H3 Streaming VAE Decoder

Blog: https://solarx-team.github.io/H3-Streaming-VAE-Decoder/

Streaming decode for the MiniMax H3 video VAE. Each latent is decoded as soon as it arrives. Supports both the official H3 VAE decoder and the StreamRender-H3 VAE decoder.

Install

pip install -e .            # StreamingDecoder and CausalViTDecoder
pip install -e ".[fused]"   # also the fused kernels

h3_vae also needs a MiniMax-H3 download that contains FL2VA/:

hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "FL2VA/*" --local-dir MiniMax-H3

H3 VAE

Without fused kernels. torch.compile runs on the machine that executes it. --no-compile turns that off. --official also writes the official decode. Saved normalized DiT latents need --normalized.

python -m h3_vae.reconstruct --h3-root /path/to/MiniMax-H3 --video input.mp4 --output stream.mp4
python -m h3_vae.decode_latents --h3-root /path/to/MiniMax-H3 --latents latents.pt --output stream.mp4

from h3_vae import load_vae, StreamingDecoder

vae = load_vae("/path/to/MiniMax-H3", "cuda")
decoder = StreamingDecoder(vae)
for latent in generator:          # [1, C, 1, h, w]
    frames = decoder.push(latent) # [1, 3, n, H, W], n is 1 or 4

With fused kernels. Needs a Blackwell GPU and pip install -e ".[fused]". The kernels compile on that GPU at the first call.

from h3_vae import load_vae, FusedFFNDecoder

vae = load_vae("/path/to/MiniMax-H3", "cuda")
decoder = FusedFFNDecoder(vae)   # push() is the same as above

StreamRender-H3

The student weights are NVIDIA's distilled 24-layer decoder, causal_decoder/causal_vit24_step6000.safetensors in Efficient-Large-Model/StreamRender-H3. That file stays under the MiniMax H3 community license. The rest of the repo is the renderer and is not used here.

hf download Efficient-Large-Model/StreamRender-H3 --include "causal_decoder/causal_vit24_step6000.safetensors" --local-dir StreamRender-H3

Without fused kernels. chunk=1 by default; chunk=2 is the original pair step. latents is normalized, [1, 1, 24, H, W].

import torch
from safetensors.torch import load_file
from streamrender_h3 import CausalViTDecoder, load_config

student = CausalViTDecoder.from_config(load_config(), chunk=1)
checkpoint = load_file("StreamRender-H3/causal_decoder/causal_vit24_step6000.safetensors", device="cpu")
student.load_state_dict(checkpoint, strict=True, assign=True)
student = student.cuda().eval()
state = student.init_state()
with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
    frames = student.decode_chunk(latents, state)

With fused kernels. Same checkpoint and the same normalized latents. H and W are the latent size; 1344×768 is 48×84. Needs Blackwell, FlashAttention-4, and quack.

from safetensors.torch import load_file
from streamrender_h3 import CausalViTDecoder, FusedStreamingCausalViT, load_config

student = CausalViTDecoder.from_config(load_config(), chunk=2)
checkpoint = load_file("StreamRender-H3/causal_decoder/causal_vit24_step6000.safetensors", device="cpu")
student.load_state_dict(checkpoint, strict=True, assign=True)
student = student.cuda().eval()
decoder = FusedStreamingCausalViT(student, H=48, W=84)
frames = decoder.decode_chunk(latents)

This repo is Apache-2.0. The official H3 weights and FL2VA stay under the MiniMax H3 community license. streamrender_h3/decoder follows the public StreamRender-H3 deployment path. streamrender_h3/h3_video_vae keeps the Apache-2.0 headers those modules were published with. The student checkpoint above is not in this repo.

Citation

@misc{chen2026h3streaming,
  title        = {H3 Streaming VAE Decoder: Training-Free Streaming without Future Latents},
  author       = {Chen, Bohong and Chen, Yajing and Du, Yihua and Huang, Junchao and Fang, Guian and Qian, Shengju},
  year         = {2026},
  month        = {oct},
  howpublished = {\url{https://solarx-team.github.io/H3-Streaming-VAE-Decoder/}},
  note         = {SolarX-Team blog post}
}