Looks and LoRAs
Applying one look to a set
Stack your trained LoRAs, weight each one, and use a card built for firing twenty takes rather than one.
If you need twenty images that belong together, generating them one at a time and hoping is the slow way to find out they do not. A trained LoRA plus a card built for volume is the fast way. ## Generating with one **lora-generate** takes one or more LoRAs from your library, each with **its own scale**, and a prompt carrying your trigger words. Scale is the dial worth learning. Low, and the adapter suggests; high, and it dominates and starts to overwhelm everything else in the prompt. The useful range is usually narrower than the slider suggests, so it is worth a few runs to find. :::caution[FLUX.1 and FLUX.2 LoRAs cannot mix] They run through different endpoints, so a single run takes one generation or the other. Two adapters from the same generation stack fine; one from each does not. ::: ## The card is locked by default **lora-generate** ships locked, which is unusual and deliberate. It stays open, every run scatters into a grid, and you can fire the next while the last is still rendering. That is built for the actual LoRA workflow, which is twenty takes and pick three rather than one considered attempt. Seeing the twenty side by side is what tells you whether a look is holding across a set rather than in one lucky frame. ## Stacking two A product LoRA and a style LoRA together, each at its own scale, gives you your object rendered in your house look. Start with the subject high and the style low. A style adapter at full strength will happily redraw your product into something that merely resembles it, which defeats the point of having trained it. ## Keeping a set consistent The habits that make twenty images belong together: - **Freeze the prompt except for one clause.** Change the scene, keep everything else identical, so the differences between images are the ones you chose - **Use a `{ }` variable** for that clause and fire the set as one batch, which guarantees every other word is identical by construction - **Keep the scales fixed** across the set. A scale change is a look change - **Judge them together**, in the grid, rather than one at a time [Variables and batch runs](/docs/using-plnty/using-a-tool/variables-and-batch-runs/) covers the batch mechanics. ## In 3D **3D LoRA Generate** does the same for meshes: drop an image, pick a trained style, and the model comes out in that visual language. One constraint to know before you plan around it: when a style is active the resolution locks to 512, because the adapters train at 512 and inference has to match. The detail steps, Fast, Standard and High, cost nothing extra and are the dial you have instead. :::tip[A LoRA is a style guide that executes itself] A written brand guideline describes a look and relies on everyone reading it. A trained adapter produces it. For a team generating a lot, that is the difference between a document people cite and a result people ship. :::If you need twenty images that belong together, generating them one at a time and hoping is the slow way to find out they do not. A trained LoRA plus a card built for volume is the fast way.
Generating with one
Section titled “Generating with one”lora-generate takes one or more LoRAs from your library, each with its own scale, and a prompt carrying your trigger words.
Scale is the dial worth learning. Low, and the adapter suggests; high, and it dominates and starts to overwhelm everything else in the prompt. The useful range is usually narrower than the slider suggests, so it is worth a few runs to find.
The card is locked by default
Section titled “The card is locked by default”lora-generate ships locked, which is unusual and deliberate. It stays open, every run scatters into a grid, and you can fire the next while the last is still rendering.
That is built for the actual LoRA workflow, which is twenty takes and pick three rather than one considered attempt. Seeing the twenty side by side is what tells you whether a look is holding across a set rather than in one lucky frame.
Stacking two
Section titled “Stacking two”A product LoRA and a style LoRA together, each at its own scale, gives you your object rendered in your house look.
Start with the subject high and the style low. A style adapter at full strength will happily redraw your product into something that merely resembles it, which defeats the point of having trained it.
Keeping a set consistent
Section titled “Keeping a set consistent”The habits that make twenty images belong together:
- Freeze the prompt except for one clause. Change the scene, keep everything else identical, so the differences between images are the ones you chose
- Use a
{ }variable for that clause and fire the set as one batch, which guarantees every other word is identical by construction - Keep the scales fixed across the set. A scale change is a look change
- Judge them together, in the grid, rather than one at a time
Variables and batch runs covers the batch mechanics.
3D LoRA Generate does the same for meshes: drop an image, pick a trained style, and the model comes out in that visual language.
One constraint to know before you plan around it: when a style is active the resolution locks to 512, because the adapters train at 512 and inference has to match. The detail steps, Fast, Standard and High, cost nothing extra and are the dial you have instead.