Home Creators Posts Import Register Favorites Logout

Downloads

Content

Ultimate Image Captioner Pro writes JSON and text captions with Qwen3 VL 8B Instruct and Joy Caption models for Ideogram 4 JSON prompting and training.

Full JSON and text captioning with Joy Caption Beta 1, Alpha 1 & 2 and Pre Alpha or Qwen3-VL 2B to 30B and Qwen3.6 27B, with a fully working JSON Prompt Builder, saved outputs and Torch Compile that speeds up Qwen models 84%. 1-click install on Windows, RunPod, Massed Compute, SimplePod and Linux with Torch 2.13 and CUDA 13.

Our Patreon exclusive posts index list all of the apps we have (over 100+ AI apps, scripts, trainers, presets and more). Use CTRL+F to find whatever app you looking on our index.

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)

=======

Latest installer zip file : Ultimate_Image_Captioner_Pro_v2.zip

  • [click here to choose a membership and Join to download zip files]

  • With Ideogram 4 model as you know JSON prompting is now a thing and we need JSON prompts for both inference and training

  • Therefore, a new app was necessary to solve this issue and I built the very best local app out there for this task

  • Fully supported model list is as below with fully working robust Torch Compile

    • Qwen Vision Models

      • Qwen3-VL 8B Instruct (default)

      • Huihui Qwen3-VL 8B Instruct Abliterated

      • Qwen3-VL 4B Instruct

      • Qwen3-VL 2B Instruct

      • Qwen3-VL 30B-A3B Instruct

      • Qwen3.6 27B

      • Huihui Qwen3.6 27B Abliterated

    • Joy Caption Models

      • Joy Caption Beta 1

      • Joy Caption Alpha 2

      • Joy Caption Alpha 1

      • Joy Caption Pre Alpha

  • Full features of the app introduced below with screenshots so please read

    • The installer will auto download all the necessary models with 16 connections + SHA256 hash verification

  • Full tutorial video : https://youtu.be/TW3MRdd0MV4

  • SwarmUI and ComfyUI zip files updated for Ideogram 4 model, model downloads and presets and workflows already added

  • Windows Requirements

  • For RunPod, SimplePod, Massed Compute and Linux please follow:

    • Massed_Compute_Instructions_READ.txt

    • Runpod_SimplePod_Ultimate_Caption_Instructions.txt

  • The application runs on Torch 2.13 with CUDA 13, supports literally every GPU out there including server GPUs

    • Moreover, we are using latest libraries that I compiled as below

19 July 2026 V2.1

  • Application upgraded to newest Torch 2.13 with above seen pre-compiled wheels that works perfect

    • So either do a fresh install or, extract latest zip file, overwrite installers, delete venv folder inside Ultimate_Image_Captioner_Pro and run installer again for update

    • The following models are fully supported now

      • Qwen3-VL 8B Instruct (default), Huihui Qwen3-VL 8B Instruct Abliterated, Qwen3-VL 4B Instruct, Qwen3-VL 2B Instruct, Qwen3-VL 30B-A3B Instruct, Qwen3.6 27B, Huihui Qwen3.6 27B Abliterated

  • We have implemented fully working torch compile feature that speeds up Qwen Vision Models 84% and works on all Joy Captions as well

  • Moreover, now we are using Python 3.12.10 so please install Python 3.12.10 if you don't have yet

  • For Torch Compile to work, you need to have installed Visual Studio Community Edition with All c++ options

  • New Qwen Image models implemented that fully works

  • They will be downloaded fully automatically when you first time use them, only default Qwen3-VL 8B Instruct is auto downloaded with the initial installation

  • Joy Caption Alpha 2, Joy Caption Alpha 1, Joy Caption Pre Alpha models will not be auto downloaded anymore with initial installation

  • They will be auto downloaded when you first time use them

  • All Joy Caption outputs are now displayed with text-wrapping and has copy generated prompt button feature

3 July 2026 V1.2

  • Qwen image captioning made more robust

    • Such as in some cases it was adding imgur links and not anymore this bug exists

  • Apply Box Edits button is now Apply Box Edits & Save so every edit is automatically saved in the outputs folder

    • Overwrites generated json file and re-generates boxed image

  • The changes you made in JSON Box Preview or JSON Elements were not being saved in outputs folder and now they will be saved

  • JSON Prompt Builder significantly improved

    • Now it will accurately recognize selected file's accurate outputs folder path and all changes will be saved

    • If your file is not in outputs folder, use Browse File to load file

    • Now lets say you started empty design and saved, it will be saved in a new folder inside outputs folder and keep using that folder as long as you work on that json prompt

    • When you were switching between folders, it was not properly updating displayed values and this issue fixed

  • Now when you switch Saved Outputs tab it will auto refresh and show latest

    • Now Saved Outputs tab is auto sorted by latest but you can re-sort by clicking display headers

  • To update just run Windows_Install_Update_App.bat file

    • Zip file is still same

Ultimate Image Captioner Pro Features

Click on images to see them full resolution

  • 1-Click to install on Windows, RunPod, Massed Compute (Linux users please use Massed Compute scripts) and SimplePod

  • Fully support JSON prompt generation

  • Full custom user preset save and load, after restart remembers last saved / used one

