roop unleashed Gradio - WIP Thread

[C0untFloyd]

For the brave and adventurous among you, I just made the gradio branch public.

I think you're in for a treat because this has several cool new additions, to name just a few:

• Drag'n Drop Support of Images/Files
• Scrubbing through videos to select faces from frames or just preview the deepfake with current settings
• Clipseg masking to restore stuff from the original image
• Live Fake Webcam (select your new face in the tab before)
• Lots of sliders for the dynamic settings like blend ratio
• Proper GIF Support
• Faster processing through memory processing, no temp file writing anymore

Be aware that there is still lots of work to be done, some stuff might break, it's in no way optimized or clean code. However in the future the gradio version will replace the main branch. Before that happens I will create another small QoL and bugfixed version of the TkInter Version (without the new stuff), just in case people want to continue using that one for dubious reasons.

To checkout:
git clone -b gradio --single-branch https://github.com/C0untFloyd/roop-unleashed.git
cd roop-unleashed
conda create -n mynewshinyroop
conda activate mynewshinyroop
pip install -r requirements.txt

Run it with the same parameters as previously e.g. python run.py --execution-provider cuda --video-encoder libvpx-vp9
For the GUI version you don't need the frame processor parameters anymore.

To update to newer versions:

cd roop-unleashed
conda activate mynewshinyroop
git pull
pip install -r requirements.txt

Please limit posts about this version to this thread only, thank you!

Google Colab https://github.com/C0untFloyd/roop-unleashed/blob/gradio/roop-unleashed.ipynb

Changelog & Issues:

Known Issues

• GPU Setting not used, still uses CLI-Arguments
• Using Controls in parallel leads to undetermined behaviours
• CLI needs revamp
• Memory Limit currently ignored
• Adding destination files not possible afterwards
• Re-choosing the same input image doesn't open face selection
• Source/Destination faces selected for a whole batch
  - Defaults for public server & GPU might be problematic Extra colab config should solve this
• Very slow live swapping as gradio doesn't support streaming to controls other than webcam
• Huge list of destination files destroys layout (issue created @ gradio)
• Big images not displayed in preview container (seems to be gradio issue as well)
• DMDNet not working with higher resolutions/face bboxes
• DMDNet Specific missing implementation
• Lots of extras on the Extras Tab missing
• Missing progress display in GUI
• Progress display bug lately with video swapping without keeping frames

v2.6.9 2023-08-09

Bugfix: Deleting wrong videofile if no audio (thanks to @eniora for finding)
Added det_size checkbox to toggle between default and 320x320 size (has to be currently used BEFORE any swapping action!)
Added Mask Preview to text masking
Added colab config to have public server defaulting to false
Some layout changes
More code cleanup

v2.6.8 2023-08-08

Bugfix: Possible Memory Leak
Swapping multiple faces now possible
Added selection of multiple faces from same frame until done
Added setting for live-cam activate from start
Better handling of changing settings and writing cfg
Added most missing arguments to settings
Provider,Max. Threads, Video Quality and Encoder can be changed on the fly now

v2.6.5 2023-08-06

Bugfix: Clear output toggle
Checkbox & Dropdown values in settings are applied at selection
Added Refresh Button to Preview (for re-applying changes made)
Added cutting of videos to Extras Tab

v2.6.4 2023-08-05

Bugfix: Empty Frames in Preview
Bugfix: Torch not using GPU
Default config should have public server enabled for colab
Added checkbox for permanently scrubbing with faked content
Added DMDNet generic version, currently works only for low quality faces smaller than 512x512.

v2.6.2 changes 2023-08-03

Bugfix: GFPGAN not using GPU
Bugfix: Clear output toggle not working
Bugfix: Num threads ignoring arguments
Bugfix: Blend Ratio
Added face enhancement only processing faces bbox for speed increase

v2.6.0 changes 2023-08-02

Bugfix: CodeFormer missing upscaler
Bugfix: Slider values not being used
Restore Audio re-added
Added Google Colab
Setting for output clearance
Error handling for target faces not selected
txt2clip using cuda when available

v2.5.4 changes 2023-07-28

Bugfix: Removed upscaler from CodeFormer
Bugfix: Removed txt2clip from default chain
Bugfix: Missing some global vars
Added Settings for Image & Video Output Formats
Layout change in Settings
Extract Frames from Video implemented
Info Messages for small Utility Functions e.g. Clean Temp

[RealCalumPlays]

Live camera doesnt work. Heres the traceback:

