
Edits each input Image under a text prompt with a HuggingFace diffusers image-to-image checkpoint (default `stabilityai/sdxl-turbo`) and emits one edited Image at the same resolution. strength sets how far the edit departs from the source.
Restyle a live webcam frame with a prompt and stream the edited result to a browser.
Apply a prompted edit to an uploaded image file and return the result over HTTP.
model, the iteration count and guidance scale actually used, the working width and height, precision, device, the random initialiser (null when unset, with digests of the global CPU and CUDA random-number states) and options holding strength. Text inputs are logged only as the SHA-256 and codepoint length of the exact text passed to the model (the prompt and negative_prompt; an unset negative prompt has null digest and length); the input picture is logged as a digest of its decoded RGB pixels with its shape. The sources entry holds that pixel digest and, when the position input is wired, the position of the received picture, else null. Timings are synchronized seconds: startup, model_load, total, encode, denoise and decode; stage and load times are subsets of the total, not additive. Memory holds the GPU allocator peaks, the GPU name and a host RSS peak sampled every 10 ms; an unavailable probe is null. The result holds the status (ready, generated, passthrough or failed), the error class name only, output_count, generated_count, output_bytes and one entry per output with worker_id, job_index, sequence, pixels_sha256, position (the PictureIdentity emitted with the picture), generated, width, height and bytes. A passthrough output counts no generated picture. The attempt key is worker_id with job_index and an output adds its sequence; digests and positions can repeat across identical jobs, so a sink keeps its accept or reject decision against the attempt key. No usable count is recorded, the sink decides that. The worker API has no stop hook, so a hard termination prints no final summary, and a record longer than the platform's line limit raises at its job instead of printing.Image to exactly one output Image at the same resolution. Unset width and height give SDXL Turbo and SD Turbo a working image of about 512x512 pixels at the source's aspect ratio, each side a multiple of 8; other checkpoints keep the source dimensions. Both set win as given; one set derives the other from the source's aspect ratio. The source is resized before diffusion, and the result is resized back to the source dimensions before emit.strength and num_inference_steps interact: the pipeline runs the integer product of their values in denoising steps. The scalar sentinel 18446744073709551615 means automatic; every other positive value is an explicit override. Automatic num_inference_steps selects 2 for SDXL Turbo and SD Turbo, so the default strength 0.5 yields one effective step. Other checkpoints use their pipeline's iteration default. Explicit values win. The product must remain at least 1; lowering both can leave no effective step.guidance_scale selects 0.0 for SDXL Turbo and SD Turbo, which were trained without classifier-free scaling. Other checkpoints use their pipeline's own default. Explicit values, including zero, win. The negative instruction has no effect when classifier-free scaling is disabled.model 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. A checkpoint that is not a local directory is fetched from the Hugging Face hub into the platform cache unless the hub is offline, when only cached files are used; gated or private repositories need hf_token.prompt both pass the input frame through unchanged without running diffusion. A diffusion fault, CUDA out-of-memory included, fails the job with its whole error and emits no picture. device is read once at startup; a value starting with `cuda` fails startup when CUDA is unavailable.Automated release
Automated release