HunyuanVideo Decode

This is the last step of the pipeline: it takes the LATENT samples HyVideoSampler produced and runs them through HunyuanVideo's VAE to get actual IMAGE frames back out. Straightforward in concept, but the tiling settings here are where people get tripped up, because this VAE isn't just spatial (compressing width/height) - it's a 3D causal VAE, meaning it compresses across time as well, and the tiling options reflect that.

The inputs that matter

• vae - from HyVideoVAELoader.
• samples - the LATENT from HyVideoSampler.
• enable_vae_tiling (default on) - decodes in tiles instead of all at once, drastically reducing memory use at the cost of potentially visible seams where tiles meet.
• temporal_tiling_sample_size (default 64) - the tooltip is unusually direct here: "model default is 64, any other value will cause stutter." That's not a suggestion, it's a stated fact from the author. Leave it at 64 unless you're specifically troubleshooting VRAM and are prepared to accept the stutter as a tradeoff.
• spatial_tile_sample_min_size (default 256, steps of 32) - the spatial tile size in pixels. Smaller uses less VRAM but risks more visible seams.
• auto_tile_size (default on) - picks sensible tile sizes automatically and ignores the two settings above when enabled. Leave this on unless you're deliberately hand-tuning for a specific VRAM budget.

Optional: skip_latents (default 0) - skips a number of latents from the start of the sequence, and the tooltip notes it "can help with flashing," a real artifact some users hit at the very start of a clip. balance_brightness (default off) - attempts to even out brightness across the decoded frames, useful if you're seeing a flicker or drift in exposure across the clip.

Output: images, ready for a video-combine node or straight to disk.

How to install it

Ships with the pack - no separate download for this node.

• ComfyUI Manager - search ComfyUI-HunyuanVideoWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, then pip install -r ComfyUI-HunyuanVideoWrapper/requirements.txt, restart.

You need HyVideoVAELoader working first - the VAE weights are a manual download from Kijai's HunyuanVideo_comfy repo on HuggingFace, no auto-download.

Common issues & troubleshooting

Output stutters or has odd frame pacing. Check temporal_tiling_sample_size is still at 64 (or that auto_tile_size is on). The author's own tooltip states plainly that any other value causes stutter - this is the first thing to check if playback looks wrong.

Visible seams in the decoded frames. Turn off auto_tile_size and raise spatial_tile_sample_min_size if you have VRAM headroom, since larger tiles mean fewer seams. If you're VRAM-constrained and can't raise tile size, some seaming is the accepted cost of enable_vae_tiling.

Flashing or a jarring flicker right at the start of the clip. Try a small nonzero skip_latents - the author added this specifically to help with that symptom.

Out of memory during decode, even though sampling finished fine. Decode is a separate memory spike from sampling. Make sure enable_vae_tiling is actually on (it is by default, but double-check nothing turned it off) and let auto_tile_size pick conservative tiles rather than hand-setting a large spatial_tile_sample_min_size.