HunyuanVideo Sr Transformer

HunyuanVideo 1.5's super-resolution isn't a filter you slap on at the end. It's a distilled diffusion pass - a second transformer trained specifically to take your low-res generation and regenerate it with real detail at 720p or 1080p - and HyVideoSrTransformer is the node that runs its denoising loop. It's the SR-branch mirror of HyVideo15Transformer, and it's notably faster because it's designed for very few steps: the default is just 8.

What it needs

• transformer - the SR diffusion transformer from HyVidelSrTransformerUpsamplerLoader (that loader's model output).
• hyvid_cfg - from HyVideo15CFG; it supplies the scheduler, seed, and guidance setting. The SR branch uses the hyvid_sr_cfg output on the CFG node in the example workflows.
• vae_concat - from HyVideoSrVaeEncode, which packages the upsampled low-quality latents, the noise-strengthened condition, and the reference-image condition the SR transformer consumes.
• hyvid_embeds / vision_states / extra_kwargs - the same text, vision, and glyph conditioning the base pass used, reused here.
• lantens_dict - yes, that's the actual input name, a typo in the pack's schema for "latents_dict." It's the SR noise from HyVideoSrLatentsPrepare. This is a known quirk - wire the latents_dict output from HyVideoSrLatentsPrepare into the lantens_dict port and don't let the spelling throw you.
• n_tokens and steps (8) - the token count from HyVideoSrLatentsPrepare and your SR step count. 8 is the distilled default and the sweet spot.
• target_dtype (bfloat16) - match the transformer.
• embedded_guidance_scale (None) - extra guidance for models that support it; leave off. guidance_rescale (0) and eta (0) work like they do on the base transformer.

Output latent (HYVID15TRANSFORMERLATENT) feeds HyVidelSrVaeDecoder, which decodes it and hands the frames to HyVideo15VaeDecode's sr_out.

How it works

Same loop as the base pass - timesteps from the flow scheduler, CFG when guidance > 1, a transformer forward pass each step, scheduler step to advance - but with a key trick in the conditioning: early in the loop the SR transformer sees the full noise-augmented low-quality latents as its guide, and once the timestep drops below a threshold it switches to a "zero condition," trusting the model to have already locked the structure. That's how a few-step pass can pull off a resolution jump without the whole thing falling apart.

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. The SR transformer and upsampler download through HyVidelSrTransformerUpsamplerLoader (paths "None" auto-fetch from tencent/HunyuanVideo-1.5).

Common issues

The typo'd lantens_dict input is the #1 stumbling block - it looks like a missing node, not a misspelling, and it's easy to chase a wiring ghost. The second is step count: cranking steps up "for quality" on this distilled pass doesn't help much and makes the SR branch nearly as slow as the base pass - the whole point is that 8 steps is right. And VRAM: this transformer runs alongside the base one, so the two-pass complete workflow wants real headroom; if it OOMs, the move is dropping the SR branch entirely (disable create_sr_pipeline in the simplified samplers) and upscaling in post. That's the trade, and it's a legitimate one - the SR pass is 1.5's headline feature, but it's an optional one.