<!-- Generated by tools/build_node_reference.py from the node definitions. Do not edit by hand. -->

# Upscale nodes

AI upscaling, repair, enhancement, and restoration.

10 nodes. [All sections](README.md)

- [AI Upscale](#ai-upscale)
- [Bit Depth Convert](#bit-depth-convert)
- [Downscale 32-bit](#downscale-32-bit)
- [Upscale 32-bit](#upscale-32-bit)
- [Upscale By Size](#upscale-by-size)
- [Upscale Face Restore](#upscale-face-restore)
- [Upscale Image](#upscale-image)
- [Upscale Router](#upscale-router)
- [Upscale Tiler](#upscale-tiler)
- [Upscale Video](#upscale-video)

## AI Upscale

`RadianceAIUpscale`

AI-powered upscaling using neural network models. Supports tiled processing for large images.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Input image to upscale. |
| `model_name` | choice | `RealESRGAN_x4plus` | `RealESRGAN_x4plus`, `RealESRGAN_x4plus_anime_6B`, `RealESRGAN_x2plus`, `ESRGAN_4x`, `4x-UltraSharp`, `4x-AnimeSharp`, `SwinIR_4x`, `HAT_4x`, `SUPIR-v0F_fp16`, `SUPIR-v0Q_fp16` | AI upscaling model. RealESRGAN_x4plus is recommended for general use. |
| `mode` | choice | `Standard` | `Standard`, `Refine (HDR)`, `Normalize (HDR)` | Processing mode. 'Standard' = direct upscale. 'Refine (HDR)' = log-compression for highlights. 'Normalize (HDR)' = scales to safe range. |
| `tile_size` | int | 512 | 128 to 1024, step 64 | Tile size for processing. Smaller = less VRAM, slower. |
| `tile_overlap` | int | 32 | 0 to 128, step 8 | Overlap between tiles to avoid seams. |
| `auto_download` | boolean | on |  | Automatically download models if not found. |
| `unload_model` (optional) | boolean | off |  | Unload model from VRAM after processing to free memory. |
| `sdxl_model_name` (optional) | string |  |  | SUPIR only: filename of your SDXL base checkpoint (e.g. sd_xl_base_1.0_0.9vae.safetensors). Leave empty to auto-detect from the checkpoints folder. Ignored for all non-SUPIR models. |
| `supir_prompt` (optional) | string |  | multi-line text | SUPIR only: text description for SUPIR upscaling. Leave empty for default conditioning. Ignored for all other models. |
| `vae` (optional) | VAE |  |  | Reserved for future SUPIR v2 loader support. Currently unused. |
| `clip` (optional) | CLIP |  |  | Reserved for future SUPIR v2 loader support. Currently unused. |

**Outputs**

| Output | Type | What it is |
| :--- | :--- | :--- |
| `image` | IMAGE | Upscaled image. |
| `info` | STRING | Information about the upscaling process. |

## Bit Depth Convert

`RadianceBitDepthConvert`

Convert between bit depths with professional dithering to reduce banding.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Input image in float32 format. |
| `output_depth` | choice |  | `32-bit Float`, `16-bit Float`, `16-bit Int`, `10-bit`, `8-bit` | Target bit depth. |
| `dithering` (optional) | choice | `None` | `None`, `Floyd-Steinberg`, `Ordered`, `Blue Noise`, `Random` | Dithering algorithm. Floyd-Steinberg = error diffusion, Ordered = Bayer, Blue Noise = high-frequency, Random = simple noise. |
| `dither_strength` (optional) | float | 1 | 0 to 2, step 0.1 | Dithering intensity. 1.0 = standard. |
| `seed` (optional) | int | 0 | 0 to 18446744073709551615 | Random seed for reproducible Blue Noise and Random dithering. |

**Outputs**

| Output | Type | What it is |
| :--- | :--- | :--- |
| `converted_image` | IMAGE | Image quantized to target bit depth. |
| `bit_depth_info` | STRING | Information about the conversion. |

