HunyuanVideo Sr Latents PrePare

The super-resolution branch of HunyuanVideo 1.5 doesn't upscale pixels - it runs a second, distilled diffusion pass that regenerates detail at a higher resolution. And like every diffusion pass, it needs its own starting noise shaped for the target size. HyVideoSrLatentsPrepare is the node that makes that noise: the SR-branch cousin of HyVideo15LatentsPrepare, working out the upscaled latent dimensions and generating fresh latents the SR transformer can refine.

What it does

Given hyvid_cfg (which carries both the base transformer_config and the SR sr_transformer_config - so it knows both the source and target resolutions), the vae (for compression ratios), and an aspect_ratio, it:

1. Takes the low-quality latents from the base pass (transformer_latent, optional input) and maps their resolution up to the SR target using a bucket map from the official pipeline (e.g. 480p → 720p).
2. Computes the SR latent size - note it uses 32 channels for the SR noise, double the base model's channel count, because the SR transformer conditions on more than plain noise.
3. Generates the noise with the CFG node's seed and the flow scheduler's noise scaling.

Outputs are latents_dict, height, width, and n_tokens - the same shape as the base HyVideo15LatentsPrepare, feeding HyVideoSrTransformer.

Where it sits

The SR branch wiring: HyVidelSrTransformerUpsamplerLoader → HyVideoSrLatentsPrepare → HyVideoSrVaeEncode → HyVideoSrTransformer → HyVidelSrVaeDecoder → back into HyVideo15VaeDecode's sr_out. If you're only generating through the simplified HyVideo15I2VSampler / HyVideo15T2VSampler with create_sr_pipeline on, this entire branch is built for you internally - these SR nodes exist for the hand-built workflow.

Installing it

Part of 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. Needs the VAE present and a hyvid_cfg with sr_transformer_config wired from the SR loader's config output - without it, this node has no idea what resolution to target.

Common issues

The aspect_ratio input here has no default and is required, which trips people who expect to inherit the base pass's ratio - set it explicitly to match. The bigger one: because this node reads sr_transformer_config to pick the target resolution, an unwired config produces a confusing failure or a silently wrong output size. And worth knowing for VRAM planning: this node allocates the SR noise at 32 channels at the upscaled resolution, so on a 12GB card the SR branch's memory peak can approach the base pass's. If you're tight, that's a legitimate reason to run without SR and upscale in post instead.