Welcome to my 💫🔼 Friendly FLASHVSR Super-Resolution
✨ Less mess, more magic
🆕 FLASHVSR v1.1 support added - enhanced stability + fidelity
⚡FLASHVSR is a new single-step 4K video upscaler that restores detail even without increasing resolution. It also works with long videos more than 1 minute length. Topaz Starlight now has a strong competitor! According to the developers, the upscaler can generate in real time on an Nvidia A100. This is all thanks to its unique architecture and a modified Wan 1.3b model.
At the same time, it runs quite fast even on low-end graphics cards and offers flexible settings.
Video upscaling to 4K works even with 10 GB of video memory! (tested on the 7-second video).
📌 6GB of VRAM and 32GB RAM (tested on RTX 4050 Laptop) are sufficient for operation, but the more the better.
If you have 16GB+ VRAM use 4x upscale for better quality
For less than 12GB VRAM 2x upscale recommended for first run, and 2x upscale again after. Finally it will be 4K video. 2x upscale is much lighter even with high resolution input videos
For Low VRAM use mode "tiny" and "tiny long" (for long videos). You can use "full" mode with better quality but it's slower. You can switch on "tiled vae" for decrease VRAM peaks
📌 Detailed tips and links in the workflow
✨ Workflow features:
User-friendly interface with detailed tooltips and links to everything you need to use the upscaler
Ability to downscale the source video to reduce system load
Ability to upscale long videos (longer than 1 min)
Post-processing: interpolation, adding grain, brightness, contrast control, stylization, and much more
FLASHVSR requires Triton and SageAttention installed. It's also important to install requirements.txt from the node folder
🤗🙏🏼 Thanks to OpenImagingLab group and lihaoyun6
Original repo — GitHub
Description
added new FLASHVSR v1.1 model support (you can download new version from Links section in the workflow)
fix frame count according to video input
added links to new Lightx2v alternative VAE decoders with better perfomance (not working in Full mode for now; waiting for node update)
FAQ
Comments (16)
Hi. The small version works for me, but when I select the full version I get 'NoneType' object has no attribute 'model on the FlashVSR Init Pipeline. Could I be missing some models..?
Btw, the lihaoyun6/ComfyUI-FlashVSR_Ultra_Fast node seems to work fine with the Full version. But that's a different node I'm using in a separate simplified workflow.
Ah it seems I have to pick a different VAE in the alt_vae parameter, then it works. Cheers!
@drey08 Hi! Yes, alternative VAE works only in tiny and tiny long modes for now
Hello!
When i run workflow im getting error in the end of resizing:
Exception in callback ProactorBasePipeTransport.call_connection_lost(None)
handle: <Handle ProactorBasePipeTransport.call_connection_lost(None)>
Traceback (most recent call last):
File "asyncio\events.py", line 84, in _run
File "asyncio\proactor_events.py", line 165, in callconnection_lost
ConnectionResetError: [WinError 10054] Удаленный хост принудительно разорвал существующее подключение
Exception in callback ProactorBasePipeTransport.call_connection_lost(None)
handle: <Handle ProactorBasePipeTransport.call_connection_lost(None)>
Traceback (most recent call last):
File "asyncio\events.py", line 84, in _run
File "asyncio\proactor_events.py", line 165, in callconnection_lost
ConnectionResetError: [WinError 10054] Удаленный хост принудительно разорвал существующее подключение
Exception in callback ProactorBasePipeTransport.call_connection_lost(None)
handle: <Handle ProactorBasePipeTransport.call_connection_lost(None)>
Traceback (most recent call last):
File "asyncio\events.py", line 84, in _run
File "asyncio\proactor_events.py", line 165, in callconnection_lost
ConnectionResetError: [WinError 10054] Удаленный хост принудительно разорвал существующее подключение
Can you help with it?
Привет! Почему-то браузер прерывает соединение с сервером ComfyUI. Может случаться в т.ч. из-за нехватки видеопамяти. На какой именно ноде ошибка происходит? И только на этом воркфлоу такая проблема? И какая видеокарта?
@RusselX 4070 ti 12gb
Прерывание происходит на ноде Image Concatenate Multi и только там...
После нескольких попыток "подключиться" (хз, у меня комфи локально крутится) флоу идет дальше...
Но, в любом случае я пытаюсь получить хоть какой-то внятный результат, но все генерации\апскейлы имеют отвратительное качество...
@kalavatz103 Видяха хорошая, по крайней мере дело в ней не должно быть. Вообще Image Concatenate считается на процессоре и расходует мало VRAM и RAM, поэтому странно. Как вариант могу предложить запустить в другом браузере. На крайний случай можно отключить ноду Image Concatenate Multi, а к ноде Video Compare протянуть узел от исходного видео. Таким образом можно будет всё равно видеть изменения рядом друг с другом. Как я понимаю сама генерация FlashVSR выводится?
Насчёт качества FlashVSR это конечно не идеальная волшебная палочка, но по сравнению с другими апскейлерами, которые просто накидывают шарпинга, этот частично восстанавливает детали
@RusselX спасибо за развернутый ответ. Да вот странно, что в локальной сборке pinokio апскейлер выдает норм качество О-о
@kalavatz103 Пожалуйста) Портабельный комфи просто капризный, чуть что и можно вообще всё навернуть, например, после установки каких-нибудь новых зависимостей
Works fine. Thanks a lot!
You are welcome 🤗🙏🏼
I always get a black video or a memory error (without downsizing). Any ideas what could be the reason for the black video?
Hi! As for memory error it depends of your input video size. If it's too big so you need to set downscale to avoid memory errors. If you have less then 12GB VRAM it is recommended 2x max scale specifically to 2K resolution like 2560×1440 or 2048×1080 or less if you still have problems
@RusselX Yeah that I already know. But why I got only black videos? Even without memory error.
@Pressydent It's hard to say right away. Do you get any errors during generation in the terminal? And did you update ComfyUI and all nodes in the workflow? It is also important to make sure that the models are placed in the correct location, be guided by models storage locations in the links section of the workflow. Also you can try reupload models because they could have been damaged during downloading