Noise Removal

AI-powered noise and grain removal with multi-tier quality options, using the Python sidecar (SCUNet model).

API Endpoint

POST /api/v1/tools/noise-removal

Processing: Asynchronous (returns 202, poll /api/v1/jobs/{jobId}/progress for status via SSE)

Model bundle: upscale-enhance (4-5 GB)

Parameters

Parameter	Type	Required	Default	Description
file	file	Yes	-	Image file (multipart)
tier	string	No	"balanced"	Quality tier: quick, balanced, quality, maximum
strength	number	No	50	Denoising strength (0-100)
detailPreservation	number	No	50	How much detail to preserve (0-100). Higher values keep more texture
colorNoise	number	No	30	Color noise reduction strength (0-100)
format	string	No	"original"	Output format: original, png, jpeg, webp, avif, jxl
quality	number	No	90	Output encoding quality (1-100)

Example Request

curl -X POST http://localhost:13490/api/v1/tools/noise-removal \
  -F "[email protected]" \
  -F 'settings={"tier":"quality","strength":60,"detailPreservation":70,"colorNoise":40}'

Response

Initial Response (202 Accepted)

{
  "jobId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "async": true
}

Progress (SSE at /api/v1/jobs/{jobId}/progress)

event: progress
data: {"phase":"processing","stage":"Denoising...","percent":65}

Final Result (via SSE)

{
  "phase": "complete",
  "percent": 100,
  "result": {
    "jobId": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "downloadUrl": "/api/v1/download/{jobId}/noisy-photo_denoised.jpg",
    "originalSize": 500000,
    "processedSize": 380000
  }
}

Notes

• Requires the upscale-enhance model bundle to be installed (4-5 GB).
• Quality tiers trade speed for quality: quick is fastest with basic denoising, maximum uses the most thorough multi-pass approach.
• The detailPreservation parameter is critical for textured subjects (fabric, hair, foliage). Higher values prevent the denoiser from smoothing away fine detail.
• When format is set to "original", the output format matches the input file format.
• Supports HEIC/HEIF, RAW, TGA, PSD, EXR, and HDR input formats via automatic decoding.