  • Fully edit generated json values content or boxes and reconstruct json and easy 1-click to copy

    • Hide / display boxes to easy work, fully drag to change position or resize and make bigger or smaller from interface

  • We have got 35 Ideogram 4 presets ready to select and use

  • Fully working automatically selected GPU VRAM presets for both Qwen and Joy Caption models

  • You can run as subprocess thus it will leave 0 VRAM or RAM usage after captioning

  • Moreover, you can set which GPU ID to run captioning on, or set multiple GPUs to distribute batch captioning

  • Auto save box drawn images

  • Add suffix, prefix or word replaces to generated captions automatically

  • All Joy Caption Models are supported with full features like Qwen (e.g. batch captioning, VRAM presets, various preset prompts and save options, etc.)

  • Fully working JSON prompt builder that you can build from scratch or load existing image

    • Add boxes, write info, modify boxes, move them, resize them, etc. all fully working

  • If you load existing image, if that image has JSON file, it will be auto loaded

    • This way, you can load your previous generations and modify and work further on them

  • Every generated output is saved in outputs folder with full metadata as well

  • Fully working view saved outputs screen that you can quickly navigate and find your previous generations and see them with full details, info, etc.

    • The page has full filtering and pagination features so use them as well

  • CMD screen shows full details of what is happening even token / second as well

    • Following CMD outputs is especially useful for batch folder processing

  • Qwen JSON prompt generator is so amazing that it has a specific field for written text on images, analyze below image to understand logic

  • Pay attention to JSON Elements table you will see text

  • You can fully edit the JSON Elements table and re-generate, auto generated JSON as you wish

    • When you click Apply Box Edits it will overwrite generated JSON file and boxes drawn image in respected output folder

  • Saved output json files are beautified and saved - can be disabled if you wish

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

Ec Jep

thx for this tool! do you recommend performing new training (QWEN 2512 for example) to take advantage of the json captioning? If I understand you, this makes the trained model more flexible with prompting

SECourses

well json prompting is useful for Ideogram 4. but i am not sure if can help Qwen 2512 need to be tested

Ec Jep

I have not been able to run an image file. I have installed it and keep getting errors with the joy options or qwen3: Traceback (most recent call last): File "D:\SD\image_captioner_v1-1\Ultimate_Image_Captioner_Pro\venv\Lib\site-packages\gradio\queueing.py", line 867, in process_events response = await route_utils.call_process_api( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ... File "D:\SD\image_captioner_v1-1\Ultimate_Image_Captioner_Pro\venv\Lib\site-packages\torch\__init__.py", line 268, in _load_dll_libraries raise err OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed. Error loading "D:\SD\image_captioner_v1-1\Ultimate_Image_Captioner_Pro\venv\Lib\site-packages\torch\lib\c10.dll" or one of its dependencies. Errors listed above image input: Loaded universal preset 'try01'. CUDA GPU was not detected. Defaulting VRAM preset to 6 GB.

SECourses

what is your nvidia driver version? open cmd type nvidia-smi tell me this info : NVIDIA-SMI 596.21 Driver Version: 596.21 CUDA Version: 13.2

Ec Jep

NVIDIA-SMI 591.86 Driver Version: 591.86 CUDA Version: 13.1

SECourses

ok then do this. make a fresh install and email me install logs : monstermmorpg@gmail.com entirely

EvanGuo

do u have preset for trainning style

SECourses

for ideogram not yet. or you are interested in other models too?

Douse

I tried Ace 1.5 everything loaded and seemed fine, but when I clicked Generate, an error popped up saying something wasn't installed. Background generation failed: Initializing DiT service: ACEStep_1_5_XL_Turbo_BF16 Error initializing model: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device. Traceback: Traceback (most recent call last): File "/workspace/ACE-Step_Premium/acestep/core/generation/handler/init_service_orchestrator.py", line 243, in initialize_service text_encoder_path = self._load_text_encoder_and_tokenizer( File "/workspace/ACE-Step_Premium/acestep/core/generation/handler/init_service_loader_components.py", line 111, in _load_text_encoder_and_tokenizer self.text_encoder = self.text_encoder.to(device).to(self.dtype) File "/workspace/ACE-Step_Premium/venv/lib/python3.11/site-packages/transformers/modeling_utils.py", line 3587, in to return super().to(*args, **kwargs) File "/workspace/ACE-Step_Premium/venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1384, in to return self._apply(convert) File "/workspace/ACE-Step_Premium/venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 934, in _apply module._apply(fn) File "/workspace/ACE-Step_Premium/venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 965, in _apply param_applied = fn(param) File "/workspace/ACE-Step_Premium/venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1377, in convert raise NotImplementedError( NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

SECourses

hi please delete venv folder run installer and send me entire CMD log as email : monstermmorpg@gmail.com

Douse

i fix it just restart thx)

Glen Carpenter

One suggestion - instead of downloading all of the models with the installer, you could defer the download to when the user first uses a model. Personally I don't have much use for JoyCaption pre alpha / alpha 1 / alpha 2 and they do consume a large amount of disk space and take a long time to download.

SECourses

sure will try to make this way when updating in future.