Traceback (most recent call last): File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\gradio\routes.py", line 442, in run_predict output = await app.get_blocks().process_api( File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\gradio\blocks.py", line 1389, in process_api result = await self.call_function( File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\gradio\blocks.py", line 1094, in call_function prediction = await anyio.to_thread.run_sync( File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\anyio\to_thread.py", line 33, in run_sync return await get_asynclib().run_sync_in_worker_thread( File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\anyio\_backends\_asyncio.py", line 877, in run_sync_in_worker_thread return await future File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\anyio\_backends\_asyncio.py", line 807, in run result = context.run(func, *args) File "C:\Users\Calum\.pyenv\pyenv-win\versions\3.10.11\lib\site-packages\gradio\utils.py", line 703, in wrapper response = f(*args, **kwargs) File "C:\Users\Calum\Documents\roop-unleashed\roop\ui.py", line 497, in on_stream_swap_cam current_cam_image = live_swap(camimage, "all") TypeError: live_swap() missing 2 required positional arguments: 'use_clip' and 'clip_text'

[C0untFloyd]

Thanks, was missing the last minute clip parameters. Fixed now, please do a git pull.

[tfis2]

First:
Thank you very much for your work!!!
It's great that you can drag and drop files, even *.png with multiple faces.

Clip Masking leads to an error:

Traceback (most recent call last):
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\routes.py", line 442, in run_predict
output = await app.get_blocks().process_api(
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\blocks.py", line 1389, in process_api
result = await self.call_function(
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\blocks.py", line 1094, in call_function
prediction = await anyio.to_thread.run_sync(
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio\to_thread.py", line 33, in run_sync
return await get_asynclib().run_sync_in_worker_thread(
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio_backends_asyncio.py", line 877, in run_sync_in_worker_thread
return await future
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio_backends_asyncio.py", line 807, in run
result = context.run(func, *args)
File "C:\Users\info\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\utils.py", line 703, in wrapper
response = f(*args, **kwargs)
File "D:\StableDiffusion\roopunl\beta\roop-unleashed\roop\ui.py", line 448, in start_preview
current_frame = live_swap(current_frame, roop.globals.face_swap_mode, use_clip, clip_text)
File "D:\StableDiffusion\roopunl\beta\roop-unleashed\roop\core.py", line 202, in live_swap
temp_frame, _ = roop.globals.IMAGE_CHAIN_PROCESSOR.run_chain(frame,
File "D:\StableDiffusion\roopunl\beta\roop-unleashed\chain_img_processor\image.py", line 85, in run_chain
img = self.processors_objects[proc_id][thread_index].process(img,params)
KeyError: 'txt2clip'

---

When scrubbing "Preview face swap" is off. Only works with original face.

[jmp909]

some quick observations (Windows 11, RTX 3080)

• when doing a batch image convert of say 400 images, the terminal output stops at 8/400 images (maybe related to 8 execution threads?) for the rest of the process, and never updates past that even though the swapping process is continuing in the background

• would help if Target file(s) list was a fixed-height scrollable list because currently it stretches out the whle UI vertically with a lot of files - making it harder to get to eg "Use face from this frame"

• for images would be useful to allow either eg "Keep original file format" or an option for the save file type. My 100K JPGs are being re-saved as swapped 2MB PNG's, but JPG would be fine and smaller.

thanks

[TrickleDownClown]

Awesome. I'm looking forward to trying it. I'll report back tomorrow.

[maxbizz]

