Skip to main content

Image Upscaling & Enhancement with chaiNNer

This guide covers setting up node-based batch image upscaling workflows in chaiNNer, understanding model architectures (such as 4x-Nomos8kDAT), optimizing VRAM usage, and mastering advanced model chaining.

1. Quick Start: Batch Upscaling Workflow​

The workflow below takes a folder of images, passes each image through a PyTorch model (4x-Nomos8kDAT.pth), and saves the upscaled results to an output folder.

Node Flow Diagram​

Step-by-Step Setup​

  1. Verify PyTorch Dependency

    • Open chaiNNer.
    • Click the Dependency Manager (puzzle piece/download icon in the top-right header).
    • Ensure PyTorch is installed. (Required for running .pth models).

  2. Add Required Nodes Search for and place the following four nodes onto the canvas grid:

    • Load Model (PyTorch) (PyTorch > Load)
    • Load Images (Image > Batch Processing)
    • Upscale Image (PyTorch > Processing)
    • Save Image (Image > Input & Output)

  3. Connect Node Handles

    • Load Model (PyTorch) → Model Output ➔ Upscale Image → Model Input
    • Load Images → Image Output ➔ Upscale Image → Image Input
    • Upscale Image → Image Output ➔ Save Image → Image Input

  4. Configure Node Inputs

    • Load Model (PyTorch): Select your downloaded 4x-Nomos8kDAT.pth file.
    • Load Images: Browse and select the source directory containing input images.
    • Save Image: Choose the output destination directory and configure file naming (e.g., [name]_4x_DAT.[ext]).

  5. Execute Workflow

    • Click the green Play button ($lacktriangleright$) in the top navigation bar to begin processing the batch.

2. Model Architecture Breakdown​

Understanding the architecture of your .pth model helps in choosing the right model for specific image types.

ArchitectureDescriptionStrengthsBest Used For
DAT (Dual Aggregation Transformer)Combines spatial and channel attention mechanisms across dual transformer branches.Exceptional detail preservation, high fidelity, sharp text & fine textures without heavy artifacts.High-quality photos, textures, fine line art, architectural shots.
ESRGAN (Enhanced Super-Resolution GAN)Classical Convolutional Neural Network (CNN) based on residual-in-residual dense blocks.Fast inference, crisp lines, strong hallucinatory reconstruction for blurry sources.General photos, textures, game textures.
RealESRGANVariant of ESRGAN trained on realistic synthetic degraded images.Heavy noise removal, compression artifact removal.Low-res internet images, compressed JPEGs, legacy photos.
SwinIR / HATSwin Transformer / Hybrid Attention Transformer architectures.Top-tier perceptual quality and structural accuracy.Maximum quality renders, archival photo restoration (High VRAM required).
Compact / UltraCompactLightweight convolutional networks.Extremely lightweight, low memory overhead, very fast execution.Video frames, low VRAM GPUs, real-time preview batching.

4x-Nomos8kDAT is the model used in the Quick Start workflow above, so it's worth knowing what it's good at:

  • Source: OpenModelDB Model Page
  • Architecture: DAT (Dual Aggregation Transformer)
  • Scale: 4x
  • Trained on: High-resolution 8K photography
  • Best Used For: High-resolution photography, raw camera renders, detailed digital paintings, and fine textures.
  • Key Strength: Preserves sharp natural details (skin texture, fabric, hair, foliage, stone) without producing the "waxy" or over-smoothed artifacts typical of older GAN architectures.

3. Model File Formats Explained​

AI upscaling models come in several file formats:

  • .pth / .pt (PyTorch Checkpoint)
    • Pros: Native format for most super-resolution research models; widely compatible in chaiNNer.
    • Cons: Requires PyTorch runtime; slower initial load times.
  • .onnx (Open Neural Network Exchange)
    • Pros: Highly portable across runtimes (DirectML, TensorRT, NCNN, OpenVINO); faster execution on non-Nvidia GPUs.
    • Cons: Requires conversion from PyTorch; may lock batch sizes or resolution constraints if improperly converted.
  • .safetensors
    • Pros: Safe binary format that prevents arbitrary code execution risks inherent to Python pickling (.pth). Fast loading.

4. VRAM & Performance Optimization​

Large transformer models like DAT can easily consume significant VRAM at high resolutions. chaiNNer provides built-in mechanisms to handle large images safely.

Tile-Based Processing​

Instead of feeding a full 4K image into the GPU at once (which can crash with Out-Of-Memory / OOM errors):

  • Enable Tile Size inside the Upscale Image node settings.
  • Recommended Tile Sizes:
    • 4GB – 6GB VRAM: 256 or 384
    • 8GB – 12GB VRAM: 512 or 768
    • 16GB+ VRAM: 1024 or Auto
  • Tile Margin / Padding: Always set a small overlap (e.g., 16 or 32 pixels) to ensure seamless blending between tiles and prevent visible grid seams.

5. Advanced Workflow Patterns​

Pattern A: Two-Stage Upscaling (Restore + Scale)​

For heavily degraded or noisy images, running a 4x DAT model directly can amplify noise. Chain a restoration model before scaling:

Pattern B: Face Restoration Pipeline​

When upscaling photos with human faces, add a face restoration node after the main upscaler:

6. Troubleshooting Common Issues​

PyTorch Dependency Error

Symptom: Load Model (PyTorch) node displays a warning or fails to load. Fix: Open Dependency Manager ($\downarrow$ icon top-right) $\rightarrow$ Reinstall PyTorch $\rightarrow$ Restart chaiNNer.

CUDA Out of Memory (OOM)

Symptom: Workflow halts halfway through with torch.cuda.OutOfMemoryError. Fix: Reduce the Tile Size on the Upscale Image node (e.g., lower from 512 to 256) and ensure no heavy background GPU tasks are running.

Seams / Grid Lines on Output

Symptom: Grid patterns appear on the upscaled output image. Fix: Increase the Tile Padding / Overlap setting in the processing node to 32px or higher.

7. Where to Find Models​

  1. OpenModelDB

    • Centralized community database for finding and comparing upscaling models with side-by-side visual comparisons and architecture tags.

  2. Hugging Face

    • Search for model repos under tags like super-resolution, esrgan, dat, or hat.

  3. GitHub Releases

    • Original research repositories (e.g., Philip Hofmann's Nomos series, Xinntao's Real-ESRGAN, etc.) host raw .pth checkpoints under their Releases tabs.

8. Alternative GUI Tools to chaiNNer​

ToolTypeLicense / PriceBest For
UpscaylDesktop AppFree & Open-SourceQuick, one-click batch processing without setting up visual node trees.
QualityScalerDesktop AppFree & Open-SourceSimple Windows UI focusing on PyTorch / BSRGAN models with automatic tiling.
ComfyUINode-Based UIFree & Open-SourceDeep pipelines integrating upscaling models directly with Stable Diffusion / Flux generation pipelines.
Topaz Gigapixel AICommercialPaidDedicated commercial desktop suite for RAW photo processing and print workflows.