HyVidelSrTransformerUpsamplerLoader

HunyuanVideo 1.5 doesn't just upscale video with a generic model at the end - it runs a whole second, distilled diffusion pass whose job is turning a 480p or 720p generation into 720p or 1080p. HyVidelSrTransformerUpsamplerLoader is the node that loads both halves of that pass at once: the SR diffusion transformer and the upsampler network. It's one of the "HyVidel" nodes (the pack's slightly inconsistent naming for the SR branch), and it's only relevant if you're assembling the complete workflow by hand instead of letting the simplified sampler build the SR pipeline for you.

What it loads

Two models, one call:

• The SR transformer - a HunyuanVideo_1_5_DiffusionTransformer trained specifically for the few-step super-resolution pass, loaded from models/diffusion_models/hyvideo15/<resolution>_sr_distilled.
• The upsampler - the SRTo720pUpsampler or SRTo1080pUpsampler network from models/upscale_models/hyvideo15, which does the actual resolution bump of the latents before the SR transformer refines them.

The three settings are resolution (480p / 720p / 1080p) and task_type (t2v / i2v), which together pick which distilled SR variant to load (e.g. 720p_sr_distilled for the i2v 720p pass), plus transformer_dtype, defaulting to bfloat16. Both path dropdowns default to None, which triggers auto-download from tencent/HunyuanVideo-1.5 (upsampler/* and the matching transformer/*_sr_distilled/* folders). Note the resolution here means the target of the SR pass: pick 720p to go from 480p up to 720p.

It outputs three things: model (the SR transformer), config (its config object, which the CFG node can consume for sr_transformer_config), and upsampler.

Where it sits in the workflow

This is the SR branch of the complete workflow: load this, then HyVideoSrLatentsPrepare builds the upscaled noise latents, HyVideoSrVaeEncode conditions them on the low-quality latents from the base pass, HyVideoSrTransformer denoises them (default just 8 steps - it's distilled), and HyVidelSrVaeDecoder decodes the result back into frames, which then merges into HyVideo15VaeDecode via its sr_out input. If all you want is the finished clip, the simplified HyVideo15I2VSampler / HyVideo15T2VSampler build this whole pipeline internally whenever create_sr_pipeline is on, so you'll never touch any of these nodes.

Installing it

Same pack, same story: ComfyUI Manager → search "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 ComfyUI afterwards. The SR models are part of the tencent/HunyuanVideo-1.5 download, so a successful auto-download of the main transformer and VAE generally means these are nearby too.

Gotchas

Watch the transformer_path vs upsampler_path distinction when you point them at real folders - the transformer lives under diffusion_models and the upsampler under upscale_models, and they're separate dropdowns for a reason. Also, the SR pass is genuinely heavy: you're loading a second transformer alongside the base one, which is why the simplified sampler only spins it up when you ask. If you're on a 12–16GB card and VRAM pressure is real, disabling create_sr_pipeline in the simplified sampler (or skipping this branch in the complete workflow) and upscaling in post is a legitimate call. The pack's own README treats the SR pass as the differentiator for 1.5, and it is - but it's an optional differentiator.