https://upscale.wiki/wiki/Official_Research_Models#SwinIR
This might be a good jumping off point to find what you're looking for

Cheating a little, but here's the answer from ChatGPT:

1. Upscaling Factor (x2, x3, x4, x8)

The numbers x2, x3, x4, and x8 refer to the scaling factor of the model. For example:

x2: The model will upscale the image by a factor of 2.x4: The model will upscale the image by a factor of 4.This allows flexibility depending on how much enlargement you need.

2. Training Dataset (DF2K vs DIV2K vs BSRGAN)

DF2K and DIV2K are popular image super-resolution datasets.

DIV2K (Diverse 2K) is a large dataset used for super-resolution tasks, known for its high-quality, diverse images.DF2K is a combination of the DIV2K and Flickr2K datasets, used to improve training.

BSRGAN is a dataset and model used for "real-world" super-resolution tasks, addressing degradation like blur, noise, and compression.

3. Task Type (Classical SR, Real SR, Denoising, Compression Artifact Removal)

ClassicalSR: Refers to super-resolution tasks on images with artificial, well-known degradation types (ideal conditions).

Examples: 001_classicalSR_DF2K_s64w8_SwinIR-M_x3.pth

RealSR: Focuses on real-world super-resolution, often involving more complex, real-world degradations (like blur, noise, etc.).

Examples: 003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth

GrayDN and ColorDN: Models for denoising tasks on grayscale or color images.

Examples: 004_grayDN_DFWB_s128w8_SwinIR-M_noise15.pth (gray denoising, noise level 15).

CAR (Compression Artifact Removal): Deals with removing JPEG compression artifacts.

Examples: 006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth (for JPEG quality level 10).

4. Model Size (SwinIR-M, SwinIR-S, SwinIR-L)

SwinIR-M: Medium model, balanced in terms of performance and resource usage.

SwinIR-S: Small model, faster and more lightweight, designed for low-resource scenarios.

SwinIR-L: Large model, slower but yields the highest quality due to more parameters.

Example: 003_realSR_BSRGAN_DFO_s64w8_SwinIR-L_x4_PSNR.pth (Large version for real-world super-resolution).

5. Objective (GAN vs PSNR)

GAN: Generative Adversarial Network-based model, used for generating more realistic textures. Great for perceptual quality but might not score well on traditional quality metrics like PSNR.

Example: 003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_GAN.pth

PSNR: A model trained to optimize Peak Signal-to-Noise Ratio. It focuses more on pixel accuracy and produces higher PSNR scores but might be less perceptually realistic compared to GAN-based models.

Example: 003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_PSNR.pth

with-dict-keys-params-and-params_ema: These indicate versions of the model that include exponential moving average (EMA) parameters, which can stabilize training and improve inference quality.

6. Noise Levels (for Denoising)

In the denoising models, the numbers like noise15, noise25, and noise50 refer to the level of noise the model is trained to handle. Higher numbers represent images with more noise.

Example: 004_grayDN_DFWB_s128w8_SwinIR-M_noise50.pth (handles grayscale images with high noise levels).

7. JPEG Compression Levels

In compression artifact removal models, jpeg10, jpeg20, etc., refer to the JPEG quality level the model targets. Lower numbers correspond to lower-quality images (heavily compressed).

Example: 006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth (handles JPEG images compressed at quality level 10).

Summary:

Upscaling Factor (x2, x4, etc.): How much the model will upscale the image.

Dataset (DF2K, DIV2K, BSRGAN): Indicates the dataset used for training.

Task Type (ClassicalSR, RealSR, Denoising, Compression Removal): Specifies the type of task the model was trained for.

Model Size (SwinIR-S, M, L): Refers to small, medium, and large versions of the SwinIR model, affecting speed and quality.

Objective (GAN vs PSNR): Whether the model is optimized for perceptual quality (GAN) or pixel accuracy (PSNR).

Noise/JPEG Levels: Specifies the level of noise or compression the model targets for denoising or compression artifact removal.