Bad quality on larger image size

[minh-nguyenhoang]

The model is strictly trained on 512x512 input, show when dealing with image size larger than 512, the performance degraded significantly, while diffusion prior model have an advantage of having much better pre-training knowledge, wiith much better performance on different image size.
Furthermore, I've quick-tested the 1.4B models and doesn't get good result: The super-resolutioned image look very flat and there are artifact (the multi-step model is even worse than the one-step).

[cswry]

Thanks for the feedback.

Yes, the current released models were trained only at 512x512 resolution, so direct inference on images larger than 512 can degrade noticeably. This is an expected limitation of the current checkpoint.

However, I would not attribute the advantage of diffusion-prior-based methods mainly to “better pre-training knowledge” in this case. Their stronger behavior at different resolutions is largely because those models are trained or fine-tuned with multi-resolution / larger-resolution data. This is also something our method can support in future training, so we do not consider it a fundamental technical limitation.

For the current checkpoint, if you need to process larger images, tiled inference is recommended. Directly feeding much larger images may lead to flat textures or artifacts, especially for the multi-step model.

We may add larger-resolution or multi-scale training in future releases.

[minh-nguyenhoang]

Wish you could overcome the overly flat look on most non-diffusion models in the future release. IMO this is the most critical thing compared to other diffusion model like HYPIR or SUPIR, while sometimes they do alter the object significantly, we can't deny the advantage of superior textual quality, while only giving up small fidelity.

[cswry]

Thanks for sharing your thoughts. The T2I-based restoration methods you mentioned can produce stronger textures, largely due to the priors from large pre-trained T2I backbones.

Our goal is stable image enhancement/restoration, where fidelity and avoiding text/object changes are important.

We currently train at a much lower cost, roughly ~1/10 of SD2-scale training, and will continue scaling to improve perceptual quality while preserving stability.