FlashVSR dimensions

[PGCRT]

Hi, is there a max resolution or specific things to know about FlashVSR dimensions? At 2x (here 2048px on longer side) I got some image shift, I can see it on the inference preview. This doesn't happen at 1.5x wich is around 1536px on longer side, but don't really get benefits from the upscale at this resolutin

WanVideoWrapper_FlashVSR_00005.mp4

[RuneGjerde]

Oh i didnt even notice this was added ;-) and it works very well. Restored some old "crappy" LTX videos to new heights (pun intended)... :-p

Really nice upscaler and so fast.

So far i just did 2x, so those old LTX videos didnt go beyond 2048px.
Will give it a try to upscale beyond 2048px.... (if my gpu is capable)

As far as "restoring" low quality small videos its works really well on a few i tested

WanVideoWrapper_FlashVSR_00002.mp4

[RuneGjerde]

2496px wide.. seemed ok here, at least the first video i tried (will try height too)

WanVideoWrapper_FlashVSR_00002-audio.1.1.mp4

[RuneGjerde]

ah yes, with more than 2048px height, there is a mirroring effect. Not sure if its a limitation of the model or not.

WanVideoWrapper_FlashVSR_00001.1.mp4

(compressed the video after render, for github to allows upload and embed with its max 10mb)

[PGCRT]

Yeah I have this issue when using higher side than 1600px, unfortunately there is alot of motion artifacts

[zhuang2002]

Hi, is there a max resolution or specific things to know about FlashVSR dimensions? At 2x (here 2048px on longer side) I got some image shift, I can see it on the inference preview. This doesn't happen at 1.5x wich is around 1536px on longer side, but don't really get benefits from the upscale at this resolutin

WanVideoWrapper_FlashVSR_00005.mp4

Hi @PGCRT, thanks for reporting — I’m the original author of FlashVSR 👋

The “frame shift / reflection artifacts” you’re seeing above ~1600–2048 px aren’t an inherent limitation of the model. They come from the ComfyUI node implementation, which omits our Block-Sparse Attention and Locality-Constrained Sparse Attention (LCSA) modules.

In our official version, LCSA restricts each query to attend only within a local spatial window. This design prevents the periodicity of positional encodings (e.g., RoPE) from causing texture repetition or aliasing at large resolutions, keeping high-res inference stable and consistent with training conditions.

The current ComfyUI node uses plain dense attention without LCSA, so artifacts appear once the positional range exceeds what the model was trained on. When proper LCSA + block-sparse attention are enabled (as in our paper), these artifacts disappear — see the “Locality-Constrained Sparse Attention” section and Ablations 1 video on our project page for comparisons:
🔗 https://zhuang2002.github.io/FlashVSR/

💻 https://github.com/OpenImagingLab/FlashVSR

Recommendation:
I strongly recommend running inference with the official FlashVSR code, which includes the correct LCSA and sparse attention kernels.
Alternatively, you can send me a short sample video — I’d be happy to run it through the original FlashVSR pipeline and share the enhanced result for your preview.

[zhuang2002]

ah yes, with more than 2048px height, there is a mirroring effect. Not sure if its a limitation of the model or not.

WanVideoWrapper_FlashVSR_00001.1.mp4
(compressed the video after render, for github to allows upload and embed with its max 10mb)

Hi @RuneGjerde, glad to hear FlashVSR worked well for your LTX videos — really appreciate the feedback! 🙏

Yes, the “reflection” you noticed beyond ~2K height is also due to the missing Locality-Constrained Sparse Attention (LCSA) in the current ComfyUI node.
With LCSA and block-sparse attention (as in our official repo), this issue disappears completely.

If you’d like, feel free to send me one of your test clips — I can run it through the original FlashVSR pipeline and share the corrected high-res result for comparison.

[RuneGjerde]

@zhuang2002 I tried a lot of "normal size" videos, and it works really well, and super fast. As for the extreme resolutions, maybe something that can be fixed in the repro here, if Kijai got the time, and its not something too complicated to implement.
And for more "normal" restoration and upscale (not going beyond 2K), it already works great.

I'll try some more. It works specially well for videos generated with a bit low quality, like I did a while back when trying LTX (no bad words about LTX, it was just my PC at the time that had a very slow gpu ;-). The "restoring" part is impressive.
As well as the regular upscaling.

Yes, the “reflection” you noticed beyond ~2K height is also due to the missing Locality-Constrained Sparse Attention (LCSA) in the current ComfyUI node. With LCSA and block-sparse attention (as in our official repo), this issue disappears completely.

@kijai maybe something for you

[kijai]

@zhuang2002 I tried a lot of "normal size" videos, and it works really well, and super fast. As for the extreme resolutions, maybe something that can be fixed in the repro here, if Kijai got the time, and its not something too complicated to implement. And for more "normal" restoration and upscale (not going beyond 2K), it already works great.

