Home Creators Posts Import Register Favorites Logout

Downloads

Content

JoyCaption in four improved Gradio apps: Pre-Alpha, Alpha One, Alpha Two and the newest, fastest Joy Caption Beta One, all with batch and multi-GPU captioning.

1-click installers for Windows, RunPod and Massed Compute with Torch 2.8, Triton 3.4 and CUDA 12.9. All Joy Captions apps support batch processing of sub-folders, 4-bit quantization and caption prefix and suffix; Beta One adds saved presets and liger_kernel speed.

Patreon exclusive posts index to find our scripts easily, Patreon exclusive posts index to see which scripts updated or added last and amazing Patreon special generative scripts list that you can use in any of your task.

Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord

Please also Star, Watch and Fork our Stable Diffusion & Generative AI  GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)

=======

App Screenshots Gallery here (Public) : https://www.reddit.com/r/SECourses/comments/1mfedbw/all_joy_captions_joycaption_apps_1click/

For mass editing captions > https://www.patreon.com/posts/108992085

Latest installers : Joy_Caption_v41.zip

I have developed 4-seperate amazingly advanced Gradio APPs for

  • Joy Caption Pre Alpha

  • Joy Caption Alpha One

  • Joy Caption Alpha Two

  • Joy Caption Beta One - Newest and Best

    • Beta One is extremely fast, advanced, and supports full batch processing

  • When doing multi-GPU batch captioning displayed statistics could be wrong

29 November 2025 Update

  • With v41 timeout increased from 10 seconds to 60 seconds

21 September 2025 Update

  • With v40, we have added full preset save and load feature to Joy Caption Beta One

  • You can now even add new Extra Options (for non-custom prompts)

  • It will remember all your saved settings in preset and auto load next time you restart your app

  • Presets will be saved inside presets folder

  • Extract zip file and overwrite previous files for update that is sufficient to upgrade from V18

18 September 2025 Update

  • With v39, we have significantly improved Joy Caption Beta One

  • Now it fully and properly utilizes liger_kernel

  • The processing speed of BF16 is extremely fast

  • You will lose a lot of speed in Int8 or NF4 - so use only if you get out of VRAM error

  • Multiple GPU processing of Joy Caption Beta One completey rewritten and now works amazing and ultra fast

  • Installation libraries upgraded to Torch 2.8, Triton 3.4, CUDA 12.9 for Windows, RunPod and Massed Compute

    • I am using all pre-compiled libraries so it should work fairly well as long as you have Python 3.10.11 and Git installed

  • Extract zip file and overwrite previous files. Moreover, i recommend delete venv folder and make a fresh install of venv

Windows Requirements

  • Python 3.10, FFmpeg, CUDA 12.9, C++ tools, MSVC and Git

  • If it doesn't work make sure to below tutorial and install everything exactly as shown in this below tutorial

  • https://youtu.be/DrhUHnYfwC0

22 August 2025 Update

  • With V38 now there is extra option for Beta One app

    • If you enable Unload model and clear VRAM after each caption before first run, it will fully unload model after each caption and clear VRAM

17 August 2025 Update

  • With V36 now it will process with sorted by name all folders and sub-folders in all apps

    • So it will sort files and folders by name first and then will do batch process

      • E.g. A > B > C etc

  • To update to V36 just extract files from zip and overwrite older files that is it

17 May 2025 Update

  • Joy Caption Beta One app added to our installers

  • Now fully supports RTX 5000 series GPUs with all kind of speed improvements

    • Installs with xFormers, Flash Attention, Sage Attention, Triton, DeepSpeed

    • Thus, Joy Caption Beta One works ultra fast

    • If you get Triton error use Clear_Triton_Cache.bat

  • Delete older venv and make a fresh install

5 January 2025 Update

  • Joy Caption Alpha Two stop batch processing problem fixed

  • Joy Caption Alpha Two batch statistics display on CMD fixed

  • Joy Caption Alpha Two batch "Error processing image d.JPG: 'd.JPG' and 'd.JPG' are the same file" error fixed

  • Joy Caption Alpha Two even multi-GPU batch processing displays accurate statistics and progress on CMD right now

29 December 2024 Update

  • Batch processing now will process sub-folders as well

  • When using Pre Alpha or Alpha One, batch processing was slower on single GPU and was loading model again. This issue fixed

Features

All apps has the following amazing features. If any of them are broken please report and let me know

  • Use 4-bit Quantization

  • Input Folder for Batch Processing

  • Output Folder for Batch Processing (Optional)

  • Multi-GPU Batch captioning (e.g. 0,1,2)

  • Overwrite existing caption file - by default skips

  • Append new caption to existing caption

  • Remove newlines from generated captions

  • Cut off at last complete sentence

  • Discard repeating sentences

  • Don't save processed image

  • Resize Input Image Max Resolution

  • Caption Prefix and Caption Suffix

How To Install And Use

Using 4-bit Quantization reduces VRAM usage but also slows down.

When using multi-GPU batch captioning, the displayed statistics on CMD could be wrong but still should work perfectly fine.

Massed Compute (Recommend Cloud) :

RunPod (Cloud):

Patreon exclusive posts index to find our scripts easily, Patreon exclusive posts index to see which scripts updated or added last and amazing Patreon special generative scripts list that you can use in any of your task.

Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord

Please also Star, Watch and Fork our Stable Diffusion & Generative AI  GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)

=======

App Screenshots Gallery here (Public) : https://www.reddit.com/r/SECourses/comments/1mfedbw/all_joy_captions_joycaption_apps_1click/

For mass editing captions > https://www.patreon.com/posts/108992085

Latest installers : Joy_Caption_v40.zip

I have developed 4-seperate amazingly advanced Gradio APPs for

  • Joy Caption Pre Alpha

  • Joy Caption Alpha One

  • Joy Caption Alpha Two

  • Joy Caption Beta One - Newest and Best

    • Beta One is extremely fast, advanced, and supports full batch processing

  • When doing multi-GPU batch captioning displayed statistics could be wrong

21 September 2025 Update

  • With v40, we have added full preset save and load feature to Joy Caption Beta One

  • You can now even add new Extra Options (for non-custom prompts)

  • It will remember all your saved settings in preset and auto load next time you restart your app

  • Presets will be saved inside presets folder

  • Extract zip file and overwrite previous files for update that is sufficient to upgrade from V18

18 September 2025 Update

  • With v39, we have significantly improved Joy Caption Beta One

  • Now it fully and properly utilizes liger_kernel

  • The processing speed of BF16 is extremely fast

  • You will lose a lot of speed in Int8 or NF4 - so use only if you get out of VRAM error

  • Multiple GPU processing of Joy Caption Beta One completey rewritten and now works amazing and ultra fast

  • Installation libraries upgraded to Torch 2.8, Triton 3.4, CUDA 12.9 for Windows, RunPod and Massed Compute

    • I am using all pre-compiled libraries so it should work fairly well as long as you have Python 3.10.11 and Git installed

  • Extract zip file and overwrite previous files. Moreover, i recommend delete venv folder and make a fresh install of venv

Windows Requirements

  • Python 3.10, FFmpeg, CUDA 12.9, C++ tools, MSVC and Git

  • If it doesn't work make sure to below tutorial and install everything exactly as shown in this below tutorial

  • https://youtu.be/DrhUHnYfwC0

22 August 2025 Update

  • With V38 now there is extra option for Beta One app

    • If you enable Unload model and clear VRAM after each caption before first run, it will fully unload model after each caption and clear VRAM

17 August 2025 Update

  • With V36 now it will process with sorted by name all folders and sub-folders in all apps

    • So it will sort files and folders by name first and then will do batch process

      • E.g. A > B > C etc

  • To update to V36 just extract files from zip and overwrite older files that is it

17 May 2025 Update

  • Joy Caption Beta One app added to our installers

  • Now fully supports RTX 5000 series GPUs with all kind of speed improvements

    • Installs with xFormers, Flash Attention, Sage Attention, Triton, DeepSpeed

    • Thus, Joy Caption Beta One works ultra fast

    • If you get Triton error use Clear_Triton_Cache.bat

  • Delete older venv and make a fresh install

5 January 2025 Update

  • Joy Caption Alpha Two stop batch processing problem fixed

  • Joy Caption Alpha Two batch statistics display on CMD fixed

  • Joy Caption Alpha Two batch "Error processing image d.JPG: 'd.JPG' and 'd.JPG' are the same file" error fixed

  • Joy Caption Alpha Two even multi-GPU batch processing displays accurate statistics and progress on CMD right now

29 December 2024 Update

  • Batch processing now will process sub-folders as well

  • When using Pre Alpha or Alpha One, batch processing was slower on single GPU and was loading model again. This issue fixed

Features

All apps has the following amazing features. If any of them are broken please report and let me know

  • Use 4-bit Quantization

  • Input Folder for Batch Processing

  • Output Folder for Batch Processing (Optional)

  • Multi-GPU Batch captioning (e.g. 0,1,2)

  • Overwrite existing caption file - by default skips

  • Append new caption to existing caption

  • Remove newlines from generated captions

  • Cut off at last complete sentence

  • Discard repeating sentences

  • Don't save processed image

  • Resize Input Image Max Resolution

  • Caption Prefix and Caption Suffix

How To Install And Use

Using 4-bit Quantization reduces VRAM usage but also slows down.

When using multi-GPU batch captioning, the displayed statistics on CMD could be wrong but still should work perfectly fine.

Massed Compute (Recommend Cloud) :

RunPod (Cloud):

Related posts

All posts in this topic: Captioning, Dataset and Image Processing Tools · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

Comments

Ant-2014

Hi. Its unbelievable. You are so efficient, so we just recently talked about this project. I really appreciate your work, a lot of time and effort spent for this /and for us, your subscribers. Thank you and All the best to you, Master!

Markus

Amazing, thanks a lot! Which JoyCaption Model works best?

SECourses

i think Joy Caption Pre Alpha but you can compare all 3

cool1

Thanks a lot. In the Alpha 2 version it has 2 new checkboxes, one "Use FP16 Model Type" and one "Use 4-bit Quantization". What does it use for those if you don't select them? If they're not selected does it use a higher bit model type (like FP32?) and no quantization? I assume we should leave both those blank for the most accurate captions if we have enough VRAM and we are okay with it using that higher VRAM? In one of the top new options for Alpha 2 there's an option to "not include information about people/characters that cannot be changed (like ethnicity, gender etc)...". Have there been any tests anywhere to see if doing that actually helps? I don't know if by adding those descriptions in the captions if it might make you be more likely to need to specify them when prompting (after training), but I thought including them might also help it train them quicker (by starting learning with learned concepts that are similar to that). When you use the "stop batch processing" option, if you kept doing that a lot does that ever cause problems? Like is it closing all processes and freeing things fully? eg. a few times after using the previous Alpha 1 version I've had one of the windows processes start using about 2.5% CPU after stopping batch processing a few times (which made things feel slower) but I don't know if it was caused by that or other things (since I'd been using multiple apps), but it went to 0 after rebooting (task manager said it might cause instability if I end-tasked that windows process which I was trying to restart to get it to 0% CPU).

SECourses

you should test all. also fp16 is for older gpus dont use unless you needed. 4 bit is good for lower vram gpus especially if you dont have 24 gb gpu

DarkView AI

Extra options causes error, spits out random caption and takes over a minute. i selected the first two options on extra options, also selected Do Not mention the image's resolution, and Do not use any ambigous language.

SECourses

ye this model is not that great. but since people requested i added :D

H M

Perfect timing on this release. Feature request for batch processing; the capability of working through subfolders within the target folder. Edit: A way to offload the model without closing and reopening it would also be appreciated.

