Training & Using LoRAs

Train a custom AI model on your own images, then call it in generations with a trigger word. LoRA training lives in the dashboard; applying a LoRA happens in the chat bar.

Create a LoRA Model

To start a new LoRA:

1. Open the LoRA Training screen in the dashboard.
2. Click Create New LoRA (or the Menu button at the top of the center panel, then New).
3. Enter a descriptive name (e.g. "My Custom Style") and press Enter or click Create.

In the left LoRA Controller panel, pick the Training Model:

• Flux LoRA — requires a trigger word.
• Z-Image Turbo — trigger word optional.
• Z-Image Turbo V2 — trigger word optional.

Set a Trigger Word (the word you'll type in prompts to activate the LoRA). Click the wand icon next to the field to generate a random one. For Flux it's required; for Z-Image models it's an On/Off toggle.

Collect & Organize Training Images

The center panel holds your training set; the right Image Library panel is your source.

To add images:

• Click images in the right Image Library to add them. Shift+click adds a range. Filter the library by AI Generated, Rendered, Upscaled, or Uploaded badges, or use the search box.
• Click the empty drop area or the Add More tile to upload files or pick from your library.
• Paste with Ctrl+V — from the clipboard, from a copied canvas image node, or from the Framo Chrome extension.

Remove an image with the X on its tile. Use Clear All (top right) to empty the set. Recommended count is 15–25 images (minimum 10, maximum 30).

Pointer: training images come from the same library as your generated/uploaded images across Framo, so anything you make on the canvas can be reused here. See Projects, media library & stock.

Crop Non-Square Images

Training needs square (1:1) images. Non-square images are center-cropped to 1:1 automatically, but you can adjust the crop.

To crop: click an image tile (or its Crop button, bottom-left — amber when an image is non-square). In the Image Crop Dialog, drag to reposition the 1:1 crop, then save. The dialog can also AI Extend a non-square image to 1:1 or Remove Watermark (both cost credits); accept the result to replace the image in your set.

Auto-Caption (Flux Training)

Good captions describe everything except what the LoRA should learn. Click Generate Captions in the left panel (or Auto Caption from an image's caption dialog).

For Flux, pick a caption category describing what the model learns implicitly:

• Design Language — shape, form, proportions, silhouette.
• CMF — color, material, finish.
• Light & Mood — lighting, atmosphere, post-processing.
• Artistic Medium.

Choose a captioning model (Gemini 2.5 Flash, GPT-5 Mini, Claude Sonnet, or GPT-5). Auto-captioning costs credits per image; Skip existing avoids re-captioning images that already have a caption. You can edit any caption by clicking the message icon on its tile.

Auto-Caption (Z-Image Training)

Z-Image captioning is built around a Training Focus (set in the left panel):

• Style — captions describe what is depicted, not the visual style.
• Content — captions describe context/variations, not fixed identity (use a trigger word; for content you also pick Person or Object).
• Balanced — captions describe both content and style.

Z-Image supports two Caption Modes:

• Single — one master caption applied to all images.
• Per Image — an individual caption per image.

Captions are optional for Z-Image. If a trigger word is enabled, it's prepended to generated captions. Use the same Generate Captions button; choose a captioning model and cost applies per image.

Configure Training Parameters

In the left LoRA Controller:

• Training Steps — drag the slider or type a value. A Recommended range (about 30–50× your image count) shows below; click Apply to use it. More steps = better quality but longer training.
• Z-Image models expose extra options (e.g. learning rate) under the Training Parameters section.

A Training Status Summary shows whether Name, Trigger, Images, Captions, and Steps each meet requirements (green = ready, red = missing).

Submit a Training Job

When everything is green, click Start Training at the bottom of the left panel. The button shows the estimated credit cost. Confirm you have a name, a valid trigger word (where required), 10–30 images, and captions where required.

Track Training Progress / Status

Once submitted, the left panel shows a live Training Progress bar with status: In queue (with position), Training in progress, or Processing results, plus the latest log line and a percentage.

A model's status is shown as a banner: Draft (editable), Queued, Training, Successfully Trained (read-only), or Training Failed (read-only). Only Draft models can be edited; submitted/completed models open read-only.

Browse Completed LoRAs & Manage Drafts

Open the Menu button at the top of the center panel:

• Open — lists your models split into WIP (preparing/queued/training/processing) and Completed. Pick one to load its images, captions, and settings.
• Save — saves the current draft (images, captions, crops, settings also auto-save).
• Load Dataset — copy the image set + captions from another of your models into the active draft.
• Delete — permanently removes the model and its training data.

From the empty-state screen you can also use Create New LoRA or Open Existing LoRA.

Apply a LoRA in Generation (Trigger Word)

To use a finished LoRA in a generation, go to the chat bar / generation controls (see Chat bar: generating images):

1. Open the LoRA Models selector and click Add LoRA (the + / browser).
2. Pick from your own completed LoRAs or community LoRAs. Up to 4 LoRAs can be combined.
3. Each selected LoRA shows its Trigger word (copy it with the copy icon) and a Scale control. The slider covers the comfortable range; type a higher value (up to the API max of 4) to over-push. A Total Strength indicator warns when the combined scale gets high.
4. Put the trigger word in your prompt to activate the LoRA. When a LoRA with a trigger word is added, Framo can auto-insert the trigger phrasing (e.g. "in the style of <trigger>") for you.

Pointer: trigger words for Z-Image content LoRAs control when your subject appears; without one the subject may bleed into every generation. Style/object type also affects how the trigger is phrased in the prompt.