LoRAs
Train a LoRA for character or object consistency in AI renders. Upload references, attach to a Build-mode object, reuse the same look across every shot.
A LoRA is a small fine-tuning file that biases the diffusion model toward a specific style, character, or product. Attach one to a Build-mode object and the visualizer applies the LoRA's behavior at render time, on top of whatever the base model would otherwise do.
Image Reference vs LoRA vs Custom Style
Three different ways to lock visual consistency. They serve different jobs.
Scope
One specific object
One object, model-level bias
Whole render, every shot
Controls
What this object looks like
A learned style or character bias
Project-wide brand treatment
Setup
Drop in a photo
Train or download .safetensors or .pt
Upload references in Visualize, save
Best for
Brand-true product, hero with photos
Illustrator style, internal aesthetic, characters without photos
Brand- or campaign-wide look
Cross-shot consistency
Yes, automatic
Yes, automatic
Yes, automatic
Plan requirement
All plans
All plans
All plans
Reach first when
Almost always start here
Can't capture the look in a single photo
The style is bigger than any one object
The fastest correct answer for most agency briefs: try Image Reference first. Layer in LoRA when one specific character or style needs more than a photograph can carry. Switch to Custom Style when the brand language has to apply globally.
What it does
Image references handle most consistency cases (see Image reference) and they don't require a fine-tuning step. LoRAs cover the cases image reference can't:
Cross-object style consistency. A whole rendered scene in a specific illustrator's pen-and-ink style isn't a per-object reference – it's a model-level bias. LoRA on a global "scene style" object handles it.
Character that needs more than one reference image can capture. A custom character with many expressions, costumes, and angles benefits from a trained LoRA more than from a stack of reference images.
Brand-specific surface treatments. A material library that the team has trained internally as a LoRA and wants to apply consistently.
If you don't already have a LoRA on hand, you probably don't need one. Image reference is the right starting tool.
How to use it
LoRAs attach per object, via the contextual menu's More Options menu.
Select the object in the viewport. The contextual menu opens above it.
Open More Options – the three-dot (⋯) icon at the right end of the contextual menu.
Click Add LoRA. A dialog opens with inputs for the file (or URL) and an optional trigger word.
Upload from computer. Browse to a
.safetensorsor.ptfile. The file gets stored alongside the project; subsequent renders reuse it without re-uploading.Add via URL. Paste a URL from a service that hosts LoRA files (Civitai, Hugging Face, internal artifact storage). The system fetches and caches.
Trigger word. Some LoRAs are trained to activate only when a specific word appears in the prompt. Enter the trigger word here; the visualizer prepends it to the prompt block for this object.
Once attached, the LoRA is bound to that object. The visualizer's auto-prompt notes the LoRA's presence in the [Subjects] block at render time.
LoRAs apply per-object, not per-scene. Attach the LoRA to the object whose look should be biased. To bias the entire scene's style, attach to a "global style" placeholder object, or use the Custom styles feature in Visualize mode for the scene-wide case.
How LoRAs interact with image references
Both can apply to the same object. The model receives the prompt text plus the image reference plus the LoRA bias at generation time. In practice:
Image reference dominates for "what does this specific thing look like". A photograph of the actual car constrains the rendered car's appearance directly.
LoRA dominates for "what is the style of this rendering". The pen-and-ink LoRA bends every render through that aesthetic regardless of the reference image's surface.
Use both when both apply. They don't fight; they layer.
Limits and known issues
Not every base model honors LoRAs. Some of the visualizer's models (notably some video models) don't accept LoRA conditioning. The LoRA attaches successfully but is silently ignored at render time. Test against your target model before committing.
LoRA files are large. Multi-hundred-megabyte LoRAs slow project sync and increase storage costs. Use compressed formats where possible.
Trigger words must be exact. A LoRA trained on the trigger "robocar" won't activate from "robo car" or "robotic car". Match the training data exactly.
Where to find LoRAs
Civitai – community library for image-model LoRAs. Most are stylistic biases.
HuggingFace – broader model repository; LoRAs live alongside their parent base models.
Internal artifacts – if your agency or team has trained custom LoRAs, host them in your own storage and share URLs.
Related
Last updated
Was this helpful?