Choose where captions run
Use local vision models or connect Claude and OpenAI. Cloud requests go straight to the provider; your output files are saved locally.
1.3.0 Image preparation for AI creators
Prepare the images.
Then train.
Caption with Claude, OpenAI or local models. Upscale, tag and organize images for LoRA and other AI training workflows, all from one desktop app.
TrainKit turns chelsea.png into chelsea.txt with the caption: a photograph of a cat sitting on a bed.
New in TrainKit 1.3.0
Give Claude the context.
Get captions for your dataset.
A caption for a character LoRA needs different details than one for a product collection. Tell Claude what to describe, choose an image or a folder, and save a text caption for each image.
Connect your Anthropic API key in TrainKit’s API tab, then select Claude in Caption. You can also use OpenAI or run a local vision model.
See the API workspaceCAPTION INSTRUCTIONExample
Caption this image for a character LoRA. Describe the pose, clothing, expression, camera angle and background. Return a concise natural-language caption.
Your instruction is applied to each image in the batch.- 01
Choose your images
One image or a folder, with your caption instruction.
- 02
Claude reads each image
TrainKit sends a resized image and your instruction directly to Anthropic.
- 03
Keep the captions
UTF-8
.txtfiles are saved in your chosen output folder.
Cloud captioning uses your provider’s API credits. Local captioning, upscaling, tagging and renaming run on your computer. How data is handled ↗
TrainKit 1.3.0 · API workspace
Inside the toolkit
From an image to useful output.
Real inputs. Real runs.
Explore the results below.
Text that belongs with the image.
Use Claude, OpenAI or a local vision model to caption an image or a folder. TrainKit writes a UTF-8 text file for each image. This sample shows a local run with BLIP base.
chelsea.txt
a photograph of a cat sitting on a bed
BLIP base · prompt: “a photograph of”Download the sidecar ↓See the captioning workspace
More pixels. A visible difference.
This 128 × 85 image was processed through TrainKit with the NCNN Real-ESRGAN x4plus model. Drag the divider below to compare the input and the actual 512 × 340 result.
The Real-ESRGAN x4plus model upscales a 128 by 85 image to 512 by 340. The animation compares these real input and output files.
Input · 128 × 85TrainKit · 512 × 340Drag to compare. Use arrow keys when the slider is focused.Open the output PNG ↗
See the upscaling workspace
Labels you can take into training.
Run a local image classifier, filter by score and choose how many tags to keep. Export text, scored JSON or both. These are the unedited predictions from the sample run.
Modelmicrosoft/resnet-18SettingsThreshold 0.01 · Top-K 5ExportJSON + text sidecars
chelsea.tags.json
Egyptian cat
0.397884tiger cat
0.394540tabby, tabby cat
Download the actual JSON ↓0.203333Scores are this model’s predictions; results depend on the model you choose.
See the tagging workspace
A folder you can work with.
Rename files in natural order, choose number padding and keep the originals. This replay shows the verified mappings from a three-image batch run.
OrderingNatural filename orderNumber width3 digitsOperationAtomic copies
IMG_2.png becomes 001.png, IMG_10.png becomes 002.png, and IMG_23.png becomes 003.png. The original images are preserved.
See the batch renaming workspace
Sample photo: Chelsea by Stefan van der Walt, CC0 via scikit-image. Batch preview photo: NASA, public domain. These local processing examples were recorded with TrainKit v1.2.1.
Keep the control
Work through the whole folder.
Batch tools with progress
you can come back to.
Resume a stopped run
Cancel a run and resume it using a saved manifest. Completed and skipped files keep their progress.
Handle existing files
Choose how output collisions are handled, keep track of each file’s status and preserve your original images.
Choose your plan
Free or Pro.
Pick the tools that fit your workflow.
Free
Available · 1.3.0$0 to download
The open-source desktop app. Run local models or connect your own cloud API key.
- Local captioning, upscaling, tagging and renaming
- Claude and OpenAI captioning with your own API key
- Batch processing and resumable runs
- Full source code under the MIT license
Cloud usage is billed by your API provider.
Download FreePro
$25 / month
$240 / year · $20 per month
A customized TrainKit app with account login and hosted image processing.
- Everything in Free
- Cloud captioning through your TrainKit account
- Cloud image upscaling
- Access to more hosted models
- 1,000 processing credits each month
Install the essentials
Install uv and the Visual C++ x64 runtime.
Extract and launch
Unzip the release into a writable folder. Open
TrainKit.exeand let setup finish.Choose your caption provider
Connect your Claude or OpenAI API key, or select a local model. Choose your images and output folder, then start a run.
First-launch requirements
First-launch setup needs internet access and several gigabytes of free space. Local model files are selected separately. An NVIDIA GPU is recommended for local captioning and Spandrel; NCNN supports CPU and compatible Vulkan GPUs. Cloud captioning needs an API account with credits and does not require a local GPU.
Read the installation guide ↗