Hunyuan Video 1.5 Super Resolution

Hunyuan Video 1.5's super-resolution is a two-stage trick, and this node is the bridge between the stages. You generate a short, low-res clip, upscale its latent, and then let the model re-denoise the bigger version to add real detail - not just blurry interpolation. HunyuanVideo15SuperResolution is the conditioning node that packages everything the second stage needs: the upscaled latent, the text conditioning, and optionally a first-frame image and CLIP vision signal, all glued into a form the 1.5 super-res model understands.

If you've only ever used the single-pass Hunyuan 1.5 text-to-video, this is the "latent upscale + refiner" pattern done the official way. Think of it as the video-model version of what SD upscalers do for images: hallucinate detail at a higher resolution instead of just resampling pixels.

How it works

The mechanism is worth understanding because it explains why the node looks the way it does. Hunyuan 1.5's refiner doesn't take your latent directly - it wants a concatenated latent with extra channels. The node builds that: it takes your upscaled latent, pads it into a wider-channel latent, and stores the low-res content in the upper channels. If you provide a start_image, it upscales that image to match and VAE-encodes it into the first channels too (which is why the VAE input is only required when start_image is connected).

Then it stamps two things onto both positive and negative conditioning: the concat latent (concat_latent_image) and your noise_augmentation value. That number (default 0.7, range 0–1) controls how much fresh noise the refiner starts from - higher means more freedom to invent detail and diverge from the source, lower means a more faithful-but-soft result. An optional clip_vision_output gets attached as well, which lets the second stage reference visual context beyond the latent alone.

The inputs that matter

• latent - your upscaled latent from the first stage (output of the base model's sampling, run through a latent upscaler like the Hunyuan 1.5 latent upscale model).
• positive / negative - text conditioning from the Hunyuan 1.5 text encoder.
• noise_augmentation - detail-vs-fidelity dial (default 0.7). Lower it toward 0.3–0.4 if the refiner is drifting too far from the source.
• vae + start_image - optional; needed together if you want to lock the first frame.
• clip_vision_output - optional visual context.

Outputs: the packed positive and negative conditioning, plus the latent passed through unchanged. All three feed the second-stage sampler.

Where it fits

It ships with ComfyUI core; the underlying Hunyuan 1.5 super-resolution model and its latent upscaler go in your ComfyUI model folders. Honestly? Hunyuan Video's local usage has shrunk a lot as Wan took over, and the super-res path is extra VRAM on top of an already-heavy model. Reach for it when you actually want 720p+ out of Hunyuan 1.5 rather than a low-res draft - it's the official way to get that, and it's not particularly fiddly once the two-stage flow is in place.

Common issues

The classic trip-up is connecting start_image without a VAE - the node will error, because it has to encode the image. Also, don't skip the latent upscale step before this node and expect magic: this node packages an upscaled latent, it doesn't upscale one for you. And if your output looks mushy, that's usually noise_augmentation too high (invented detail that doesn't fit) or too low (nothing new added) - it's the parameter you'll actually tune here.