sd-webui-replacer
sd-webui-replacer copied to clipboard
light-and-ray
Return frame by frame video mode
In the last enormous video update https://github.com/light-and-ray/sd-webui-replacer/pull/94, I've removed all video functions from the main pipeline, so frame by frame has gone. However I will recreate this mode from scratch later!
Any ETA?, My pc is not powerful enough to use the animatediff method, and I used a lot the frame by frame mode, do you know when it will be back or if it is possible to use a previous version?
Of course yo can use an old version, this commit is the last before new video mode https://github.com/light-and-ray/sd-webui-replacer/commit/12e67ebfde5b9afd3075ae8f2026d1e73168e640
Use command git checkout 12e67ebfde5b9afd3075ae8f2026d1e73168e640. For return use git checkout master
Any ETA?, My pc is not powerful enough to use the animatediff method, and I used a lot the frame by frame mode
I can't say when, I code this on inspiration. I think I will implement this, sam2, rife and freeInit in the same time, but idk when
But I can recommend you to try animatediff again, I've noticed that on the last sd-webui version it requires much low vram then before. On default settings on my system (without medvram) now it uses about 8GB. You can try lower batch size and medvram option
Any ETA?, My pc is not powerful enough to use the animatediff method, and I used a lot the frame by frame mode
I can't say when, I code this on inspiration. I think I will implement this, sam2, rife and freeInit in the same time, but idk when
But I can recommend you to try animatediff again, I've noticed that on the last sd-webui version it requires much low vram then before. On default settings on my system (without medvram) now it uses about 8GB. You can try lower batch size and medvram option
I don't use A1111, but forge/reforge, because the XL models don't work for me on A1111
I tried to install Replacer + Animatediff in reforge (An updated version of the old Forge), everything works fine but the only problem I have is that it does not detect the control net model of inpaint to test animatediff, everything else works perfectly, but just that does not detect it.
I know you no longer offer support for forge, but I think you should take a look at the new version, many people are using it right now and I say without fear of being wrong, that is even more popular now than A1111, for its speed, Vram handling and FLUX compatibility, and the daily updates it offers.
This is the error I'm having, I have no idea about programming, but it seems to be having problems loading the inpaint models, which is in the correct folder since the old installation of A1111 I have, if it detects it.
Thanks anyways
Yes, ControlNet had issues in API and Inpainting models in the old Forge. I tried to fix it, but Iliyasivil ignored any PRs (Forge was abandoned). Then he reworked everything, and I don't even want to learn it
because the XL models don't work for me on A1111
What do you mean? I use sdxl in replacer (not video) with no problems. They are just needed to be converted to inpainting, and baked with fixed sdxl vae. https://github.com/light-and-ray/sd-webui-replacer/blob/master/docs/tips.md#how-to-get-an-inpainting-model
Yes, ControlNet had issues in API and Inpainting models in the old Forge. I tried to fix it, but Iliyasivil ignored any PRs (Forge was abandoned). Then he reworked everything, and I don't even want to learn it
because the XL models don't work for me on A1111
What do you mean? I use sdxl in replacer (not video) with no problems. They are just needed to be converted to inpainting, and baked with fixed sdxl vae. https://github.com/light-and-ray/sd-webui-replacer/blob/master/docs/tips.md#how-to-get-an-inpainting-model
I mean in general, with 8gb of vram, the XL models do not work in A1111, it takes me about 2 minutes to generate an image, filling the VRAM and virtual memory, when in forge it uses 6.8GB and takes 25 seconds.
That's why I stopped using A1111 and consequently replacer, even if it was a tool that I loved.
You need --medvram cmd argument. Your GPU uses CPU's ram so it's very slow. Forge sets it automatically
You need --medvram cmd argument. Your GPU uses CPU's ram so it's very slow. Forge sets it automatically
I don't need that command in forge, and it is not selected automatically either, just the optimization is superior and it works perfectly. Forge is proving to be the future of UI's, if you want to stay behind and continue using A1111 is a shame, your tool is very good but I think that as you close the door of forge, it will be forgotten, and it is a real shame.
You are not right, forge selects medvram or lowvram automatically. It writes about selected mode in console. And it can cause problems with Replacer in one more way, for example if you want to use SDXL on 12GB GPU, it forces fullvram, and there is no free space for Segment Anything. I faced with this problem, and it's really nasty. You didn't face because you have 8GB VRAM. Manual setting up vram policy is better as for me, to avoid such ridiculous problems
You are not right, forge selects medvram or lowvram automatically. It writes about selected mode in console. And it can cause problems with Replacer in one more way, for example if you want to use SDXL on 12GB GPU, it forces fullvram, and there is no free space for Segment Anything. I faced with this problem, and it's really nasty. You didn't face because you have 8GB VRAM. Manual setting up vram policy is better as for me, to avoid such ridiculous problems
I don't know how forge will work now, the “med_vram” option is not used anymore, I use stability matrix and the option doesn't appear, it automatically selects this option:
Also now you have a selector in the interface that you can determine yourself how much Vram you want to use, and it automatically assigns you the optimal one, so I think that option is no longer used.
I understand that when you have a pc with 12GB+ of VRAM both programs will work similar and you don't care which pc has the rest, but for people who have 8GB OR LESS, forge is the only solution, and believe me, we are the vast majority of people who use these applications.
Still, here are tests of the 3 modes so you can see the difference in someone who has 8GB and the reason why using A111 is absurd.
-A1111 without “med_vram”.
(xd)
-A1111 with “med_vram”.
-Forge with default options
I think the difference is more than obvious.
A1111 has not been updated for more than 2 months, it lacks many options that forge is implementing in a record time, like flux compatibility and its loras.
Forge is updated almost daily several times fixing bugs, improving compatibility and performance and incorporating new improvements, something that a1111 does not do since its development, although it was the first, today leaves much to be desired.
And forge....
I understand that you don't want to “reprogram” the extension because it is a lot of work, but believe me, time is proving that forge is going to end up being the king of image generators, and it is a pity that your extension doesn't work.
About controlnet that you commented yesterday, the author of forge has said that in the next few days he will be completely rewriting the integrated control net extension, to update it and support more things like controlunion, who knows, maybe that will make it easier for you to work on it.
And my comments are not to offend you or anything, I just would like you to reconsider (again), try to make the extension compatible with forge, because I think it is one of the best extensions that exist, and when it is compatible with SAM2, even more.
Nothing else, sorry if you are bothered by anything I have said and I hope everything goes great.