ComfyUI: 1080P Latent Upscaling with Hunyuan SR Model

Step-by-step

1. Set the main model shift value to 2.0.
2. Add the Hunyuan Video Latent Upscale with Model node.
3. Connect the original positive and negative conditions into the upscale node to generate corrected conditions.
4. Configure the first sampler with 'add noise' set to 'enable' for the first 4 steps.
5. Configure the second sampler with 'add noise' set to 'disable' and set the start step to 4.
6. Connect the corrected conditions to the first sampler and the original conditions to the second sampler.
7. Decode the final latent output to view the upscaled video.

Watch out for

• The shift value of 2.0 is fixed regardless of whether you are doing text-to-video or image-to-video.
• The final upscaled effect will not be identical to the original video; expect about 90% similarity.

Tips

• Use the SR (Super Resolution) model specifically for the sampling stage.
• For this distilled model, 8 sampling steps are sufficient.
• Think of this dual-sampler structure as high and low noise sampling, similar to Wan 2.2 workflows.

Highlights

“I also optimized it... 8 sampling steps are sufficient.”