Some Questions about Inference Time and Prompts

[HaojunChen663]

Excellent work!

• I was wondering about the inference time of your super-resolution model, and whether you provide scripts for batch parallel inference as well as any tips for speeding up inference.
• Additionally, I'd like to ask if setting a prompt for image super-resolution yields better results compared to not using one.

Looking forward to your reply!

[Feynman1999]

Excellent work!

• I was wondering about the inference time of your super-resolution model, and whether you provide scripts for batch parallel inference as well as any tips for speeding up inference.
• Additionally, I'd like to ask if setting a prompt for image super-resolution yields better results compared to not using one.

Looking forward to your reply!

Thank you for your interest in our work.

1. The model in our paper follows the paradigm of mainstream papers, i.e., training and inference on 512×512 resolution. For images at 512×512 resolution, inference on an L20 machine takes less than 1 second (you may need to test it yourself for the exact figure).
   For higher resolutions, the current script employs a tiled processing approach, and the speed can be estimated accordingly by scaling.

2. Regarding acceleration methods, I do not believe batch parallelism would be effective, as a single sample already achieves very high GPU utilization with a large 20B-parameter model like Qwen-Image. Possible acceleration options include:
   ① Leveraging NVIDIA's FP8 TensorRT acceleration (offering roughly 1.x× speedup compared to BF16, though it requires compatible hardware such as H20 architecture or above);
   ② Employing lower-bit attention kernels, such as SageAttention (however, at 512×512 resolution, the speedup may not be substantial since attention does not dominate the computation).

3. As for the prompt component, this represents one of the primary innovations in our paper. Previous one-step super-resolution methods have explored prompts only to a limited extent, largely because under conventional single-step RGB MSE loss + LPIPS loss, the model's responsiveness to prompts remains constrained. We propose a dynamic fidelity weight adjustment mechanism (Fidelity Weight $f$), which encourages stronger prompt adherence when $f$ is small.