HyVidelSrVaeDecoder

HyVidelSrVaeDecoder is a small, single-purpose node at the end of the super-resolution branch: it takes the SR latents that HyVideoSrTransformer produced and runs them through the VAE decoder to get actual video frames. If you're building the complete HunyuanVideo 1.5 workflow and you want the 720p/1080p SR output, this is the node that closes the loop.

What it does

It's basically the decode half of HyVideo15VaeDecode, specialized for the SR path. Inputs are the latents (HYVID15TRANSFORMERLATENT) coming out of HyVideoSrTransformer, the vae from HyVideo15VaeLoader, the hyvid_cfg from HyVideo15CFG (used for the RNG), plus two settings: output_type (pt for pixel frames, latent to get the tensor back raw) and vae_dtype, defaulting to float16. The decode runs with tiling enabled and the VAE offloaded, which is the pack's standard approach to keeping decode memory in check on consumer GPUs.

The output is sr_out (HYVID15SROUT). Here's the slightly confusing part: it does not hand you a preview-ready IMAGE directly. It produces the SR frames as a packaged result that you feed into HyVideo15VaeDecode's optional sr_out input - that node then merges the base-pass decode and the SR decode and returns the final IMAGE. So the wiring is: HyVideoSrTransformer → HyVidelSrVaeDecoder → HyVideo15VaeDecode (sr_out). Easy to get backwards, and the first time you hook this up you will.

Why it exists instead of one decode node

In the complete workflow the SR branch is deliberately kept separate so you can run the base pass and the SR pass with different settings, step counts, or even different VAEs. The simplified samplers (HyVideo15I2VSampler, HyVideo15T2VSampler) do all of this internally whenever create_sr_pipeline is on - including the merged decode - so most users will never see HyVidelSrVaeDecoder. It's the "I'm building it by hand" path.

Installing it

It ships in comfyui_hunyuanvideo_1.5_plugin. ComfyUI Manager → "HunyuanVideo-1.5 nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yuanyuan-spec/comfyui_hunyuanvideo_1.5_plugin
cd comfyui_hunyuanvideo_1.5_plugin
pip install -r requirements.txt

Restart, and have the VAE present (HyVideo15VaeLoader with path "None" fetches it automatically).

Common issues

The one real trap is expecting a finished image out of this node. Remember the output is sr_out, not IMAGE - wire it into HyVideo15VaeDecode and read the frames there. The other practical gotcha is VRAM: this decode runs at the upscaled resolution, so a 480p→1080p SR decode is a heavier VAE operation than the base decode. vae_dtype defaults to float16 precisely so it fits; if you're crashing on decode with plenty of headroom, bumping it to bfloat16 or float32 is a legitimate experiment, but expect the memory use to climb with it.