pyiqa.archs.fgresq_arch¶

FGResQ metric implementation.

Reference:: Sheng, X., Pan, X., Yang, Z., Chen, P., and Li, L. Fine-grained Image Quality Assessment for Perceptual Image Restoration. AAAI 2026.
Reference URL:: https://github.com/sxfly99/FGResQ

Module Contents¶

pyiqa.archs.fgresq_arch.default_model_urls[source]¶

pyiqa.archs.fgresq_arch.load_checkpoint(model_path)[source]¶

Load and normalize a checkpoint state dict.

Parameters:: model_path (str) – Path to the checkpoint file.
Returns:: Cleaned state dict.
Return type:: dict

pyiqa.archs.fgresq_arch.get_pooler_output(model, x)[source]¶

Extract pooled CLIP visual features.

Parameters:

model (CLIPVisionModel) – CLIP vision backbone.
x (torch.Tensor) – Input tensor.

Returns:

Pooled feature tensor.

Return type:

torch.Tensor

class pyiqa.archs.fgresq_arch.FGResQ(clip_model='openai/clip-vit-base-patch16', task_clip_model='openai/clip-vit-base-patch16', clip_freeze=True, pretrained=True, pretrained_model_path=None, degradation_model_path=None, input_size=224, resize_size=256, default_mean=OPENAI_CLIP_MEAN, default_std=OPENAI_CLIP_STD, score_scale=0.3)[source]¶

Bases: torch.nn.Module

FGResQ no-reference image quality model.

Parameters:

clip_model (str) – HuggingFace CLIP vision backbone id.
task_clip_model (str) – CLIP backbone for the degradation-aware branch.
clip_freeze (bool) – Whether to freeze the main CLIP backbone.
pretrained (bool) – Whether to load official pretrained weights.
pretrained_model_path (str | None) – Optional local checkpoint path for the main FGResQ weights.
degradation_model_path (str | None) – Optional local checkpoint path for the degradation branch weights.
input_size (int) – Final center crop size.
resize_size (int) – Resize size before center crop.
given (The network returns a single quality score when only x0 is)

:param : :param and returns quality0: :param quality1: :param rank: :param and rank_prob: :param when both x0 and x1 are provided.: :param default_mean: Input normalization mean. :type default_mean: tuple[float, float, float] :param default_std: Input normalization std. :type default_std: tuple[float, float, float] :param score_scale: Scale factor used before the sigmoid output head. :type score_scale: float

load_degradation_weights(model_path)[source]¶

Load degradation-branch weights.

Parameters:: model_path (str) – Path to the degradation checkpoint.

load_pretrained_weights(model_path)[source]¶

Load main FGResQ weights.

Parameters:: model_path (str) – Path to the FGResQ checkpoint.

preprocess(x)[source]¶

Preprocess an input image tensor.

Parameters:: x (torch.Tensor) – Input tensor with shape (N, C, H, W).
Returns:: Preprocessed tensor.
Return type:: torch.Tensor

get_quality_features(x)[source]¶

Extract FGResQ quality features.

Parameters:: x (torch.Tensor) – Preprocessed image tensor.
Returns:: Concatenated quality features.
Return type:: torch.Tensor

forward_single(x)[source]¶

Predict single-image quality.

Parameters:: x (torch.Tensor) – Preprocessed image tensor.
Returns:: Predicted quality score.
Return type:: torch.Tensor

forward_pair(x0, x1)[source]¶

Predict pairwise quality and comparison logits.

Parameters:

x0 (torch.Tensor) – First preprocessed image tensor.
x1 (torch.Tensor) – Second preprocessed image tensor.

Returns:

Single-image scores for both inputs and pairwise comparison logits.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor]

get_pair_rank(compare_logits)[source]¶

Convert comparison logits to a discrete rank label.

Parameters:: compare_logits (torch.Tensor) – Pairwise comparison logits.
Returns:: Rank tensor with shape (N, 1) where 0=image2_better, 1=image1_better, and 2=similar_quality.
Return type:: torch.Tensor

get_pair_result(quality0, quality1, compare_logits)[source]¶

Format pairwise prediction outputs.

Parameters:

quality0 (torch.Tensor) – Quality of the first image.
quality1 (torch.Tensor) – Quality of the second image.
compare_logits (torch.Tensor) – Pairwise comparison logits.

Returns:

quality0, quality1, rank, and rank_prob.

Return type:

tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]

forward(x0, x1=None)[source]¶

Forward pass for FGResQ.

Parameters:

x0 (torch.Tensor) – First input tensor.
x1 (torch.Tensor | None) – Optional second input tensor.

Returns:

Output depends on whether x1 is given.

Return type:

torch.Tensor | tuple