SECourses

just made v28 update 1 : fixed 2 : please read changes and see if fixes or elaborate more

Sphiwe April

Is it possible to add the prompt description box - where I can specify what the model should keep in "mind" as it captions. Prefix and suffix are the best but very limiting

SECourses

the apps has custom prompt boxes you mean that?

cool1

When I mentioned the windows process sometimes going to about 2.5% CPU, I think now it was probably nothing to do with JoyCaption or "stop batch processing". It's "Shell Infrastructure Host" that goes to that much and perplexity.ai says that could be caused by a memory leak in the windows photo viewer and I'd used that (legacy version). So it was probably that. I'll use a different image viewer in future so it shouldn't have that problem.

cool1

In their page about it they scored JoyCaption Alpha 2 at 22.8 for avg score (top) but that's only slightly higher than the 22.6 that they scored Pre Alpha. Though alpha 2 has more options and they did add more training stuff to it. Though they said standard deviation for Pre Alpha and Alpha 2 are both 5.32 so they think that's equal (I would have thought one would be better if the alpha 2 is scored slightly better). Also they only got those scores from checking 15 images. Ideally the could have tested more. It still sometimes mixes up the left and right side (eg. talking about the person's right hand when it's actually their left hand, and re-running it can make it correct it. It might get that wrong 25% of the time though I haven't done a proper test). I noticed in Alpha 1 version, when captioning a single image that you drag to the input image section, it output the images with the same filename as the source image to the output folder (so you'd have multiple images being output there). But the new Joycaption Alpha 2 version is outputting them all as "temp_input_image.png". I assume that's intentional, though it's different than before. In JoyCaption Alpha 2 it shows "Gradio" in the browser tab. Ideally that could show "JoyCaption Alpha 2" there so it's clearer when switching between tabs/windows though it's not vital.

cool1

It seem like there might be a bit of a bug/issue with batch captioning with JoyCaption Alpha 2. I batch captioned a folder that already had .txt files with short captions for each image - so captioning to make them longer. I had "Overwrite existing caption file" checked at the time I selected "batch caption". It seems to have captioned every image in the folder, updating the .txt files with the longer captions, but after each one it shows in the command window "Error processing image [folder & filename]: [folder & filename] is the same as [folder & filename].". Then at the end of the batch it shows in the command prompt "Processed 0/45 | skipped: 0| Failed: 45" (the GUI "batch processing progress" window also mentions the same number failed) but that's incorrect as it seems to have added longer captions to all the existing .txt files in the folder and not failed on them.

cool1

Also another issue with JoyCaption Alpha 2 is pressing the "stop batch processing" button doesn't stop it from processing the images in the folder. It says the following error: ...\Joy_Caption_Alpha_Two.py", line 649, in stop_batch_processing_fn stop_flag.value = True NameError: name 'stop_flag' is not defined then continues captioning the rest of the images in the folder.

cool1

Thanks a lot

cool1

I extracted the v29 files. It shows v29 in the gradio top of screen for Alpha 2, but when I press "stop batch processing" now while it's processing a batch, nothing seems to happen. It doesn't give the error message now though. The code in alpha 2 for checking for stopping seems different to alpha 1's checking in process_gpu_batch. I don't know if copying the way it was doing it in Alpha 1 might make it stop the batch processing in the same way as alpha1 as that one worked ok.

SECourses

i tried regular batch it worked. start processing let it process like 1 image and then stop you will see.

cool1

I've tried that. I've let it go past 1 too and then kept pressing it, nothing happens, it just seems to keep going on the next .txt files until it gets the end, no matter how many times I press "stop batch processing". Alpha 1 worked ok for this. Gpu IDs box is set to 0 (the default). I don't know if that affects it. I only have 1 GPU (3090).

SECourses

you were right. for some reason it had overwritten older file. please try Joy_Caption_v29_v2.zip

cool1

Thank you. I downloaded 29_2 and extracted that. It shows as V29 in the gradio top of the screen but it works differently so it must have extracted okay. When I ran it it and then pressed "stop batch processing" it immediately said in the command window "Stop Clicked Stopping batch processing. Please wait for current tasks to finish...". So I waited and then it stopped soon after (probably not processing more than the first image). Though it seemed to take longer to actually stop than Alpha1. It still gives the error messages about existing .txt files being there but I can ignore those. But after it has fully stopped (it's said "Batch processing complete. Batch processing complete. Processed: 0..." and I try to to start the batch processing again by pressing "Start batch processing" it immediately says "Starting batch processing. Found 1 GPU. what the input & output folders are, and immediately "Batch processing complete. Batch processing complete. Processed: 0...". So each time I press "start batch processing" (some time after it's fully stopped after using "stop batch processing") it won't restart again. "Overwrite existing caption file" option is still checked so it shouldn't be failing because of existing .txt files.

SECourses

try fresh folder no txt files and see if works

cool1

I've just tried deleting all the .txt files from that folder and re-pressing "start batch processing" and it still does the same thing. It still immediately shows "Starting batch processing. Found 1 GPU(s).. Batch processing complete. Batch processing complete. Processed: 0..." (if I use that some time after stopping batch processing. I assume if I ctrl C out of the program and re-start it it would work again, but only until I stopped batch processing once - which seems to take longer than alpha 1 did). I can go back to Alpha 1 or Pre Alpha for now though. I assume if the same stop/start method was used that was used in alpha 1 was used it might work the same way as that, which was working. But if you changed if for other things like multi-GPU maybe the default way could work the same way as alpha 1 (most people probably only have 1 GPU).

SECourses

download v31 fixed all errors. also read changelogs

eman eliforp

How to eliminate duplicated/similar images? Which tool is best?

