Upscale Image (Real ESRGAN)

How it fits

Typical backends

Caveats

• Hard constraintmodel fixes the backbone and native scale; the filename must be one of the known checkpoints or loading fails at startup before any inference.
• Parameter interactionoutscale is the final factor relative to the input — the network runs at its native scale and the result is resampled to outscale, so an x4 checkpoint with outscale 2 yields a 2x output.
• Parameter interactionmax_input_side caps the longest input side before inference to bound VRAM, and tile_size (with tile_pad) tiles large frames; both trade a little quality for memory headroom. half enables 16-bit inference on CUDA.
• Parameter interactionmin_width and min_height gate the skip jointly — upscaling is skipped only when every non-zero floor is met — while max_width and max_height clamp the output size.
• LatencyA GPU device is effectively required for live use; half lowers VRAM and speeds CUDA inference.
• I/O contractThe checkpoint is supplied through the bound model file slot selected by model; it loads from the staged artifact at startup with no runtime download.