ComfyUI workflow
How to use lora in comfyui becomes straightforward once you separate model loading from LoRA application. This guide shows where each node belongs, how to set weights, and how to verify the result before changing your prompt.
Choose your route
Start with the situation that matches your graph. The right path depends on whether you are building a new workflow or adding Lora AI to a checkpoint that already works.
Pick your starting point
Choose the card that describes your current graph, then apply the linked guidance without rebuilding more than necessary.
You have a checkpoint file but no working graph yet. Start with a basic text-to-image workflow, then place Lora AI between checkpoint loading and sampling.
You get a clean baseline that makes it easier to tell whether the LoRA or the rest of the graph is responsible for a change.
how to use lora ai redditYour prompt and checkpoint already work in another interface, and you want the equivalent ComfyUI node arrangement.
You can compare the same model, prompt, and approximate weight rather than troubleshooting several variables at once.
how to use lora stable diffusionYou are experimenting with a trained adapter for a character, product, or visual style and need a repeatable loading pattern.
You can keep the base checkpoint stable while testing one Lora AI file at a time and recording its weight.
lora training guideA ComfyUI graph already produces images, but the added LoRA appears to do nothing or changes the image too aggressively.
You can inspect file compatibility, node connections, and model and CLIP strengths in a controlled order.
ai lora modelsBuild the graph
This is the dependable three-part sequence for adding a LoRA to a working ComfyUI text-to-image workflow.
Use CheckpointLoaderSimple to load the base model. Keep its MODEL and CLIP outputs visible, because both need to pass through the LoRA node before they reach the rest of the graph.
Add Load LoRA and connect the checkpoint MODEL to its model input and the checkpoint CLIP to its clip input. Select the LoRA file from the node's model_name field.
Send the LoRA node's MODEL output to the sampler model input and its CLIP output to the text-encoding nodes. Keep the VAE path from the checkpoint unless your workflow specifically uses another VAE.
Connection check
Compare the two sides before you render. The left side is the base graph; the right side is the same graph with one Lora AI adapter applied.
Base checkpoint only
Checkpoint plus LoRA
Base checkpoint only
CheckpointLoaderSimple supplies the base MODEL.
Checkpoint plus LoRA
The same MODEL enters Load LoRA first.
Base checkpoint only
Checkpoint CLIP goes directly to text encoding.
Checkpoint plus LoRA
Checkpoint CLIP passes through Load LoRA before text encoding.
Base checkpoint only
No LoRA file is selected.
Checkpoint plus LoRA
One compatible LoRA file is selected in model_name.
Base checkpoint only
No adapter strength is applied.
Checkpoint plus LoRA
Set strength_model deliberately and change one variable at a time.
Base checkpoint only
Text conditioning uses the checkpoint CLIP unchanged.
Checkpoint plus LoRA
Set strength_clip separately when the adapter supports or needs it.
Base checkpoint only
The sampler receives the original checkpoint MODEL.
Checkpoint plus LoRA
The sampler receives the MODEL output from Load LoRA.
Base checkpoint only
Useful for confirming the checkpoint and prompt work alone.
Checkpoint plus LoRA
Useful for testing whether the LoRA changes the intended subject or style.
Base checkpoint only
There is no adapter chain to inspect.
Checkpoint plus LoRA
Chain Load LoRA nodes only after one adapter works by itself.
Node anatomy
Know the limits
The adapter modifies a compatible base model; it is not a complete image model and cannot produce a useful result without the right checkpoint family.
WorkaroundLoad a compatible checkpoint first, then test the adapter on a simple baseline prompt.
If the base workflow has disconnected conditioning, an invalid sampler route, or a missing VAE, adding an adapter will not repair those problems.
WorkaroundRender the checkpoint without LoRA, confirm that baseline works, and only then insert Load LoRA.
Some files respond to a trigger word, some respond to descriptive prompting, and some depend heavily on the training captions.
WorkaroundRead the model notes, try the recommended trigger, and compare one short prompt with one descriptive prompt.
A file made for a different architecture or checkpoint family may load poorly, have little effect, or produce distorted output.
WorkaroundMatch the adapter to the model family and keep a known-compatible checkpoint for testing.
Visual comparison
The useful comparison is not simply prettier versus worse. Check whether the intended subject, clothing, pose, texture, or style changes while composition and prompt meaning remain understandable.
Try the workflow
Once the ComfyUI graph is connected, use a focused prompt and a single compatible adapter to establish a baseline. You can then carry the same idea into Lora AI generation, compare outputs, and return to ComfyUI with a clearer target.
Try Lora AICommon questions
Load your checkpoint with CheckpointLoaderSimple, add a Load LoRA node, and connect the checkpoint MODEL and CLIP outputs to that node. Then connect the adapted MODEL to the sampler and the adapted CLIP to your text-encoding nodes before selecting the LoRA file.
Place it after the checkpoint loader and before the sampler and text encoders. The model and CLIP paths both need to pass through Load LoRA, while the VAE connection can usually remain on its existing route.
Check that the file is stored where ComfyUI can detect it, that the selected checkpoint belongs to the same model family, and that both MODEL and CLIP are connected. Also test the recommended trigger word and compare against a render made without the adapter.
strength_model controls how strongly the adapter changes the denoising model, while strength_clip affects the text-conditioning side. They are separate controls, so change one at a time and keep notes about which setting produced the visible difference.
Yes, you can chain Load LoRA nodes, passing the adapted MODEL and CLIP from one node into the next. Start with one compatible adapter first, because multiple files make weight conflicts and attribution much harder to diagnose.