SECourses

coding it right now for you stay tuned

cool1

Thanks a lot. I'll try that. edit: It seems to work okay now thanks. The stopping and starting batch processing seems okay and it's not showing those errors on each file now. It takes quite a long time for the first image in the batch after re-starting the batch (about 2 mins) but I assume it's loading the model then. I don't remember the alpha 1 taken that long. Maybe the model is a bigger than alpha 1. But it's working. It's a lot faster at re-starting batch processing when it wasn't stopped with the "stop batch processing" option (doesn't take mins on the first image). edit: the first image in the batch has sped up a lot now after re-starting batch processing after stopping it part way through (it doesn't take 2 mins now just seconds). Maybe there was something else running like Adobe's creative cloud UI helper that was making it slower before.

cool1

For now, one way is with Digikam (a free image organizer/viewer). You can right click on a folder (that's in an album) and select "Find duplicates" and you can specify a similarity range %. Though you first need to click "update fingerprints" on the images as far as I know. (if it hadn't already been done for those images). Though it would be easier if there was just a simple way for a basic option in a program to just select a folder and it told you the duplicates (ie. a more basic/lightweight option than using digikam for it).

Anshul Gupta

You need to include or put the update bat file inside the one-click installer script with all of your one-click installer scripts because it is a hard and time-consuming process. Second, in this way, ssd stroage tbw performance get expired fastly because ai tools are large file tools and it makes ssd obsolete

SECourses

i agree with you. i will try to make better hopefully

Rikard

i'm a bit dense when it comes to these things, but you said this would work for 8GB GPU systems? i have an 8 GB gaming laptop, but whenever i've tried to run this even with fp16 and 4-bit Quantization i get the error "The paging file is too small for this operation to complete." i assume i have to do something with the RAM memory, but i wanted to hear what suggestions you have

SECourses

yes please set your virtual RAM to 100 GB. look youtube so easy to set. your RAM is not being sufficient

cool1

Thanks for the update. Though when I select to start (eg. 3 or 4) it says: "ModuleNotFoundError: No module named 'gradio'" Also when running the download models.bat it didn't seem to get to 100% on everything but then said they're all downloaded and rerunning "download models" again again says they're all downloaded. So I don't know if I have partially downloaded model files now. Ideally it could be able to detect that and continue where it left off if any are incomplete.

Harkonnen

Apparently venv is not installed correctly, modules are missing

SECourses

sorry for this error fixed with v33

SECourses

sorry for this error fixed with v33

SECourses

sorry for this error fixed with v33

SECourses

sorry for this error fixed with v33

SECourses

you are welcome. also there was an error fixed with v33

EvanGuo

No matching distribution found for triton>=2.3.1

cool1

Thanks. I'll try the new version. How can I tell which of the model files are incomplete after using the "download models.bat" file? It didn't seem to get to 100% on all model files when it said it finished the download. edit: The joycaption beta 1 is working for me now (with v33) thanks so at least the model for that must have downloaded okay.

cool1

In the batch caption option of Joycaption beta 1, it seems to create .txt files for each but for quite a few times when processing the images in batch mode it says "[date & time] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib'" Do those matter? Is that something to do with trying to do CUDA acceleration and it failing but then still continuing to caption but not in an accelerated way? Also quite a few times when processing the images in batch mode it says " NOTE: Redirects are currently not supported in Windows or MacOs.". I don't think it needs to output that multiple times if it's not really needed. It doesn't seem to be outputting to the console the number of images it's processed in batch mode so far and which one it's on and how many are left to do like Joycaption alpha 2 did. And it might be good if it also output a message to the console when it's finished processing the batch if possible (I think the alpha 2 did that). Otherwise there's nothing shown to indicate it's finished. It could list at the end the number processed, number skipped etc. how long it took, like before if possible. Also when you press "start batch processing" that button doesn't change or become disabled until it finishes and no "cancel/stop batch processing" button shows. It would help if a cancel/stop batch processing button showed in case needed like you could with Joycaption Alpha 2 (eg. in case you've not set the right things and need to stop it quickly). edit: though it does show a processing % at the top of the GUI in batch mode. But showing some processing info in the console could help too.

Cemil Hacimahmutoglu

Enter your choice (1-5): 4 [2025-05-18 20:53:26,278] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: 'aio.lib' giriş dosyası açılamıyor test.c LINK : fatal error LNK1181: 'cufile.lib' giriş dosyası açılamıyor W0518 20:53:34.810000 15380 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. 🛠️ System configuration: Python : 3.10.11 (H:\Joy_Caption_v33\venv\Scripts\python.exe) PyTorch : 2.7.0+cu128 ‣ CUDA build : 12.8 transformers : 4.51.3 bitsandbytes : 0.45.5 liger_kernel : 0.5.8 GPUs (total 1): • [0] NVIDIA GeForce RTX 3080 Ti | compute 8.6 | 12.0 GiB * Running on local URL: http://127.0.0.1:7860 * To create a public link, set `share=True` in `launch()`.

RayHell

With all your release lately I have this: LINK : fatal error LNK1181: cannot open input file 'aio.lib' issue and I'm not alone. Are you planning to fix that or at least point at how to fix it.

SECourses

it is related to linux. 100% safe to ignore on windows. pytorch related not me :D

SECourses

it is related to linux. 100% safe to ignore on windows. pytorch related not me :D

SECourses

it is related to linux. 100% safe to ignore on windows. pytorch related not me :D

SECourses

fixed the error. please use latest zip file. sorry for this.

SECourses

it should just download missing files when you click download model - it will skip existing ones

cool1

Thanks. What about the "stop batch processing" option, could we have an option for that please? Though if it's faster than before it might not be needed as much but you still might make a mistake with settings or not need to stop soon before it captions all files. And though it shows a % shown in the top of the UI (that I didn't notice at first) it still might help to have a few stats shown in the cmd prompt like when it's finished (eg. x captioned, y skipped).

cool1

Thanks. I was just worried I might have ended up with some less than complete (like 80% or 90% complete) model files after the "download models.bat" file running, so they existed but weren't fully complete. And wondered if it might have stopped the program running or caused crashes/hangs if it tried to load an incomplete one. But the beta 1 loaded okay.

