Edit Image (HuggingFace)

How it fits

Typical backends

Caveats

• I/O contractThe worker maps one input Image to exactly one output Image at the same resolution; the diffusion processor rounds the working size to the VAE scale factor internally and the result is resized back to the input dimensions before emit.
• Parameter interactionstrength and num_inference_steps interact: the pipeline runs `int(num_inference_steps * strength)` denoising steps, and a turbo checkpoint requires that product be at least 1. The defaults (steps 2, strength 0.5) clear it; lowering both at once can drop the effective step count to 0.
• Parameter interactionTurbo checkpoints are trained without classifier-free guidance, so guidance_scale should stay 0.0 and the negative instruction has no effect; a non-turbo checkpoint wants guidance_scale ~7.5 and honours the negative instruction.
• Latencymodel is hot-swappable at runtime and triggers a pipeline reload (and a GPU memory reclaim on CUDA) on the next call, producing a one-call latency spike.
• FallbackEvery input emits exactly one Image. An empty or degenerate (sub-16-pixel-side) frame and an empty prompt both pass the input frame through unchanged without running diffusion, and a diffusion fault is logged and degraded to the input frame passed through; there is no 1-to-0 drop. device is read once at startup and a value starting with `cuda` silently falls back to CPU when CUDA is unavailable.