## Downscale 32-bit

`RadianceDownscale32bit`

32-bit downscaling with the exact resampling kernel on GPU or CPU and an adjustable anti-aliasing prefilter.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Image or batch to downscale, display-encoded sRGB or linear float (see input_color_space). Values above 1.0 are kept. |
| `scale_factor` | float | 0.5 | 0.01 to 1, step 0.05 | Output size as a fraction of the input (rounded down, minimum 1 px). |
| `method` | choice |  | `lanczos`, `lanczos4`, `bicubic`, `mitchell`, `catrom`, `gaussian`, `bilinear` | Resampling kernel, run exactly as named on CUDA and CPU alike: lanczos, lanczos4, bicubic (Catmull-Rom), mitchell, catrom, gaussian, bilinear. |
| `antialiasing` (optional) | float | 0.5 | 0 to 1, step 0.05 | Width of the downscaling prefilter. 0.5 = the kernel's textbook width (stretched by the scale factor), 0 = no prefilter (sharpest, aliases and moires), 1 = twice as wide (softest). pre_blur adds a Gaussian on top. |
| `pre_blur` (optional) | float | 0 | 0 to 2, step 0.1 | Gaussian sigma in input pixels applied before downscaling to suppress aliasing and moire. 0 = off. |
| `process_in_linear` (optional) | boolean | on |  | Decode sRGB to linear before resampling and re-encode after, for gamma-correct filtering. Only acts when input_color_space is sRGB; linear/HDR input is never converted. |
| `use_gpu` (optional) | boolean | on |  | Run on CUDA when available (MPS is not used). Falls back to the CPU path, which uses the exact resampling kernels. |
| `input_color_space` (optional) | choice | `sRGB` | `sRGB`, `Linear`, `Auto` | Colour space of the input image. Linear skips the sRGB-linear conversion for HDR or already-linear input. Auto: Linear when any value is outside 0-1, else sRGB. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `downscaled_image` | IMAGE |
| `width` | INT |
| `height` | INT |

## Upscale 32-bit

`RadianceProUpscale`

Professional 32-bit upscaler optimized for Flux with HDR support.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Image or batch to resize, display-encoded sRGB or linear float (see input_color_space). Alpha is resized separately. |
| `scale_factor` | float | 2 | 0.1 to 8, step 0.1 | Output size as a multiple of the input (rounded down to whole pixels). Below 1.0 downscales. |
| `preset` | choice |  | `Custom`, `Flux Default`, `Flux Sharp`, `Flux Smooth`, `Flux HDR`, `Flux Print`, `Flux Cinematic`, `Flux Maximum` | Custom uses the widgets below. Any other preset overrides method, sharpening, detail_enhancement and antialiasing, and the HDR/Cinematic presets force process_in_linear on. |
| `method` (optional) | choice |  | `lanczos`, `lanczos4`, `bicubic`, `mitchell`, `catrom`, `hermite`, `gaussian`, `bilinear`, `nearest` | Resampling kernel, run exactly on every path (GPU, CPU and tiled): lanczos (3 lobes), lanczos4, bicubic and catrom (Catmull-Rom), mitchell (B=C=1/3), hermite, gaussian, bilinear, nearest. |
| `sharpening` (optional) | float | 0.3 | 0 to 2, step 0.05 | Unsharp-mask amount applied after resizing. 0 = off, 1 = add the full detail difference once. |
| `sharpen_radius` (optional) | float | 1 | 0.5 to 5, step 0.1 | Gaussian sigma in output pixels for the unsharp mask. Larger sharpens coarser detail. |
| `detail_enhancement` (optional) | float | 0.2 | 0 to 1, step 0.05 | Multi-scale luma detail boost (blur sigmas 1, 2 and 4 px) added equally to RGB after resizing. 0 = off. |
| `antialiasing` (optional) | float | 0.3 | 0 to 1, step 0.05 | Edge-aware softening applied last: blends in a 1 px Gaussian blur in proportion to edge strength. 0 = off; counteracts sharpening on edges. |
| `input_color_space` (optional) | choice |  | `sRGB`, `Linear`, `Auto` | Encoding of the input. sRGB allows the linear-light conversion; Linear leaves values untouched. Auto: Linear when any value is outside 0-1 (HDR), else sRGB; the info output says which. |
| `process_in_linear` (optional) | boolean | on |  | Decode sRGB to linear before resampling and re-encode after, for gamma-correct filtering. Only acts when input_color_space is sRGB; linear/HDR input is never converted. |
| `use_tiles` (optional) | boolean | off |  | Process image in overlapping tiles to handle large images that exceed VRAM. |
| `tile_size` (optional) | int | 512 | 128 to 2048, step 64 | Input tile size in pixels for the tiled CPU path. Used only when tiling is on (or forced above 64 MP) and the image is larger than this. |
| `tile_overlap` (optional) | int | 64 | 16 to 256, step 16 | Overlap in input pixels between tiles, blended to hide seams. Tiled path only. |
| `output_bit_depth` (optional) | choice |  | `32-bit Float`, `16-bit Float`, `8-bit` | Precision of the result (always returned as float). 32-bit keeps values above 1.0; 16-bit rounds to half precision; 8-bit clamps to 0-1 and quantises to 256 levels. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `upscaled_image` | IMAGE |
| `width` | INT |
| `height` | INT |
| `info` | STRING |

