
Restores a degraded Image with a two-stage restoration pipeline — a SwinIR restorer feeds a ControlNet-guided Stable Diffusion prior — and emits a super-resolved Image at outscale. The networks come from the bound cleaner, sd and controlnet checkpoints, their sizes from the bound architecture file. Tiling bounds VRAM; a frame that already meets the min-size floors is passed through unchanged. Licences, as published: the component's own code (worker, tiled VAE, restorer layout) is under the Pipelogic Platform License; the vendored DiffBIR code (XPixelGroup/DiffBIR) is Apache-2.0, with MIT-licensed parts from latent-diffusion, Stable Diffusion 2, CLIP, open_clip, guided-diffusion, k-diffusion and DPM-Solver; the vendored SwinIR (JingyunLiang/SwinIR) is Apache-2.0; each model's licence lives on its bound file.
Restore and enlarge a heavily degraded photo from disk.
Low-resolution restoration before OCR.
steps, the classifier-free scale, the pre-enlarged width and height the network ran at, the precision and device compute selects, whether VAE tiling is on, the seed with digests of the global CPU and CUDA random-number states, and options holding the remaining settings. The positive and negative prompts are logged only as the SHA-256 and codepoint length of their 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; encode also covers the stage-1 restorer, and 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.compute names, and CPU inference is impractically slow.architecture JSON must hold exactly the sections cleaner, cldm and diffusion with every constructor argument of their networks, the restorer's scale (its pixel-unshuffle factor, a power of two that also sets its count of x2 upsampling stages) included; a missing, unknown or null key fails startup naming the key, and so does a scale that is not a power of two of at least 2. Startup also fails when the cleaner or controlnet checkpoint lacks a tensor its network holds, holds one the network does not take, or holds one of another shape, and when the sd checkpoint lacks a tensor the networks read or holds one of another shape, naming each missing tensor. Only the text-encoder tensors the configured CLIP layer never reads may be absent from the sd checkpoint (the transformer blocks after that layer, logit_scale and text_projection); sd tensors the networks do not take are ignored. Bind the architecture that matches the three checkpoints.compute without CUDA available fails startup.outscale 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.outscale, so memory and time grow with the square of outscale and with the input size; max_input_side caps the input before the enlargement, and a frame that exhausts VRAM fails its job.max_input_side caps the longest input side before inference. The stage-1 and control-branch tiling options bound VRAM at a small quality cost. VAE tiling (vae_encoder_tiled, vae_decoder_tiled) gives the untiled result up to rounding: it normalizes with whole-frame statistics, keeps each level's activations in host memory (several GB of RAM for a multi-megapixel frame) and copies them to the GPU tile by tile, which adds decode time. cldm_tile_overlap_ratio sets how much of a control-branch tile its neighbours overlap. compute selects the precision and device together: fp16, bf16 or fp32 on a CUDA device, fp32 on cpu.pos_prompt and neg_prompt steer the diffusion prior; an empty positive prompt gives the text encoder an empty string, and guidance still applies against the negative prompt. min_width and min_height gate the skip jointly — a frame is passed through unchanged, at its original size, when every set floor is met — while max_width and max_height clamp the output.seed and every frame draws the next noise, so a frame's result depends on the frames before it; the same seed with the same frames in the same order draws the same noise.live-test prerelease
Automated release