This is the error its throwing btw im using venv instead of conda

"Traceback (most recent call last):
 File "G:\Github\roop\venv\lib\site-packages\gradio\routes.py", line 442, in run_predict
   output = await app.get_blocks().process_api(
 File "G:\Github\roop\venv\lib\site-packages\gradio\blocks.py", line 1389, in process_api
   result = await self.call_function(
 File "G:\Github\roop\venv\lib\site-packages\gradio\blocks.py", line 1094, in call_function
   prediction = await anyio.to_thread.run_sync(
 File "G:\Github\roop\venv\lib\site-packages\anyio\to_thread.py", line 33, in run_sync
   return await get_asynclib().run_sync_in_worker_thread(
 File "G:\Github\roop\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 877, in run_sync_in_worker_thread
   return await future
 File "G:\Github\roop\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 807, in run
   result = context.run(func, *args)
 File "G:\Github\roop\venv\lib\site-packages\gradio\utils.py", line 703, in wrapper
   response = f(*args, **kwargs)
 File "G:\Github\roop\roop\ui.py", line 419, in start_swap
   batch_process([file.name for file in target_files], use_clip, clip_text)
 File "G:\Github\roop\roop\core.py", line 266, in batch_process
   roop.globals.BATCH_IMAGE_CHAIN_PROCESSOR.run_batch_chain(imagefiles, imagefinalnames, 8, processors, params_gen_func)
 File "G:\Github\roop\chain_img_processor\batchimage.py", line 85, in run_batch_chain
   future.result()
 File "C:\Users\xyz\AppData\Local\Programs\Python\Python310\lib\concurrent\futures\_base.py", line 451, in result
   return self.__get_result()
 File "C:\Users\xyz\AppData\Local\Programs\Python\Python310\lib\concurrent\futures\_base.py", line 403, in __get_result
   raise self._exception
 File "C:\Users\xyz\AppData\Local\Programs\Python\Python310\lib\concurrent\futures\thread.py", line 58, in run
   result = self.fn(*self.args, **self.kwargs)
 File "G:\Github\roop\chain_img_processor\batchimage.py", line 62, in process_frames
   resimg, _ = self.run_chain(temp_frame, params, self.chain)
 File "G:\Github\roop\chain_img_processor\image.py", line 85, in run_chain
   img = self.processors_objects[proc_id][thread_index].process(img,params)
 File "G:\Github\roop\plugins\plugin_faceswap.py", line 47, in process
   frame = swap_face(params["input_face_datas"][0], face, frame)
 File "G:\Github\roop\roop\face_helper.py", line 90, in swap_face
   return get_face_swapper().get(temp_frame, target_face, source_face, paste_back=True)
 File "G:\Github\roop\venv\lib\site-packages\insightface\model_zoo\inswapper.py", line 50, in get
   latent = source_face.normed_embedding.reshape((1,-1))
AttributeError: 'NoneType' object has no attribute 'normed_embedding'
Traceback (most recent call last):
 File "G:\Github\roop\venv\lib\site-packages\gradio\routes.py", line 442, in run_predict
   output = await app.get_blocks().process_api(
 File "G:\Github\roop\venv\lib\site-packages\gradio\blocks.py", line 1389, in process_api
   result = await self.call_function(
 File "G:\Github\roop\venv\lib\site-packages\gradio\blocks.py", line 1094, in call_function
   prediction = await anyio.to_thread.run_sync(
 File "G:\Github\roop\venv\lib\site-packages\anyio\to_thread.py", line 33, in run_sync
   return await get_asynclib().run_sync_in_worker_thread(
 File "G:\Github\roop\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 877, in run_sync_in_worker_thread
   return await future
 File "G:\Github\roop\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 807, in run
   result = context.run(func, *args)
 File "G:\Github\roop\venv\lib\site-packages\gradio\utils.py", line 703, in wrapper
   response = f(*args, **kwargs)
 File "G:\Github\roop\roop\ui.py", line 274, in remove_selected_input_face
   if len(input_thumbs) > SELECTED_INPUT_FACE_INDEX:
NameError: name 'SELECTED_INPUT_FACE_INDEX' is not defined
"

[C0untFloyd]

when doing a batch image convert of say 400 images, the terminal output stops at 8/400 images
or images would be useful to allow either eg "Keep original file format" or an option for the save file type.
NameError: name 'SELECTED_INPUT_FACE_INDEX' is not defined
KeyError: 'txt2clip'

Fixed or added in the latest update.

When scrubbing "Preview face swap" is off. Only works with original face.

It's meant to position the video to a frame you want to check out. Live-swapping with all options turned on is way too slow (at least on my system).

would help if Target file(s) list was a fixed-height scrollable list because currently it stretches out the whle UI vertically with a lot of files - making it harder to get to eg "Use face from this frame"

Of course! I'd like this for every multifile box but this is currently unsupported by gradio. I'm trying to hack this in with CSS/HTML trickery like automatic1111 did with a lot of components but without luck so far. I also had bugs with lots of input files, there seems to be a magic limit of ~400 input files.

[TrickleDownClown]

Forgive my silliness but during the install I get the following:
bash: conda: command not found

This happens when doing the conda commands in the install intructions. Do I need conda in addition to python?

[C0untFloyd]

Anaconda/Miniconda is a way to isolate your Python Installations. This is always a good idea because otherwise you will have many version conflicts. The 1-click installer does this automatically for you but the branch does not of course. If you don't want/need that you can start right away with pip installing the requirements but I'd suggest installing Miniconda

[TrickleDownClown]

I'll do that. Thanks.

[eadnams22]

I’ve been exploring adding support for more ffmpeg codecs (ProRes for more lossless workflows, for example), should I do that for this gradio branch and not the main, or would that port over easily enough if I pull from main to add it?

[C0untFloyd]

The gradio branch will sooner or later become the main branch. But adding ffmpeg codecs is just a matter of extending the argument list so I think it's your choice really.

