Upscale Image (DiffBIR)

How it fits

Typical backends

Caveats

• Hard constraintGPU is effectively required; the diffusion prior runs a full image generator on the device, and CPU inference is impractically slow.
• FallbackA frame that cannot be decoded, is empty, or exhausts VRAM is passed through unchanged rather than failing the stream; only startup and configuration errors stop the component.
• Parameter interactionoutscale is the final factor relative to the input — the input is pre-enlarged by it before restoration and the result is resampled to exactly that factor. steps, cfg_scale, sampler, and seed trade quality, prompt adherence, trajectory, and reproducibility.
• Parameter interactionmax_input_side caps the longest input side before inference and the stage-1, VAE and control-branch tiling options bound VRAM at a small quality cost. precision selects the compute precision on CUDA.
• Parameter interactionpos_prompt and neg_prompt steer the diffusion prior, with an empty positive prompt running unconditioned. min_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.
• AccuracyThe diffusion prior samples a fresh latent per call, so output is not bit-deterministic unless seed and the input are held fixed.