I'll try some more. It works specially well for videos generated with a bit low quality, like I did a while back when trying LTX (no bad words about LTX, it was just my PC at the time that had a very slow gpu ;-). The "restoring" part is impressive. As well as the regular upscaling.

Yes, the “reflection” you noticed beyond ~2K height is also due to the missing Locality-Constrained Sparse Attention (LCSA) in the current ComfyUI node. With LCSA and block-sparse attention (as in our official repo), this issue disappears completely.

@kijai maybe something for you

I'm aware, as it states in the commit and in the workflow, this is only partial implementation of FlashVSR. Implementing all of the attention code etc. is probably out of scope for this repository and would rather make more sense as it's own stand-alone implementation, or like mentioned, can just use the original code.

[RuneGjerde]

No worries, its already quite usable as it is. For restoring and upscaling reasonably.
and extremely fast

Perfect for adding more fine-grain details to a video you have already generated, and at the same time upscaling it, in a fast and easy way:

WanVideoWrapper_FlashVSR_00003.mp4

[zhuang2002]

Hi, is there a max resolution or specific things to know about FlashVSR dimensions? At 2x (here 2048px on longer side) I got some image shift, I can see it on the inference preview. This doesn't happen at 1.5x wich is around 1536px on longer side, but don't really get benefits from the upscale at this resolutin

WanVideoWrapper_FlashVSR_00005.mp4

Hi @PGCRT,
I just ran the sample clip you provided — here are the results from the official FlashVSR pipeline (with proper Block-Sparse and Locality-Constrained Sparse Attention).
It runs faster, produces cleaner outputs, and needs no pick or adjustment.

issue0_concat.mp4

Due to GitHub’s file upload limit, I had to compress the video, but you can still clearly see that there’s no frame shift or reflection.

Everything remains stable even beyond 2K resolution.
Also, huge thanks to @kijai for the super-fast ComfyUI integration — really appreciate the quick support and collaboration! 🙏

[kijai]

Everything remains stable even beyond 2K resolution. Also, huge thanks to @kijai for the super-fast ComfyUI integration — really appreciate the quick support and collaboration! 🙏

Thanks for the model and the code! Sorry for not doing the full implementation at this point, it's just a lot of code to integrate into my code so I'm not sure I'm able to do that, I can still take a further look into it though.

[PGCRT]

@zhuang2002 Thanks for the reply, really appreciated. I did try inference from the original code, and it took me a bit more than an hour to compile the attention method. I don't know if there are pre-built wheels for it.
Unfortunately, I didn't manage to run the inference. I didn't conserve the console logs but I will try again for sure.

I would like to make a simple GUI for it, so I can run it as a desktop executable, could be fun to do.

[PGCRT]

@zhuang2002 I managed to made a simple GUI for FlashVSR, I think it's working great, I tried to render at 2048px, the inference is ok (with some screen stutters), but OOM happens during the vae decode, I think I need a little bit more VRAM (I run a 4090), I wonder how to reduce the usage. I tried tiling but didn't manage to avoid OOM, or maybe reduce the VAE dtype to fp16 but apparently this could create a missmatch, I will see if I can find a way to reduce the VRAM usage. But It works thanks

https://www.youtube.com/watch?v=7Nbn0CSJ3oM

[zhuang2002]

@zhuang2002 I managed to made a simple GUI for FlashVSR, I think it's working great, I tried to render at 2048px, the inference is ok (with some screen stutters), but OOM happens during the vae decode, I think I need a little bit more VRAM (I run a 4090), I wonder how to reduce the usage. I tried tiling but didn't manage to avoid OOM, or maybe reduce the VAE dtype to fp16 but apparently this could create a missmatch, I will see if I can find a way to reduce the VRAM usage. But It works thanks

https://www.youtube.com/watch?v=7Nbn0CSJ3oM

Hi everyone,
I talked with a community friend today and he shared a very helpful test case that makes the differences between implementations much more intuitive. Here is a simple visual comparison:

Fig.1 – LR Input	Fig.2 – Dense Attention (no LCSA)	Fig.3 – Original FlashVSR (with LCSA)
issue1_LR.mp4	FlashVSR_HR_woLCSA.mp4	FlashVSR_HR_wLCSA.mp4

As we can see, the original FlashVSR pipeline delivers noticeably higher image quality, mainly due to:

• Better preservation of fine structures and details
• Effectively avoiding texture aliasing and visual artifacts

I will work on a version that does not rely on the block-sparse attention library, while still matching the visual quality of the original pipeline with LCSA.
However, please note that this version may not reach the same speed as the optimized block-sparse implementation used in our paper.

I’ll keep you updated — stay tuned

[kijai]

I will work on a version that does not rely on the block-sparse attention library, while still matching the visual quality of the original pipeline with LCSA. However, please note that this version may not reach the same speed as the optimized block-sparse implementation used in our paper.

I’ll keep you updated — stay tuned

Hey, thanks, that would be helpful indeed!

A question for you if you don't mind: In your code the prompt is loaded as already encoded from .pt file, can you tell what the actual prompt used is, or is it relevant at all?