## Upscale By Size

`RadianceUpscaleBySize`

Upscale to exact dimensions with aspect ratio control.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `image` | IMAGE |  |  | Image or batch to resize, display-encoded sRGB or linear float (see input_color_space). Alpha is resized separately. |
| `width` | int | 2048 | 64 to 16384, step 8 | Target output width in pixels. |
| `height` | int | 2048 | 64 to 16384, step 8 | Target output height in pixels. |
| `method` | choice |  | `lanczos`, `lanczos4`, `bicubic`, `mitchell`, `catrom`, `hermite`, `gaussian`, `bilinear`, `nearest` | Resampling kernel, run exactly as named: lanczos (3 lobes), lanczos4, bicubic and catrom (Catmull-Rom), mitchell (B=C=1/3), hermite, gaussian, bilinear, nearest. |
| `maintain_aspect` (optional) | boolean | on |  | Keep the source aspect ratio using aspect_mode to fit the width x height box. Off resizes to exactly width x height. |
| `aspect_mode` (optional) | choice |  | `fit`, `fill`, `stretch` | With maintain_aspect on. fit: largest size inside the box. fill: smallest size covering the box (not cropped, so one side exceeds it). stretch: exactly width x height. |
| `sharpening` (optional) | float | 0.2 | 0 to 2, step 0.05 | Unsharp-mask amount (1 px sigma) applied after resizing. 0 = off. |
| `process_in_linear` (optional) | boolean | on |  | Decode sRGB to linear before resampling and re-encode after, for gamma-correct filtering. Only acts when input_color_space is sRGB; linear/HDR input is never converted. |
| `input_color_space` (optional) | choice | `sRGB` | `sRGB`, `Linear`, `Auto` | Colour space of the input. Linear skips the sRGB-to-linear conversion for HDR/linear input. Auto: Linear when any value is outside 0-1, else sRGB. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `upscaled_image` | IMAGE |
| `final_width` | INT |
| `final_height` | INT |