[eadnams22]

Yeah, that’s what it looked like.

Ideally, selecting the codec (and codec options) from within gradio would probably make the most sense, eventually, but that’s a bit beyond my abilities.

[eadnams22]

Only real issue I've encountered is hardcoded .mp4 files for the temp files, which causes ffmpeg to quit if the codec does not support the mp4 container (like prores).

If the TEMP_FILE variable in utilities.py could grab the extension from the output/target file automatically, that would fix it, but that's a bit beyond my python skills.

All this seems like it could be handled on the gradio side settings, eventually.

[eniora]

Thanks for this. First of all, I understand this is WIP so this post will serve as improving the future of Roop Unleashed.

Many issues/recommendations so far:

1. The audio isn't restored after swapping and the frame selection doesn't work most of the time (the slider is stuck and doesn't move at all), takes many tries of clicking on "Preview face swap" until it works. Edit: frame selection can be fixed by clicking on the video file name after adding it. Anyway please fix the audio not working after swapping.

2. It would be nice if we could choose the video output quality in settings. Current setting makes the video much larger than the original.

3. Can you make it that creating a new video doesn't overwrite the old one in the output folder? like if I create a video from a file named 1.mp4 and that file already exists on output folder, another file is created named 1_1.mp4 etc.
   Another edit: it actually deletes any file in the output folder regardless of its name after creating another video, please also fix this :D

4. Checking the keep frames option it doesn't swap anything.

5. Is it possible to make it faster and use the GPU instead of the CPU when using the masking (use text to clip masking) option? (Edit: this option isn't working, tried it and the video is the exact same as if not using it)

6. I noticed that using the enhancer (GFPGAN) on the main unleashed 2.0.3 is faster compared to this (5-6 FPS compared to ~4 FPS on the same video and face and same threshold settings)

7. Codeformer not working, here's what I get when trying to use it:

File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\routes.py", line 442, in run_predict
    output = await app.get_blocks().process_api(
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\blocks.py", line 1389, in process_api
    result = await self.call_function(
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\blocks.py", line 1094, in call_function
    prediction = await anyio.to_thread.run_sync(
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio\to_thread.py", line 33, in run_sync
    return await get_asynclib().run_sync_in_worker_thread(
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio\_backends\_asyncio.py", line 877, in run_sync_in_worker_thread
    return await future
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\anyio\_backends\_asyncio.py", line 807, in run
    result = context.run(func, *args)
  File "C:\Users\owner\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\utils.py", line 703, in wrapper
    response = f(*args, **kwargs)
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\roop\ui.py", line 439, in start_swap
    batch_process([file.name for file in target_files], use_clip, clip_text)
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\roop\core.py", line 283, in batch_process
    roop.globals.VIDEO_CHAIN_PROCESSOR.run_video_chain(v,videofinalnames[index], fps, 8, processors, params_gen_func, roop.globals.target_path)
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\chain_img_processor\video.py", line 60, in run_video_chain
    frame_processed, params = self.run_chain(frame,params,chain)
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\chain_img_processor\image.py", line 85, in run_chain
    img = self.processors_objects[proc_id][thread_index].process(img,params)
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\plugins\plugin_codeformer.py", line 51, in process
    temp_frame = inference_app(img, options.get("background_enhance"), options.get("face_upsample"),
  File "C:\Roop-unleashed\_Gradio\roop-unleashed\plugins\codeformer_app_cv2.py", line 219, in inference_app
    face_upsampler = upsampler if face_upsample else None
NameError: name 'upsampler' is not defined

[C0untFloyd]

1. You need to select the file to scrub as you can have multiple files. Yes, audio currently isn't added.
2. This is still possible through commandline args
3. Actually this is a feature not a bug 😃 . Remember that you're effectively running a server now and other people can download the stuff you put into temp and output. I might add a setting for this though.
4. Need verify this, you're talking about videos aren't you?
5. Sure it's possible, should be a setting. It is definitely working, but it's only active if a face was found in the input frame.
6. Ok
7. Needs fixing, CodeFormer is very slow anyway I only use it for images.

[eniora]

1. Ah ok.
2. I see, so the commandline for quality ranging 1-51 is still enabled? Anyway I think it would be handy to have it directly in the settings UI.
3. Would be really appreciated if there's a setting for it, I (and most of us I believe) always run my servers locally (as is the case for SD web UI for example), never enabled sharing or anything. So yah default for the current state and an optional UI setting to enable automatic naming without deleting anything in the folder.
4. Ye videos. When using keep frames it finishes in a second and doesn't swap anything, just creates identical video of the original in the output folder.
5. Hmmm pretty sure I put a clear, high quality face as input frame and it swapped it successfully in the output video but it didn't mask anything (I put "eyes" as input text), will try it again.
6. At first I thought it's a placebo effect but tried again and it's a bit slower, not a big deal though.
7. I always preferred GFPGAN but thought it's worth mentioning :)

Cheers!

