2048px 效果放大问题

[xilai0715]

非常感谢开源你们的工作。在测试过程中发现，恒定尺寸在10241024 15361536的效果都很不错，但是放大到20482048时出现了非常明显边缘线条的问题，是我弄错了什么参数吗？
（从左到右分别为：原图，10241024，15361536，20482048）

[cxzhou35]

I found the same problem when using the model on my data.

[Owen718]

目前训练data size是1024x1024的，所以2048x2048泛化是做不了的，flux本体本身在2048x2048下生成能力也够呛。我们已经在准备数据scale到2k修复和4k修复了。可以进群交流，感谢关注！

[FurkanGozukara]

@Owen718 SUPIR starts with 1024x1024 base resolution and can even upscale into 16 megapixels with no issues

your comparison is not fair imo since your model is only 1024x1024

[Owen718]

@FurkanGozukara Thanks for your interest in our work, but I have to push back on that conclusion. SUPIR does not produce a 16-megapixel image in a single end-to-end sampling pass; resolutions at ~2K and above are handled via tiling (sliding windows). Current real-world UIR benchmarks use very small LQ inputs (≤256 px). In our paper we compare on the standard, widely used benchmarks for this field, and we re-ran SUPIR and DreamClear under matched, fair inference settings. You can also cross-check the tables in the DreamClear paper—our comparison is, to our knowledge, the most comprehensive to date. Importantly, all comparisons in our paper consider non-tiled inference only.

You’re welcome to extend our model—our code was open-sourced even before the paper’s formal publication, reflecting our commitment to open source. We encourage anyone or any organization to improve it, and we’d be glad to review a pull request focused on higher-resolution inference. For unfounded claims, we reserve the right to respond and refute them.

[FurkanGozukara]

@Owen718 thanks for clarification

i presume you have no future plans for vram optimization and tiling

[Owen718]

@FurkanGozukara Hi, we’ve updated the offload inference code and pre-extracted prompt embeddings, removed T5 and CLIP, and reduced memory usage to 28 GB.

[FurkanGozukara]

@Owen718 thanks a lot

i was confused ignore my previous reply :D - if got email notification

28 GB still big but i may test quality on rtx 5090 maybe do more optimization myself if good

but most importantly we need upscale like SUPIR with tiling

[Vigilence]

I am trying the comfyui node for this model, but all outputs are really bad. Am I missing something? I even resized the image to 1024 px.

[FeiSong123]

你好，结合UltraFlux的VAE，现在可以做到2K修复啦，欢迎大家体验

Hey! Thanks to UltraFlux's VAE, we've got 2K restoration ready. Feel free to try it out!