Sergio Valsecchi

Hi Furkan, on Windows, when I run Joy Caption Beta One I get this error — it's quite strange because in ComfyUI, Joy Caption Beta One works just fine. Do you have any solution for this? Joy_Caption_31 runs without any issues. Enter your choice (1-5): 4 [2025-06-03 13:52:50,948] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: non è possibile aprire il file di input 'aio.lib' Traceback (most recent call last): File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\utils\import_utils.py", line 1967, in _get_module return importlib.import_module("." + module_name, self.__name__) File "C:\Program Files\Python310\lib\importlib\__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\models\llava\modeling_llava.py", line 27, in from ...modeling_utils import PreTrainedModel File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\modeling_utils.py", line 158, in import deepspeed File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\deepspeed\__init__.py", line 25, in from . import ops File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\deepspeed\ops\__init__.py", line 15, in from ..git_version_info import compatible_ops as __compatible_ops__ File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\deepspeed\git_version_info.py", line 29, in op_compatible = builder.is_compatible() File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\deepspeed\ops\op_builder\fp_quantizer.py", line 35, in is_compatible sys_cuda_major, _ = installed_cuda_version() File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\deepspeed\ops\op_builder\builder.py", line 53, in installed_cuda_version output = subprocess.check_output([cuda_home + "/bin/nvcc", "-V"], universal_newlines=True) File "C:\Program Files\Python310\lib\subprocess.py", line 421, in check_output return run(*popenargs, stdout=PIPE, timeout=timeout, check=True, File "C:\Program Files\Python310\lib\subprocess.py", line 503, in run with Popen(*popenargs, **kwargs) as process: File "C:\Program Files\Python310\lib\subprocess.py", line 971, in __init__ self._execute_child(args, executable, preexec_fn, close_fds, File "C:\Program Files\Python310\lib\subprocess.py", line 1456, in _execute_child hp, ht, pid, tid = _winapi.CreateProcess(executable, args, FileNotFoundError: [WinError 2] Impossibile trovare il file specificato The above exception was the direct cause of the following exception: Traceback (most recent call last): File "E:\APP_AI\Joy_Caption_33\Joy_Caption_Beta_One.py", line 2, in from transformers import LlavaForConditionalGeneration, TextIteratorStreamer, AutoProcessor File "", line 1075, in _handle_fromlist File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\utils\import_utils.py", line 1956, in __getattr__ value = getattr(module, name) File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\utils\import_utils.py", line 1955, in __getattr__ module = self._get_module(self._class_to_module[name]) File "E:\APP_AI\Joy_Caption_33\venv\lib\site-packages\transformers\utils\import_utils.py", line 1969, in _get_module raise RuntimeError( RuntimeError: Failed to import transformers.models.llava.modeling_llava because of the following error (look up to see its traceback):

Zow

Thanks for the caption app! I use it to caption SEO ready image ALT texts, but I have to download the picture first and then caption it. So it would be nice if it could caption single images directly from a URL input, so I can then copy paste those AI alt texts into ALT="x" code (just an idea).

SECourses

hmm interesting i dont know if possible. thanks for idea

SECourses

hi please delete venv and run installer and send me install logs so i can see install error : monstermmorpg@gmail.com

CulturedThighLover

Yeah I think i'm getting the same error. Fixed: Had to uninstall flash attention and comment out some lines. In joy_caption_v33, commented out lines 7,269/270. Changed line 267 to g_model = LlavaForConditionalGeneration.from_pretrained(MODEL_PATH, torch_dtype=torch.bfloat16, device_map=device_map_target, attn_implementation="eager"). Did the same for Line 293. (windows user here)

DarkView AI

SECourses JoyCaption Image Captioning App - Select to Start 1. Joy Caption Pre Alpha 2. Joy Caption Alpha One 3. Joy Caption Alpha Two 4. Joy Caption Beta One 5. Exit Enter your choice (1-5): 4 [2025-07-14 21:44:13,851] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:44:18.533000 26928 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. 🛠️ System configuration: Python : 3.10.11 (E:\Joy_Caption\venv\Scripts\python.exe) PyTorch : 2.7.0+cu128 ‣ CUDA build : 12.8 transformers : 4.51.3 bitsandbytes : 0.46.1 liger_kernel : 0.5.8 GPUs (total 1): • [0] NVIDIA RTX PRO 6000 Blackwell Workstation Edition | compute 12.0 | 95.3 GiB * Running on local URL: http://127.0.0.1:7860 * To create a public link, set `share=True` in `launch()`. Loading processor... Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`. Loading model with quant: bf16, device_id: 0... Loading checkpoint shards: 100%|██████████████████████████████████████████████████████| 4/4 [00:17<00:00, 4.46s/steps] [2025-07-14 21:48:33,345] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:48:36.776000 29128 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:48:42,232] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:48:45.658000 15152 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:48:51,159] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:48:54.454000 31112 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:48:59,701] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:02.996000 32852 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:08,292] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:11.600000 30828 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:16,958] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:20.217000 38976 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:25,594] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:28.869000 9120 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:34,212] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:37.482000 21636 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:42,703] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:45.925000 39536 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. [2025-07-14 21:49:51,224] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: cannot open input file 'aio.lib' test.c LINK : fatal error LNK1181: cannot open input file 'cufile.lib' W0714 21:49:54.510000 26900 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. stuck in a loop

