
Restores and super-resolves an Image with a one-step diffusion network: the VAE and UNet of a bound diffusers base snapshot are loaded with weights from a bound state-dict file and map a degraded input to a restored result in a SINGLE denoising step, emitting the enlarged Image at the requested factor.
Restore and enlarge compressed or noisy photos and publish them: a file source feeds frames, this component runs single-step diffusion restoration, and the result is streamed over HTTP.
Restore low-quality document captures before OCR so the recognizer reads a cleaner image, then publish the recognized text as JSON.
device, the width and height the network ran at after enlargement, and options holding the remaining settings; the random initialiser is null, with digests of the global CPU and CUDA random-number states. The prompt is logged only as the SHA-256 and codepoint length of its text; the input picture is logged as a digest of its decoded RGB pixels with its shape. The sources entry holds that pixel digest and a null position, because this image stream carries no PictureIdentity. 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, a null position, 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; 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.device without CUDA available fails startup.adapter file is a pickle read with tensors only (weights_only), so a file holding anything but a dict with the state dicts "vae" and "unet" fails startup rather than running code; the worker never executes what the file contains. Those state dicts must match the sd base exactly, or startup fails. Startup holds the whole adapter in host memory beside the base model before moving both to the device.native_scale has no default: set it to the enlargement factor the bound adapter weights were trained for, because no bound file records it.max_input_side set to Nothing the network runs on the full frame, so a large frame, or a large outscale on the result, can exhaust VRAM or host memory: the job then fails with the whole error and emits no image.native_scale before the network and the result is then resampled to the requested output factor, while a pre-network cap on the longest input side plus latent and VAE tiling bound VRAM at a small quality cost. The overlap of neighbouring latent tiles is a fraction of the tile, so it is always smaller than the tile.prompt is fed to the text encoder for every frame; an empty value gives the encoder an empty string and text beyond its token limit is truncated. A color-alignment step matches the output back to the input via adaptive instance norm, wavelet reconstruction, or none.sd and adapter files; nothing is downloaded.Automated release
Automated release