## Upscale Face Restore

`RadianceUpscaleFaceRestore`

Restore and enhance facial detail using a face restoration model.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `images` | IMAGE |  |  | Upscaled image batch (B,H,W,C) float32. |
| `face_model` | choice | `auto (CodeFormer → GFPGAN → skip)` | `auto (CodeFormer → GFPGAN → skip)`, `codeformer`, `gfpgan_v1.4`, `skip (detection only)` | Face restoration model. Auto tries CodeFormer first, falls back to GFPGAN, skips if neither is available. |
| `fidelity_weight` | float | 0.75 | 0 to 1, step 0.05 | CodeFormer fidelity: 0 = maximum enhancement (creative), 1 = faithful to input (precise). 0.5–0.8 is recommended for most upscaled content. |
| `blend_radius` | int | 20 | 0 to 80, step 4 | Gaussian feather radius in pixels at face crop edge. Higher = softer transition. 0 = hard paste. |
| `face_pad_frac` | float | 0.25 | 0 to 0.6, step 0.05 | Extra padding around each detected face bbox (fraction of face width/height). 0.25 = 25%. |
| `min_face_px` | int | 64 | 16 to 256, step 16 | Smallest face (in pixels) to process. Smaller faces are skipped. |
| `colour_correct` | boolean | on |  | Apply histogram-match colour correction after restoration to cancel diffusion colour drift. |
| `colour_strength` | float | 0.8 | 0 to 1, step 0.05 | Strength of histogram-match correction. 1.0 = full match to input colours. |
| `original_images` (optional) | IMAGE |  |  | Original (pre-upscale) images for colour reference. Used by histogram-match correction. Leave disconnected to use the restored images as self-reference. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `restored` | IMAGE |
| `face_mask` | IMAGE |
| `pass_info` | STRING |

## Upscale Image

`RadianceUpscaleImage`

Upscale a still image using a selected super-resolution model.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `operation` | choice | `Upscale` | `Upscale`, `Route` | Upscale: run super-resolution. Route: pass images through and output the recommended tier, content class and stats (no upscale). |
| `images` | IMAGE |  |  | Input image batch. |
| `scale` (optional) | choice | `4×` | `2×`, `4×`, `8× (tile cascade)` | Output scale. 8× runs the 4× model twice, the second pass area-downsampled by half. |
| `hdr_mode` (optional) | choice | `auto` | `auto`, `preserve`, `clamp` | Scene-linear / HDR handling. auto: preserve range when input exceeds 1.0, else clamp. preserve: Reinhard tonemap before SR and re-expand after (keeps highlights >1.0). clamp: legacy [0,1] (LDR). |
| `color_encoding` (optional) | choice | `passthrough` | `passthrough`, `linear<->sRGB`, `linear<->LogC3` | Encode scene-linear -> display (sRGB/LogC3) before SR and decode after, so the LDR-trained network sees the domain it expects. passthrough: feed pixels unchanged. |
| `mode` (optional) | choice | `precise` | `precise`, `creative`, `balanced` | precise: Real-ESRGAN fidelity-first. creative: diffusion detail hallucination (requires VRAM). balanced: GAN upscale + light sharpening. |
| `tile_size` (optional) | int | 512 | 128 to 1024, step 64 | Tile size in input pixels. Reduce if OOM. |
| `overlap` (optional) | int | 128 | 32 to 256, step 32 | Overlap between tiles in input pixels, blended to hide seams. |
| `sharpness_boost` (optional) | float | 0 | 0 to 1, step 0.05 | Unsharp mask strength applied after upscale. |
| `denoise_pre` (optional) | float | 0 | 0 to 1, step 0.05 | Gaussian pre-denoise strength. |
| `upscale_model` (optional) | UPSCALE_MODEL |  |  | Optional ComfyUI UPSCALE_MODEL; its native scale should match scale. When connected it replaces the built-in tiers (model_tier and creative mode are ignored). |
| `model_tier` (optional) | choice | `auto` | `auto`, `tier1_fast (Real-ESRGAN — GAN, ms/frame)`, `tier2_quality (HAT-L — transformer SOTA PSNR)`, `tier2_quality (SwinIR-L — transformer quality)`, `tier3_creative (SD x4 — diffusion hallucination)`, `tier3_creative (SeedVR2 — one-step video diffusion)` | Model tier. 'auto' selects based on content analysis. |
| `diffusion_steps` (optional) | int | 20 | 1 to 50, step 1 | Inference steps for Tier 3 diffusion (creative mode or a tier3 model_tier). More steps are slower; SeedVR2 is one-step. |
| `diffusion_noise_level` (optional) | int | 20 | 0 to 350, step 10 | SD x4 upscaler only: noise added to the low-res input. Higher lets the model invent more detail and drift further from the source. |
| `guidance_scale` (optional) | float | 7.5 | 1 to 20, step 0.5 | SD x4 upscaler only: classifier-free guidance toward enhancement_prompt. Higher follows the prompt more strongly. |
| `enhancement_prompt` (optional) | string |  |  | Text prompt for creative mode diffusion steering. |
| `prefer_speed` (optional) | boolean | off |  | Always recommend Tier 1 fast regardless of content. |
| `sample_frame` (optional) | int | 0 | 0 to 9999 | Index of frame to analyse (for Route operation). |