SECourses

none of these are error i also get them . just try to use it should work

eyal giron

~/AI/Joy_Caption_v34 ⌚ 11:01:46 $ python3 Joy_Caption_Beta_One.py --share [2025-07-21 11:01:54,183] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect) df: /home/eyal/.triton/autotune: No such file or directory Traceback (most recent call last): File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1967, in _get_module return importlib.import_module("." + module_name, self.__name__) File "/usr/lib/python3.10/importlib/__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "", line 1050, in _gcd_import File "", line 1027, in _find_and_load File "", line 1006, in _find_and_load_unlocked File "", line 688, in _load_unlocked File "", line 883, in exec_module File "", line 241, in _call_with_frames_removed File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/models/llava/modeling_llava.py", line 27, in from ...modeling_utils import PreTrainedModel File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/modeling_utils.py", line 158, in import deepspeed File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/deepspeed/__init__.py", line 25, in from . import ops File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/deepspeed/ops/__init__.py", line 15, in from ..git_version_info import compatible_ops as __compatible_ops__ File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/deepspeed/git_version_info.py", line 29, in op_compatible = builder.is_compatible() File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/deepspeed/ops/op_builder/fp_quantizer.py", line 35, in is_compatible sys_cuda_major, _ = installed_cuda_version() File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/deepspeed/ops/op_builder/builder.py", line 51, in installed_cuda_version raise MissingCUDAException("CUDA_HOME does not exist, unable to compile CUDA op(s)") deepspeed.ops.op_builder.builder.MissingCUDAException: CUDA_HOME does not exist, unable to compile CUDA op(s) The above exception was the direct cause of the following exception: Traceback (most recent call last): File "/home/eyal/AI/Joy_Caption_v34/Joy_Caption_Beta_One.py", line 2, in from transformers import LlavaForConditionalGeneration, TextIteratorStreamer, AutoProcessor File "", line 1075, in _handle_fromlist File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1956, in __getattr__ value = getattr(module, name) File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1955, in __getattr__ module = self._get_module(self._class_to_module[name]) File "/home/eyal/AI/Joy_Caption_v34/venv/lib/python3.10/site-packages/transformers/utils/import_utils.py", line 1969, in _get_module raise RuntimeError( RuntimeError: Failed to import transformers.models.llava.modeling_llava because of the following error (look up to see its traceback): CUDA_HOME does not exist, unable to compile CUDA op(s) error when i try to launch it

Roni Bliss

Thanks much for this amazing resource. There appears to be an issue with v34 - changing caption type has no effect. they all come out as "detailed" and none of the options seem to be taking effect?

SECourses

ok but which version of the app :D Joy_Caption_Alpha_One or Joy_Caption_Alpha_Two etc

SECourses

looks like you are using custom linux. you have to have gcc cuda nvidia drivers python all properly installed. i can guess your cuda not installed

Roni Bliss

Beta One, the others are past there sell by date! haha

Roni Bliss

My ErrorBot Ai is saying this: Error Summary: AttributeError: 'NoneType' object has no attribute 'strip' Occurs in build_prompt function at if name_input.strip() and NAME_OPTION not in effective_extra_options: Root Cause: The variable name_input is None when .strip() is called. This means the input field is empty or not initialized, and the code does not check for None before calling .strip(). How to Fix: Add a check to ensure name_input is not None before calling .strip(). This prevents the error when the input is empty.

SECourses

ok i tested and as i change dropdown or select checkboxes it fills the "Confirm or Edit Prompt" and it uses that as long as you dont write anything into Custom Prompt (Overrides Caption Type, Length, Extra Options if filled)

Roni Bliss

Also came across this a few days ago, its gives just as good output as Beta one from my testing in ComfyUI, and it uses barely any GPU, plus + mixed tags and captions + Flux version + Batch Version. Def worth building a Gradio around it, literally anyone rven with very low VRAM can use it. You will be able to make something magic from it Gradio Im sure! https://github.com/miaoshouai/ComfyUI-Miaoshouai-Tagger Edit: Well maybe not quite as good as Beta One, but enough needed for robust tags and captions and its blazing fast...

SECourses

thanks i will check . it is based on Microsoft Florence-2

Roni Bliss

Ahh on my installation, "Confirm or Edit Prompt" says "error" ight from start up. I have tried reinstalling it several times and using different pods, I have 1TB Network Volume on RunPod, So I now have every one of your amazing releases I can fire up on any GPU which is really next level when flipping on to a B200 ... Anyway I digress! haha , I am going to try it out on massed compute, My big gripe with using Massed Compute is that you spend so long setting shit up, and unless you want to have a GPU running at $1.50 + an hour you lose everything, If they had persistent disks it would be the best platform. Anyway I am going to test it on Massed. But to be honest PromptGen 2 I am liking it much more now than Joy Caption, its a hidden Gem its very very similar output to Joy Caption, yes its Florence 2 but their own trained version, made in the same way as Joy Caption (With DanBooru tags etc) Its really great being able to combine a long detailed caption with danbooru tags, works amazing for Flux and especially BigAsp base SDXL models. I batch tagged and captioned, a dataset today of 400 images in under 8 mins, its blazing fast. It would be awesome in one your custom Gradio apps, which are all really amazing! There is nothing that comes close to the value you give your subscribers! - The amount of work you put in and still you take the time to answer 100,000 error related questions. Rarely ever see anyone thanking you for the work you put in.

SECourses

thanks a lot. also what error you getting? do you see error on cmd? i am making a fresh venv install to test

SECourses

thank you for reporting. fixed with v35. it was due to new version of Gradio

Roni Bliss

Yessssss!! Straight in with the fix!, you absolutely rock amigo! - The error in that section was on the UI the standard Gradio "error" I couldn't see anything in the logs. - Anyway time to give this another spin! - I am going spend the night with B200want to see how fast it can Flux Dreambooth a checkpoint, it spits out Flux images in around 1 -2 seconds its wild!

Roni Bliss

Working absolutely PERFECT now! - Great Work! This is somehow even more accurate than the original Beta One! - Did you add the "Con firm or Edit Prompt section? - I donbt remember seeing that anywhere on beta one before!! - Whatever the case being able to edit the prompt like that is an absolute GAME CHANGER! - Thanks so much for this!