[sordidloam]

I can confirm that this is the case for my new install. First video saved, then once it started to overwrite the file, the new file it wrote, also gets removed(deleted?). I did try toggling on that setting to clear the output folder as a test, no difference though.

(update) Batch seems to work fine, at least it's enumerating the filenames such that they are unique.

[jmp909]

couple more observations

• dragging the timeline doesn't always show the correct frame (using the number spinner is fine tho)

• not sure what the "input faces" multiple thumbnnails (top left) are for at the moment? dragging them back down into "source face image" breaks because it's trying to use the cropped version to find a face and this doesn't work.. it can't find the face (needs to reference original uncropped image)

• when loading a new source image with multiple faces (after a previous source image), the "Input faces" doesn't update

• audio is missing even though "Skip Audio" is unchecked

realise it's all still WIP though!

thanks

[C0untFloyd]

1. Ok, weird
2. The plan is to have multiple input faces targetting multiple destination faces. Input & Target faces show you the cropped faces. What you probably overlooked: If you're using an image with several faces on it, a custom selection pops up where you have to select the face you want to use (clicking on the button below). Only then it will be used as input face. Same btw. for target faces. If only 1 face was detect in your image/video this will automatically be cropped and used.
3. see 2
4. Yes, will re-add this ASAP

[eniora]

Just tried the live cam option and it's working very nicely even with enhancer, very good addition! :D

[eniora]

@C0untFloyd is it possible for the live cam to be smoother? no difference for me between GTX 1070 and RTX 3090, it's like 2FPS on both (no enhancer or anything running). The GPU usage is low as well, it's not taking good advantage of the GPU. I know there's processing time involved but would it be possible to make the buffer a bit later but the cam output smoother?

[C0untFloyd]

It's the same processor as usual image/video swapping so if it's using your GPU there, it should be the same when live-swapping. It's only 1 thread though. I suspect Gradio is limiting the output frequency because I'm abusing an image component to stream the swapped images to. There's currently no way to create a custom streaming target 😞

[eniora]

Ye it's still using the GPU with the live cam but it doesn't take full advantage of it the same way as if it was swapping a video. But yah if it's only 1 thread then that explains why. Hopefully you can somehow find a workaround. Thanks.

[eniora]

@C0untFloyd regarding the new 2.6.5 update:

1. The Max. Number of Threads, Video Codec, and Video Quality (crf) settings don't make any difference even after applying the settings in the same session, they aren't even wired to the config file (and video quality doesn't have an entry in the config file at all). They also always reset after restarting the server. Same for Image and Video Output Format, they don't make a difference even after applying the settings on the same session and aren't wired to config as well.

2. Clear output folder before each run option now works but still always shows as checked after restarting the server despite it's already unchecked in the config file after unchecking it and saving the settings, it works but the check status is always set to "on" in the settings page and doesn't reflect the actual setting as shown in the config file.

3. I am sure I suggested this before but I will suggest it again 😁 : Please disable public server by default, after updating with git pull or installing from scratch it's always set to on by default. I sure can disable it quickly in settings but IMO it's better if it's set to off in general, both because of security and because it's much slower than running locally (since when it's enabled unleashed automatically launches the public server instead of the local one). People who want it can enable it but I am sure 99% of us don't want it.
   Also it's worth noting that some users will install and then right away start rooping without knowing the public server is enabled and then complain why it's slow (not the main reason why I am suggesting this).

4. (Not really important as there's an easy workaround) but if you upload a source face of someone then remove it and upload it again (same file) it doesn't get detected anymore, you have to upload any another face and then get back to it for the UI to detect it again. This is a bit annoying if you're working with a picture with more than one face (that's how I discovered the issue) because after re-uploading it the faces won't get detected unless you upload any other picture and then get back to it.

Again, I think you may already know most or all of these issues and you're probably only revamping the settings page and the new UI for now but saying all this in case you missed something.
Cheers!

[lysxelapsed]

I can confirm 1-3. Also it ignores the max-memory argument now and sucks up all 12GB of Vmem, leaving 65MB free - wasn't the case with 2.6.2.
@C0untFloyd maybe a "known issues"-section in the changelog would make sense? so people don't have to repeat (and you don't have to read) the same thing over and over.