**Outputs**

| Output | Type |
| :--- | :--- |
| `image_a` | IMAGE |
| `image_b` | IMAGE |
| `info` | STRING |
| `data1` | STRING |
| `data2` | STRING |

## Upscale Router

`RadianceUpscaleRouter`

Measure noise, sharpness and saturation on one frame and recommend a model_tier for Upscale Image or Video. Heuristic statistics only; images pass through unchanged.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `images` | IMAGE |  |  | Image or frame batch to analyse; returned unchanged on the images output. |
| `prefer_speed` | boolean | off |  | Always recommend Tier 1 fast regardless of content. |
| `sample_frame` | int | 0 | 0 to 9999 | Index of frame to analyse (for video batches). |

**Outputs**

| Output | Type |
| :--- | :--- |
| `recommended_tier` | STRING |
| `content_class` | STRING |
| `stats_json` | STRING |
| `images` | IMAGE |

## Upscale Tiler

`RadianceUpscaleTiler`

Tile large images into overlapping patches for memory-safe upscaling.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `operation` | choice | `Tile` | `Tile`, `ColourFix` | Tile: tiled upscale of images (output clamped to 0-1). ColourFix: histogram-match source to reference to remove upscaler colour drift. |
| `images` (optional) | IMAGE |  |  | Input image batch (B,H,W,C) float32. |
| `scale` (optional) | choice | `4×` | `2×`, `4×`, `8× (tile cascade)` | Upscale factor. 8× uses two cascaded 4× passes. |
| `tile_size` (optional) | int | 512 | 128 to 2048, step 64 | Tile side in input pixels. Smaller = less VRAM. |
| `overlap` (optional) | int | 128 | 32 to 512, step 32 | Tile overlap in input pixels. ≥20% of tile_size recommended. |
| `blend_mode` (optional) | choice | `laplacian_pyramid` | `laplacian_pyramid`, `gaussian_feather`, `linear` | laplacian_pyramid: best quality. gaussian_feather: fast. linear: simple. |
| `upscale_model` (optional) | UPSCALE_MODEL |  |  | Any ComfyUI UPSCALE_MODEL; its native scale should match scale. Leave empty to use the built-in model chosen by model_tier. |
| `model_tier` (optional) | choice | `tier1_fast (Real-ESRGAN — GAN, ms/frame)` | `auto`, `tier1_fast (Real-ESRGAN — GAN, ms/frame)`, `tier2_quality (HAT-L — transformer SOTA PSNR)`, `tier2_quality (SwinIR-L — transformer quality)`, `tier3_creative (SD x4 — diffusion hallucination)`, `tier3_creative (SeedVR2 — one-step video diffusion)` | Built-in model tier when no upscale_model is connected. |
| `source` (optional) | IMAGE |  |  | Upscaled image with colour drift (ColourFix mode). |
| `reference` (optional) | IMAGE |  |  | Original pre-upscale image — colour reference (ColourFix mode). |
| `cf_strength` (optional) | float | 1 | 0 to 1, step 0.05 | ColourFix strength: 0 = off, 1 = full CDF match. |
| `n_bins` (optional) | int | 512 | 64 to 2048, step 64 | Histogram resolution (ColourFix mode). |