SECourses

thanks. this is a new one's feature

JustLookingForNothin

Hi, could you please modify the installer so that it can (also) use Python 3.12? 3.10 is pretty old and most other current tools need 3.11 or newer. You did this Python selection code already nicely for the ComfyUI installer, so should not be a big deal to add this to Joy Caption, right?

SECourses

sadly not. just have python 3.10.11 installed in your system. it doesnt have to be default. my installer will use it

Christoph Behrmann

I am using batch processing with Beta One, and it is absolutely brilliant - I have for example 500 sub-folders - and it does NOT go thru (sub-)folders from A-Z, it looks like some random order... I was on the road for 2 days and it was doing its job for 200k+ images... just to let you know about "A-Z order NOT working on my end" ...better saying: I cannot make out any order :)

Ingo Weigel

trying to get joy Caption to work, even after several installations i got CUDA out of Memory errors. Anyone could help?? SECourses JoyCaption Image Captioning App - Select to Start 1. Joy Caption Pre Alpha 2. Joy Caption Alpha One 3. Joy Caption Alpha Two 4. Joy Caption Beta One 5. Exit Enter your choice (1-5): 1 Loading CLIP Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`. [2025-08-16 21:16:29,497] [INFO] [real_accelerator.py:239:get_accelerator] Setting ds_accelerator to cuda (auto detect) test.c LINK : fatal error LNK1181: Eingabedatei "aio.lib" kann nicht geöffnet werden. test.c LINK : fatal error LNK1181: Eingabedatei "cufile.lib" kann nicht geöffnet werden. W0816 21:16:37.515000 17928 venv\Lib\site-packages\torch\distributed\elastic\multiprocessing\redirects.py:29] NOTE: Redirects are currently not supported in Windows or MacOs. Loading tokenizer * Running on local URL: http://127.0.0.1:7860 * To create a public link, set `share=True` in `launch()`. Processing C:\Users\theow\AppData\Local\Temp\gradio\43d88ba8957c6994609b3a0f883f08115c192acb053be29e3d5236165d4ba990\image027.jpg on GPU 0... Loading model for GPU 0 Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:39<00:00, 9.91s/it] Error processing image C:\Users\theow\AppData\Local\Temp\gradio\43d88ba8957c6994609b3a0f883f08115c192acb053be29e3d5236165d4ba990\image027.jpg on GPU 0: CUDA out of memory. Tried to allocate 20.00 MiB. GPU 0 has a total capacity of 8.00 GiB of which 0 bytes is free. Of the allocated memory 7.05 GiB is allocated by PyTorch, and 194.77 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables) Image shape: (1365, 1024, 3) Processing C:\Users\theow\AppData\Local\Temp\gradio\210c49e2c309d72f744d0ba227a4b8cc1fdf827b312e37f6aca32b7df7840ba0\image028.jpg on GPU 0... Loading model for GPU 0 Loading checkpoint shards: 0%| | 0/4 [00:08

SECourses

did you try option 4 ? looks like option 1 cant run in 8 gb - also did you enable quantization options? check interface

SECourses

thank you for reporting with V36 now it will sort first and then process. just extract new zip file and overwrite older files

Jean Georges

Can you compare 2 image with this app? (Flux kontext training)

SECourses

can you eloborate more what you mean?

Jean Georges

I’m training a model with Flux Kontext. I have two groups of images: Group A (original) and Group B (retouched). The images are basically the same, except Group B has edits. I would like to know if Joy Caption can compare A and B and automatically generate captions describing the differences between them

SECourses

for this task i recommend use Gemini. i doubt that joy caption can do that. Gemini 2.5 pro may do it though it is best model.

Philipp Ludwig

i tried every link i could find, also this one, but all are invalid

Philipp Ludwig

"Invite Invalid" "This invite may be expired, or you might not have permission to join."

SECourses

tell me your discord username. if you account were hacked i could have banned it

Philipp Ludwig

the invite link just expired. you have to re-new it. i barely use discord and didn't joined so far but thank you taking time and invite me directl! lude1337

Joe

This is working very well where many others do not, thank you! Is it possible to have this unload the model, or clear cache / vram after generating a caption (or give a button to do this function if it isn't automatic)? In windows, when I generate the caption (using nf4), then put that caption into comfy, i see that my resources are still consumed ( I am running a 4080 super (16GB) with 64GB DDR5) and it forces me to restart joycaption. Also if I could save the checkbox preferences that's a nice to have. Thank you again for your content and contributions Edit: I should have mentioned, I am using Beta One

SECourses

it took a while but i added for you 22 August 2025 Update With V37 now there is extra option for Beta One app If you enable Unload model and clear VRAM after each caption before first run, it will fully unload model after each caption and clear VRAM

Joe

Amazing! Thank you so much Dr. Furkan Gözükara I extracted this over my current joycaption installation and replaced files, then ran the installer again to pick up any extra packages. (I also tried a fresh install). When I enabled the option to clear vram, I found this error in the console: ERROR: Subprocess failed with return code: 1 Subprocess errors: SyntaxError: Non-UTF-8 code starting with '\x97' in file C:\Users\me\AppData\Local\Temp\tmpzr5f44rl.py on line 87, but no encoding declared; see https://python.org/dev/peps/pep-0263/ for details Subprocess captioning failed: Subprocess failed with return code: 1 Starting subprocess for VRAM-isolated caption generation... ERROR: Subprocess failed with return code: 1 Subprocess errors: SyntaxError: Non-UTF-8 code starting with '\x97' in file C:\Users\me\AppData\Local\Temp\tmpjg8_d3xs.py on line 87, but no encoding declared; see https://python.org/dev/peps/pep-0263/ for details Subprocess captioning failed: Subprocess failed with return code: 1 With no other changes, I unticked the box for the clear vram option and everything ran as expected. It looks like that temp file removes as soon as the error comes up so I cannot see what is happening. I am still running nf4. My python version is 3.10.9.