[C0untFloyd]

1+2 Yes, known issue. Yesterdays version was committed in-between to get rid of the output complains.
3. And my answer is still the same: this and the cuda provider needs to be defaulted for the Colab and Hugging Face Users. If it wouldn't they would need to edit the config file manually, because there's no chance to reach the server. I agree that 99% will use this locally, but with my latest change this has become a non-issue because the config won't get overwritten anymore, your settings once saved will stay the same.
4. Yes Gradio seems to re-use the existing image and it doesn't fire a change event which I'm using.

A known issues entry would be good, however this means even more housekeeping work for me. IMO this would be better for a real release.

[sordidloam]

Where do I add '--share=true' for this to be available locally? Didn't see a launch.py file to edit and add the arguments. @C0untFloyd Tried in my bat file I made to launch with the arguments you recommended we use, didn't work there either.

[lysxelapsed]

Tried "selected face" method with multiple videos (10 vids with about 4GB total):

• Loading the vids took a while, but acceptable
• Preview did not work good. Even for the original, unswapped frame, it took up to a minute to display the newly entered frame number
• Once I got the frame, face recognition and selection worked fine. Selected a face for every video.
• Unfortunately, it didn't swap a single frame. Not even with the first vid (highlighted target face), so I stopped manually after the first three vids.
• Swap preview didn't work at all and produced this error:

File "D:\roop-unleashed\roop\core.py", line 212, in live_swap
temp_frame, _ = roop.globals.IMAGE_CHAIN_PROCESSOR.run_chain(frame,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\roop-unleashed\chain_img_processor\image.py", line 86, in run_chain
img = self.processors_objects[proc_id][thread_index].process(img,params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\roop-unleashed\plugins\plugin_faceswap.py", line 70, in process
if compute_cosine_distance(tf.embedding, face.embedding) <= dist_threshold:
^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'embedding'>

tbh I'm not a big fan of the webgui either. the more videos you add, the slower it gets. not that suitable for batch processing (which is what I'm looking for, but no roop fork seems to have that working with face selection) imo. but don't get me wrong: amazing work regardless!

[C0untFloyd]

Uh oh this seems to get out of hand, without some instructions lots of people seem to interpret things. Well, let's see

Loading the vids took a while, but acceptable
Preview did not work good. Even for the original, unswapped frame, it took up to a minute to display the newly entered frame number

Very probable public server was turned on, otherwise it would load in secs or rather ms

Once I got the frame, face recognition and selection worked fine. Selected a face for every video.

This is where I think you're interpreting things wrong. You seem to believe you can select a face for each video in your queue? That's not the case and multi-swapping also doesn't work (yet).

Unfortunately, it didn't swap a single frame. Not even with the first vid (highlighted target face), so I stopped manually after the first three vids.

Highlighting doesn't do anything except selecting the index for removal.

AttributeError: 'NoneType' object has no attribute 'embedding'>

That means no there was no target face whatsoever.

If I get you right, you would like to batch-process a couple of different videos with differing target faces and also different source faces?!?
So for example a List like:

Video1 - Source Faces A/B/C -> Dest Faces D/E/F
Video2 - " G/H/I -> J/K/L

Really?!? No, I mean...really? I doubt anyone will do something like this with a GUI, that's probably best done by batch and CLI (which I actually want to get rid off).

[lysxelapsed]

Got it, thanks for clearing that up. And yes, thats exactly why I'm specific in my test-review, so you could hopefully clear those things up. Absolutely, some instructions would be useful. I mean, what else than (mis)interpreting things is there, without sufficient programming skills and digging through the code?
Regarding preview: Yes, forgot to turn server off again after re-install. My fault.
And uhm, yes. Batch processing different source and target faces would be like the holy grail. But single source and different target faces would be a start. Isn't that, what you did with 2.0.3, minus the face selection? No idea why nobody would implement that in a GUI. But like I said: I probably lack the more-than-super-basic programming knowledge to understand. And it's your project man, you add or don't add whatever you like. Didn't mean to upset you, just giving my 2 cents. If they're not helpful, just say so.

[sordidloam]

Holy grail indeed!

[C0untFloyd]

No hard feelings, I was just disappointed that not many people seem to like the new web UI and honestly I can't figure why one would batch-process so many videos and faces as this is very time and energy consuming. But if that's your thing it's fine with me 😄
I need to focus more on the swapping part of roop, that's certainly something where Rope really shines. My project is the better all-around package though I think. Anyway, keep up the good support!

[sordidloam]

I can't imagine why people would want to batch videos... ;-)

I personally like the new web ui.