**Outputs**

| Output | Type |
| :--- | :--- |
| `image_a` | IMAGE |
| `image_b` | IMAGE |
| `info` | STRING |

## Upscale Video

`RadianceUpscaleVideo`

Upscale a video sequence using a selected super-resolution model.

**Inputs**

| Input | Type | Default | Range or choices | What it does |
| :--- | :--- | :--- | :--- | :--- |
| `frames` | IMAGE |  |  | Video frame batch (B,H,W,C) float32. B = frame count. |
| `scale` | choice | `4×` | `2×`, `4×`, `8× (tile cascade)` | Output scale. 8× runs the 4× model twice, the second pass area-downsampled by half. |
| `tile_size` | int | 512 | 128 to 1024, step 64 | Spatial tile size in input pixels. Reduce if OOM. |
| `overlap_spatial` | int | 128 | 32 to 256, step 32 | Spatial tile overlap in input pixels. |
| `window_size` | int | 16 | 4 to 64, step 4 | Frames per processing batch (VRAM). Tier 1/2 models are single-image: each frame is still upscaled on its own. |
| `overlap_temporal` | int | 4 | 1 to 16, step 1 | Frames shared between adjacent windows. Minimum 1 for seam-free stitching. |
| `flow_compensation` | boolean | on |  | At window seams each overlap frame is upscaled twice; Lucas-Kanade flow aligns the two results before they are blended. It does not compensate camera motion between frames. |
| `sharpness_boost` | float | 0 | 0 to 1, step 0.05 | Unsharp-mask strength (sigma 1.5 px) applied after upscaling. 0 = off. |
| `upscale_model` (optional) | UPSCALE_MODEL |  |  | Optional ComfyUI UPSCALE_MODEL; its native scale should match scale. When connected it replaces model_tier. |
| `model_tier` (optional) | choice | `tier1_fast (Real-ESRGAN — GAN, ms/frame)` | `auto`, `tier1_fast (Real-ESRGAN — GAN, ms/frame)`, `tier2_quality (HAT-L — transformer SOTA PSNR)`, `tier2_quality (SwinIR-L — transformer quality)`, `tier3_creative (SD x4 — diffusion hallucination)`, `tier3_creative (SeedVR2 — one-step video diffusion)` | Select 'SeedVR2' for best temporal consistency on video. Requires seedvr2 or diffusers package. |
| `enhancement_prompt` (optional) | string |  |  | Text prompt for Tier 3 diffusion steering (e.g. 'cinematic film grain, detailed textures'). |
| `diffusion_steps` (optional) | int | 1 | 1 to 50, step 1 | Diffusion inference steps. SeedVR2 uses 1 (one-step); SD x4 upscaler recommended 15-25. |
| `hdr_mode` (optional) | choice | `auto` | `auto`, `preserve`, `clamp` | Scene-linear / HDR handling. auto: preserve range when input exceeds 1.0, else clamp. preserve: Reinhard tonemap before SR and re-expand after. clamp: legacy [0,1] (LDR). |
| `color_encoding` (optional) | choice | `passthrough` | `passthrough`, `linear<->sRGB`, `linear<->LogC3` | Encode scene-linear -> display (sRGB/LogC3) before SR and decode after. passthrough: feed pixels unchanged. |

**Outputs**

| Output | Type |
| :--- | :--- |
| `upscaled` | IMAGE |
| `confidence_map` | IMAGE |
| `pass_info` | STRING |