SECourses

please try v38 and let me know just made a fix for this

Joe

Thank you again and thank you for the incredibly fast response. This is now working. Here is the output when running a caption: 🛠️ System configuration: Python : 3.10.9 (I:\Joy_Caption_v37\venv\Scripts\python.exe) PyTorch : 2.7.0+cu128 ‣ CUDA build : 12.8 transformers : 4.51.3 bitsandbytes : 0.47.0 liger_kernel : 0.5.8 GPUs (total 1): • [0] NVIDIA GeForce RTX 4080 SUPER | compute 8.9 | 16.0 GiB * Running on local URL: http://127.0.0.1:7860 * To create a public link, set `share=True` in `launch()`. Starting subprocess for VRAM-isolated caption generation... [Subprocess] Starting model loading... [Subprocess] Loading processor... [Subprocess] Loading model with nf4 quantization on cuda:0... [Subprocess] Model loaded successfully! [Subprocess] Loading image... [Subprocess] Preparing conversation... [Subprocess] Generating caption... [Subprocess] Caption generated successfully! [Subprocess] Cleaning up and terminating process (VRAM will be freed)... Subprocess completed! VRAM completely freed by process termination.

Christoph Behrmann

WARNING. For this auto installer to work you need to have installed Python 3.10.11, Git, FFmpeg, CUDA 12.8 and C++ tools --this is first line when starting 'Windows_Install.bat' -- when running with my 3090 I was thinking it does only work with CUDA 12.6 ?

SECourses

no it definitely works with cuda 12.8

SECourses

you did set hugging face token and it is expired. open a cmd and run this : huggingface-cli logout and restart download

doug m

Windows, I'm getting "❌ Error: Error during folder batch processing: 'LlamaModel' object has no attribute 'model'"

SECourses

i need install logs. probably your install had errors. please make a fresh install and email me all cmd logs : monstermmorpg@gmail.com

benjamin macauley

I haven’t upgraded since V19, and I’m really impressed with the improvements. Thank you for all the effort you’ve put into this.

Hockey

Is it possible to add an option in Extra Options of how the object such as a car is viewed. Like the front and side of the car is visible of the image.

SECourses

you can write custom prompts. or you mean something else?

Hockey

In: "Extra Options (for non-custom prompts)" You can tick various boxes, for instance: "Include information about camera angle" I would like a box that says: "Include information about the object/person angle" Hope that makes sense.

SECourses

not possible but gonna add that for you now

Hockey

Thank you, you are the best! =)

SECourses

21 September 2025 Update With v40, we have added full preset save and load feature to Joy Caption Beta One You can now even add new Extra Options (for non-custom prompts) It will remember all your saved settings in preset and auto load next time you restart your app Presets will be saved inside presets folder Extract zip file and overwrite previous files for update that is sufficient to upgrade from V18

eman eliforp

What is the best tool for dataset exploration? It seems I have too many repeating tags and very few unique ones. I’d like to thoroughly review a 'cloud' of the data so that I can sort or remove images based on it

eman eliforp

Somehow I managed to find applications: “FiftyOne” , "fastdup" and “Know Your Data” all quite complex.. If you can add something in to app you mentioned, it would be more useful than just showing the number of tags

SECourses

it has Token Counter and word check. what you need exactly?

eman eliforp

Example: there are about 100 different cows. But they all share the same characteristics: horns, texture, tail, udder, environment/location, etc. So if I now want to describe the skin texture, there’s no way for me to identify which entries repeat: white-black, brown, white with black spots, brown-white with small spots. Even if I tag them manually, I get lost and can’t remember which tags I’ve already used. I’d like to be able to visually see what percentage exists and which specific cows have been tagged. A text-based representation no longer works for me. Visual representation is missing, like a semantic word grouping cloud, can be grouped by relationships. That way, I’d understand not only which tail colors are described, but also their lengths and exactly which cows have them. At the same time, a frequency chart would help. If there are obviously too many entries, I’d know where to focus. For example, I might realize that I need a separate grouping for the cow’s nose texture, which I wouldn’t notice just by looking at numbers since it’s a completely different category. So basically.. I want as much detail as possible for a large number of images, while ensuring tags don’t repeat - neither in the tag list nor in the images themselves. Something similar to the program 'Obsidian,' where the Graph is displayed visually and can be edited directly. So, if I add a new tag, a new node should appear in the graph with the corresponding grouping. I could select a group of images that either still lack data or already have excess, which I could then remove

SECourses

can you visualize demo what you want and email me? then i, may work on it : monstermmorpg@gmail.com i need definitive clear example since this looks complex task

Tyler

Is there a 'how to best use Joycaptions' video or guide?

Tyler

ty, more specifically, it's my understanding that SDXL wants TAGS, not plain English captioning to best caption images. Is there a way to get that out of joycaption?

SECourses

joy caption supports different scenarios and custom prompting. you can try. but still i recommend ohwx man as a caption. what are you training? training multiple concepts? by the way tagging style depends on used base model as well.

Tyler

Training and SDXL finetune for lora extract of a female. Ideally, the lora would be body and face. Planning on using cyberrealisticXL or EpicRealismXL as the base to train with.

SECourses

ohwx woman probably best caption. also i am doing 32th training experiments right now for new sdxl lora training workflow. so far first 24 didnt work well. will publish hopefully when i get hopefully better results

Tyler

Do you mean use ONLY ohwx woman and don't caption the images? Or both? (for sdxl* ie include ohwx woman, + (caption from joy caption)

SECourses

just ohwx woman. this works best imo since we are training a single person / subject / concept.

Tyler

Well that makes SDXL trainings a little easier. So 15 face images, 15 portrait, 15 full body, just ohwx woman in the fields on kohya before populating the folder structure, and I'm gtg? Great ty!

SECourses

yes true. but add more closer shots because SDXL is very poor at full body shots sadly