ComfyUI PR measured for a 1344×768, 124-frame clip on an RTX 5090:

  • Stock H3 VAE, fp16: 10.7 s, 5.5 GiB peak VRAM
  • Stock H3 VAE, int8: 4.61 s, 3.2 GiB peak VRAM
  • Sol-H3-VAE, fp16: 3.28 s, 5.9 GiB peak VRAM
  • Sol-H3-VAE, int8: 1.70 s, 4.3 GiB peak VRAM

So you get roughly a 3× faster decode at the same precision, and over 6× if you compare int8 Sol to stock fp16.

The honest caveat: this is a different decoder, not a lossless shortcut. The PR reports 39 dB PSNR against the stock decoder on a generated 768p clip (NVIDIA's model card reports 38.2 dB). It looks the same to the eye, and the differences show up in fine texture.

Memory notes before you switch

  • No tiling. Sol-H3-VAE can't tile, so VAEDecodeTiled will decode the whole clip in one go.
  • Activations grow with length. About 4.3 KB per latent token: roughly 3.5 GiB of activations at 362 frames, and 5.4 GiB at 2048×1152 for 124 frames.
  • No streaming. The decoder is noncausal, so it works on whole clips, not latent by latent.

If you're tight on VRAM or rendering very long clips, keep the stock decoder handy as a fallback.

The latent upscaler, now first-class

The H3 latent upscaler (Apache 2.0) has been a community favorite since August, with about 290K downloads in the last month alone. The trick: generate small, upscale the latent directly, then refine at full size. That skips H3's heavy decode, pixel upscale, and re-encode round trip, and it avoids the ghosting that plain bilinear latent resizing causes.

The native version runs its GroupNorm and SiLU through ComfyUI's fused kernels and streams the clip in 16-frame slabs with clip-wide statistics. kijai notes that the result is bitwise identical to a single pass, so long clips don't need a giant memory spike.

Quick setup

Grab the two files:

cd ComfyUI && git pull
hf download Efficient-Large-Model/Sol-H3-VAE sol_h3_vae.safetensors --local-dir models/vae
hf download LBH-123-AI/Minimax_h3_latent_Upscaler minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors --local-dir models/latent_upscale_models

Then, in your existing H3 workflow:

  • Point Load VAE at sol_h3_vae.safetensors for the decode. If you also encode images in the same graph, keep the stock H3 VAE for encoding or use a merged encoder-plus-Sol file.
  • For a two-stage render, add Load Latent Upscale Model, then MiniMax H3 Latent Upscale With Model between your low-res sampler and your refine sampler. Set the target width and height, then re-noise and refine as usual.
  • The upscaler's Hugging Face repo ships example i2v and r2v workflow JSONs if you want a starting graph.

License check

The Sol-H3-VAE code is Apache 2.0, but the weights are distilled from the MiniMax H3 VAE and fall under the MiniMax-H3 license. Check that license's territory terms before you use outputs commercially. The latent upscaler is Apache 2.0.

Original Source

Conclusion

This is the kind of update that makes H3 feel lighter without changing a single prompt. Faster decodes mean faster previews, and a native latent upscaler makes the "draft small, finish big" workflow a two-node change. If you've been putting off a ComfyUI update this week, this is your sign. Pull, swap your VAE, and see how many extra iterations you get back tonight.

Want to try H3 without touching a graph? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-native-sol-h3-vae-h3-latent-upscaler-faster-decodes

—Aurelia ♡