LoRA not working? 7 checks, cheapest first

When a character LoRA isn’t working, run 7 checks, cheapest first: the file loaded, the base model matches, the trigger word is exact, the strength is high enough, the generation settings match, the checkpoint is right, and the dataset covers the shot. The first four take minutes, and none of them needs a retrain. Once she shows up reliably, the rest of the consistent AI character setup is prompting.

Why is my LoRA not working?

Almost always one link in a short chain is broken: the file, the model it runs on, the word that calls it, or how strongly it’s applied. Check them in order, cheapest first, and change one thing at a time with the same seed, so you can see which change fixed it.

Here’s the order we suggest, cheapest first.

1. Is the file loaded? Right folder, full file, no errors in the console.
2. Does it match the base model? A LoRA only works on the model it was trained for.
3. Is the trigger word exact? Same spelling, spacing and capitals as the captions.
4. Is the strength high enough? Test at 1.0.
5. Do the generation settings match? Steps, CFG and sampler from the template, not leftovers from another workflow.
6. Is it the right checkpoint? Too early isn’t her yet. Too late repeats itself.
7. Does the dataset cover the shot? A missing framing in training shows up as a wrong face in that framing.

A rough guide to where each symptom usually points:

How do you know the LoRA actually loaded?

Look at three things: the file is in ComfyUI’s models/loras folder, it’s the full file, and the console shows no LoRA warnings when you queue an image. If the LoRA isn’t in the loader’s list, ComfyUI can’t see the file. If it’s in the list and does nothing, read the console.

ComfyUI looks for LoRAs in ComfyUI/models/loras (ComfyUI docs, checked 2 Oct 2026). Upload the .safetensors file there, for example with JupyterLab, then refresh ComfyUI so the list re-reads the folder.

Check the file size. A real LoRA file is measured in megabytes. A file of a few kilobytes is a web page saved under the wrong name. Our Telegram bot checks for exactly this: if LORA_URL opens a download page instead of the file, it refuses and replies that the link didn’t give it a .safetensors file. A share link has to point at the file itself.

Does the LoRA match the base model?

A LoRA is a small set of changes to one specific base model. Load it onto a different model and the changes have nothing to attach to. ComfyUI tells you: it prints lora key not loaded in the console for every part it couldn’t match. Strength, prompts and seeds can’t fix that.

That warning comes straight from ComfyUI’s own LoRA code (source on GitHub, checked 2 Oct 2026). If the console fills with them, you have the wrong model.

A version of the same model can work when the model’s maker or the trainer says it’s compatible. Our current training template trains on the full version of a base model and the generator loads the LoRA on the faster turbo version of that same model, as its README says (lora-training repo). A different model family won’t work. The same goes for our Telegram bot: it runs Krea 2 Turbo, so it needs a LoRA trained on Krea 2 (bot repo). If you want to switch base models, train a new LoRA on the new one from the same dataset. Results depend on the base model, so test before you commit.

Is the trigger word exactly right?

The trigger word has to match the captions character for character. zvx woman, Zvx woman and zvxwoman are three different things to the model. One typo and it draws a generic woman. Put the trigger first in the prompt, followed by her short hair-and-eyes line, then the scene.

Our Telegram bot’s fix for “not her face” starts here: exact trigger word, then strength 1.0. If you’ve lost track of the trigger, open one of the training captions. In our setup every caption starts with it. More on choosing and finding it in our LoRA trigger word guide.

A prompt that describes her face, makeup or body also counts as a trigger problem. Those words pull the base model toward its own idea of a woman. Keep only the trigger and the hair-and-eyes line, like zvx woman, long wavy dark brown hair, middle part, hazel eyes.

Is the LoRA strength high enough?

Strength scales how much the LoRA changes the model. At 0 it’s off, and a low value gives you someone who only resembles her. Our generator template runs 0.9 by default. For a test, set it to 1.0, the top of our range. If she appears at 1.0, settle somewhere between 0.8 and 1.0.

If she doesn’t appear even at 1.0 with the exact trigger, go back to checks 1 and 2. Our templates top out at 1.0. A good rule of thumb: if 1.0 with the exact trigger can’t call her, the cause sits elsewhere in the chain. What each value does, and why our loader has one strength slider instead of two, is in our LoRA strength guide.

Why do trainer samples look fine but ComfyUI images look worse?

The trainer’s samples and your ComfyUI images aren’t made the same way. Samples use the trainer’s own sampling setup, and your generator uses its own steps, CFG, sampler and image size. If those don’t match what the model expects, the face can be right while the image can look soft, noisy or burnt.

The model can differ too. Our current template trains on the full version of the base model, while the generator runs its faster turbo version, so samples and posts don’t even come from the same model file.

Match the settings of the template that goes with your base model. Ours, for the current base model, are:

These are the right settings for our template on our current base model, not for every model. A different base model has different numbers. The common mistake is loading a LoRA into an old workflow built for another model, with that model’s step count and CFG still in place.

What if she looks right in close-ups but not in full body?

That’s a dataset gap, and it’s the one check that needs a retrain. The LoRA learned her face mostly from close-ups, so in full-body shots, where the face is small, it guesses. Tests catch this: if every checkpoint fails the same framing, no checkpoint will fix it.

Aim for about a third close-ups and upper body, a third cowboy and three-quarter shots, a third full body: the mix the course uses for prompts. If full body is thin, add more full-body images of her and train again. The free LoRA dataset planner lays out a shot list with that mix. How many images you need and what they should show is in our guide to LoRA dataset size.

If the face is fine but every image has the same background, outfit or expression, the LoRA works too well on the wrong things. That’s overtraining, covered in our LoRA overfitting guide.

What can go wrong?

• You change three things at once. Strength, trigger and checkpoint together, on a random seed. Something improves and you don’t know what. Fix the seed and change one thing per run.
• The pod is gone before the LoRA is saved. Terminating a RunPod pod deletes everything not on a network volume (RunPod docs, checked 2 Oct 2026). Download the .safetensors files from the trainer’s output folder first.
• Two files, one name. You upload a newer checkpoint under the same file name and test the old one by mistake. Keep the step number in the file name.
• You test only one checkpoint. Early ones are close but not quite her, and the final file isn’t always the best. Video 3 of our course compares the final file, 2500 and 2000. The method is in how to test a LoRA.
• The prompt fights the LoRA. Face shape, makeup or body words in the prompt pull toward a different woman, even with a perfect LoRA. Delete them.