But then again, if you couldn't imagine, it wouldn't be a feature, instead its in many iterations of generative ai tools, for good reason.

You're doing the good work my friend. Keep it up.

[eniora]

I am pretty sure the majority of us like the new browser UI way more than the limited tkinter. Keep up the good work seriously! Can't wait to see what features and enhancements you can "unleash" for us with future updates. That's why I try to search for any issues and always list them here as soon as I find them every time a new update is released so you keep improving and fixing them in case you missed any. It's all about feedback and improvement and I am sure you don't find me annoying by always listing them issues right? 😁

If only Gradio allowed smooth live cam (like the fork by GosuDRM for example which still uses tkinter UI), I really hope you can figure out a hack for Gradio to achieve that some day. 😄

[dsfargeegrafsd]

Don't take it like that, the new UI is great! Most people probably haven't found it yet and those who did like it won't go out of their way to tell you they liked it. People are reporting issues because they like the project and see potential improvements.

[ghost]

I need some help from anyone with a solution. Sadly, my output folder is still clearing after the video is processed. The folder sends a temp video file into the output folder and then as soon as it is processed it just vanishes out of thin air with no trace. I used to get something like "video success" but now it just gets stuck at [ROOP.CORE] Finished

Yes, I have the latest version. I actually reset my PC and reinstalled all the programs and files with a fresh copy of windows because I was convinced that I had somehow downloaded something from the original Roop that was blocking NSFW content. But the problem is still there. Here is a copy of the last few lines before it pulls the video and vanishes from the output folder.

['ffmpeg', '-hide_banner', '-hwaccel', 'auto', '-y', '-loglevel', 'debug', '-i', 'C:\Users\leade\OneDrive\Documents\roop-unleashed\output\Custom Video 7_fake.mp4', '-i', 'C:\Users\leade\OneDrive\Documents\roop-unleashed\temp\86153d97bc53f0c971406e97c0c14ab45a844508\Custom Video 7.mp4', '-c:v', 'copy', '-map', '0:v:0', '-map', '1:a:0', '-y', 'C:\Users\leade\OneDrive\Documents\roop-unleashed\output\Custom Video 7_fake_final.mp4']
[ROOP.CORE] Finished

[sordidloam]

Edit the ui.py file located in the roop folder.

Find the line that starts with 'if roop.globa.CFG.clear_output:' Comment that out with a # at the beginning.

Make it look like this:

[ghost]

I made it look exactly like that. Still cannot get Gradio to keep the video after processing. Still just stops at ROOP.CORE Finished
:(

[sordidloam]

Hmm. Did you save the file and relaunch roop-unleashed?

[ghost]

It finally worked after I restarted my computer. Thanks for your help on the #TEMP FIX

[raptureoh9]

The Gradio Web-UI seems to render much slower than when using the old GUI version. On my RTX 3070 I get about 40 FPS when launching using the old GUI, and I get only 2 FPS when using the Gradio Web-UI. Confirmed that public server is off, and I confirmed it's using CUDA cores as well in task manager. Not sure why it's so much slower, I tried closing/restarting multiple times but no change.

[dsfargeegrafsd]

Are you using CLIP when that happens?

[ghost]

The Gradio Web-UI seems to render much slower than when using the old GUI version. On my RTX 3070 I get about 40 FPS when launching using the old GUI, and I get only 2 FPS when using the Gradio Web-UI. Confirmed that public server is off, and I confirmed it's using CUDA cores as well in task manager. Not sure why it's so much slower, I tried closing/restarting multiple times but no change.

I have the same issue. I tried Rope yesterday and my GPU kicks in on my RTX 2060 Super and renders fast. But, using Gradio, I have 2 FPS.

I am not sure what CLIP is, can someone explain?

Also, C0ountFloyd, Roop Unleashed Gradio is my favorite app to use of them all. I think that it has the best quality. Personally, I did a lot of testing with the enhancers and I found that Codeformer with the "Original/Enhanced image blend ratio" set to 0.85 produces the most realistic deep fake. I like Codeformer because it is good at matching neck and body skin tones and the face is not overly sharpened to the point that it looks bad. GFPGAN tends to over sharpen and also adds a small amount of pixilation around the edges of the face.

Other than the slow FPS processing, Roop Unleashed Gradio is hands down the best choice.

[C0untFloyd]

Thanks for the kind words!
Rope did a lot of groundwork optimizations, one of the biggest I've seen is the conversion of the inswapper model everybody is using to a half-precision onnx model. This trades precision for a huge speed increase. He also writes video frames in a pipe instead to disc (what I'm doing too now). He also only uses the face bbox for enhancements (me too except for CodeFormer) and I bet he re-uses the landmarks of the first face recognition instead of detecting again. The latter one I'm doing too, but currently only for DMDNet Enhancer. So in short, performance-wise Rope is way ahead of the rest, you should benchmark roop-unleashed with the original roop or the 2.0.3 version to see if there are issues.

[eniora]

Just updated to 2.6.8 and as I can see most of the major issues are fixed now, very good job! I will report back if I find any issues that aren't documented in the OP.

Swapping multiple faces now possible

This is really nice, just tested it and works perfectly.

Here are two suggestions/requests:

1. Add a permanent button near the results window that opens the output folder. (similar to SD A1111 web UI button)

2. Ability to change the output file name before clicking on start, maybe something like a checkbox with "custom output name" or a drop down list with "default name" (which is the current behavior) and another option on the list with "custom output name" that let's you change the output file name before processing.

I know you're prioritizing the fixes now but I think it would be cool if these features are added some time later in the future.

[C0untFloyd]

Adding this to known issues 😞

[C0untFloyd]

Add a permanent button near the results window that opens the output folder. (similar to SD A1111 web UI button)

In a Browser and wanting to be platform-independent this is just not that easy as it seems. I tried to find it in A1111 without success. I'm guessing they do it using Javascript but I wonder if it works with every browser then.

[eniora]

Thanks for the effort, not really important but I didn't know it's hard to add the button since obviously I barely know anything about coding 😬

[eniora]

@C0untFloyd just did some inspect element on the button from A1111 and found out it uses Svelte (optimized javascript from what I read).

[C0untFloyd]

Yes, I know....A1111 is using the same Gradio Framework like me, which uses Svelte as Backend.😄
I'm using my own function in python now, works as expected in windows, just can't test this on other platforms.

[eniora]

@C0untFloyd I discovered an issue and it took me a lot of time trying to figure out why it happens. So I have been trying to swap a few videos I have and every time 2.6.8 finishes processing some of them, the video gets immediately deleted. Tried all of them on unleashed 2.0.3 and all fine. So what I discovered (after a lot of troubleshooting time) is that with the Gradio roop if a video doesn't have an audio track and you don't check the "Skip audio" option the video gets immediately deleted from the output folder after processing is finished.

Here's a short video you can test on: sample video without audio

Now I can see why some people reported this issue in the comments here and no one knew the cause.

[C0untFloyd]

Oooh thanks a lot, that explains why people thought I'm clearing the output folder!

[C0untFloyd]

Fixed and committed, as always it was just a wrong one-liner...thank you!

[eniora]

Cheers! just tested, all good! This Gradio unleashed is getting closer and closer to perfection with every new update.
Oh and I will keep looking for new issues 😈 (J/K, I just use it a lot that's why😄)

[popgoesthewizzle]

hello how to launch this one vs the 2.0.3 version from previous installment. i believe i installed everything correctly. but this is first time for me to use gradio.. the other version was simple for me to understand enough but this one is beyond my head for some reason since first time using gradio... please assist. im sorry to bug you but i would like to learn these things.

[anxiousottergames]

I am using a RTX 3080, the frame/s can reach 30-40. Some video can have an average of 20+. But sometimes, it drops down to 5-10frame/s, Does anyone have any idea what went wrong?

[eniora]

Probably your VRAM (which is 10GB for the 3080) is getting full, try with --execution-threads 6 commandline and set it to 6 in the settings page as well. It also depends on the video, face(s) size in the frame, or if it's swapping multiple faces in a set of frames etc so it's normal and not uncommon if it goes up and down during processing.

[anxiousottergames]

Jesus....I found out the issue....it was because the original input was a PNG not JPG...
It went from 5frame/s to 35frame/s

[anxiousottergames]

Your method also pushed the frame/s higher as well, thank you!

[C0untFloyd]

Closing this now as the gradio version evolved to be the main one. Thanks for all your support & input!

[TrickleDownClown]

Super excited. It's really great work. Forgive my commenting on the older thread but I am unable to install using the web.bat it gives an error about conda but it is for sure installed.

[C0untFloyd]

Can you paste the errors or a screenshot? You actually don't need to have Conda installed, it will also be installed automatically. I tested the installer twice to be sure it's working, this is weird.

[TrickleDownClown]

I actually figured it out. I had a space in the folder name. All working and looks really great with some quick tests.

[eniora]

Thanks a lot for adding the "open output folder" button, makes life easier 😄 also noticed that the enhancer is faster now in the new 2.7.0 version.

[etruscanmuffin909]

Worked perfectly from one-click installer. Seems like this release is the best to date! Thank you for all your work on this.