If you've added a LoRA and the character, style, or effect you expected isn't showing up, the cause is usually one of a handful of things: weight too low to register (or too high and taking over), a base model that doesn't match what the LoRA was trained on, a prompt describing something the LoRA is trying to override, or several stacked LoRAs fighting each other.
On PixAI, the trigger word is rarely the problem, since it attaches automatically when you select the LoRA. It can still trip you up if you edit it out, though.
This is a debugging guide, not a theory post. We'll isolate one variable at a time and show real generations so you can see exactly what changes.
First, know what your LoRA is supposed to do
Before touching any settings, figure out what your LoRA is built to do, since the fix depends on it. A character LoRA reproduces a specific character's face, hair, and signature features. A style LoRA shifts the overall art style (linework, shading). Others target a single outfit, pose, or element like cat ears.
These types behave differently, so a setting that's correct for a style LoRA can be wrong for a character LoRA. For a refresher on how models and LoRAs relate before you debug, see the model versus LoRA foundations guide.
Trigger words on PixAI are usually automatic
On a lot of platforms, a missing trigger word is the common failure mode. On PixAI it's less of an issue, since selecting a LoRA auto-appends its default trigger words to your prompt. You set a weight and generate.
That automatic behavior is editable, which is where problems start. Once the LoRA and its trigger attach to your prompt, you can change the trigger text (for example, the haruka v2 style tag that appears after adding that LoRA).
Editing it is useful when a LoRA has multiple triggers for different effects. But if you delete or alter that appended text while cleaning up your prompt, the LoRA can lose the keyword it needs and the effect weakens or disappears. If a LoRA stops working right after a prompt edit, check that its trigger is still there first.
A small number of LoRAs don't carry a default trigger, or need one added manually. For those, check the LoRA's model page for recommended words. A misspelling will keep it from firing. The LoRA trigger words guide goes deeper on activation mechanics.
When the weight is too weak or too strong
Weight is the setting you'll touch most. It controls how strongly the LoRA pushes on the image. Too low and it barely shows. Too high and it takes over, flattening your prompt's details into its own look.
To isolate this, I ran one fixed prompt with the Haruka v2 style LoRA and changed only the weight.
Prompt: 1girl, silver hair in a high ponytail, teal eyes, standing on a rooftop garden at sunset, city skyline behind her, gentle wind, anime style, masterpiece, best quality
Same prompt at LoRA weight 0.2 (left), 0.5 (middle), and 0.8 (right)
0.2: Light touch. The full prompt came through (silver ponytail, teal eyes, rooftop garden, sunset, skyline); the LoRA only softened rendering slightly. Could pass as a clean base-model image if you didn't know a LoRA was applied.
0.5: Much more visible style. Softer face, more expressive eyes, more vibrant lighting, prompt still fully intact. Best balance of style and accuracy for this LoRA/scene.
0.8: LoRA became the dominant influence. Stronger anime look, larger eyes, cleaner linework, prompt still held, but the model added a ribbon and a collared shirt with a bow that were never in the prompt. That's the tell: when a LoRA starts adding things you didn't ask for, you've pushed weight past the point where your prompt still leads.
There's no universal "correct" weight. It depends on the LoRA, the base model, and how much of its look you want, so test rather than guess. Move in small steps so you can see exactly where the effect appears and where it starts overriding you. More detail in the LoRA weight settings guide.
When the base model doesn't match
This is one of the most common "broken" LoRAs, and it has nothing to do with your prompt. A LoRA is trained on a specific base model architecture and only behaves properly on that same type. Load it on a mismatched base and you get almost no effect, or an unstable, distorted one.
On PixAI, the split that matters most is DiT models (like Tsubaki.2) vs. SDXL models (like Haruka v2 and Hoshino v2). A LoRA trained for an SDXL base will not work correctly on a DiT model, and vice versa.
PixAI shows the architecture on most model and LoRA cards. Before blaming your prompt, confirm the LoRA and base model are the same family. If a LoRA works on one model and does nothing on another, compatibility is the first thing to rule out.
Whether choosing a LoRA or training one, match it to the base you generate with. The train LoRA on PixAI guide covers that, and DiT models like Tsubaki.2 have their own requirements in the DiT LoRA training guide.
When your prompt fights the LoRA
A LoRA and your prompt can pull in different directions: a character LoRA wanting one hairstyle while your prompt asks for another, or a style LoRA pushing one look against your request.
To test this, I used the Cute Anime Style v3.0 LoRA at weight 0.7 and deliberately specified a different character look.
Prompt: masterpiece, best quality, 1girl, [LoRA tag here], short blonde pixie cut, plain white t-shirt, casual jeans, standing in a modern kitchen, bright daylight, anime style
The prompt won cleanly. Pixie cut, white t-shirt, jeans, and modern kitchen all came through as written. The LoRA's influence showed up in the rendering (expressive eyes, soft features, clean shading) without overriding content. Since this is a style LoRA, it shapes the look rather than the character, so specifying hair and clothing didn't fight it.
Then I removed the character details and let the LoRA lead:
Prompt: masterpiece, best quality, 1girl, [same LoRA], standing in a modern kitchen, bright daylight, anime style
Without a hairstyle or outfit specified, the LoRA filled those gaps itself: dark bob, large amber eyes, casual clothing, none of it in the prompt. Kitchen and daylight stayed intact. This shows the LoRA carries its own character bias that only appears when the prompt leaves room for it.
Practical approach: test the LoRA alone with a minimal prompt first to learn its default behavior, then add details gradually. If results destabilize, remove descriptions one at a time until you find what's fighting the LoRA. With a strong character LoRA this conflict is much sharper, since it has a fixed look baked in.
Character LoRA vs. style LoRA: different failure modes
Character LoRA: built to reproduce a specific design. Wants a stable base prompt plus its trigger, and pushes back hard when you contradict its built-in traits.
Style LoRA: built to shape overall look. Usually needs a lower weight to avoid flattening character detail, and lets your prompt handle content.
When a LoRA misbehaves, ask what type it is first. A character LoRA ignoring your requested outfit is often working as designed. A style LoRA erasing your character's features is probably set too high.
When you've stacked too many LoRAs
PixAI allows up to 5 stacked LoRAs, which is useful but also a common source of muddy, unstable results. Each one pushes the image in its own direction, and several pulling at once interfere with each other.
Two or three is a practical ceiling for most work. Past that, you're usually compensating for the wrong base model rather than filling a real gap. To debug a stack: load one LoRA at a time, confirm each does what you expect, then combine and lower weights so no single one dominates. For multi-character scenes specifically, see the multi-character LoRA guide.
The debug checklist
Repeatable process, one variable at a time:
- Start with a simple prompt and the LoRA alone, at a moderate weight, to see baseline behavior.
- Confirm the trigger is still attached to your prompt (re-add it if edited out).
- Sweep the weight in small steps, low to high, and watch where the effect appears vs. where it overrides your prompt.
- Try the same LoRA on a compatible base model, and check architecture match if it does nothing.
- Simplify your prompt if the LoRA and description seem to be fighting, then add details back gradually.
- If stacking, add one LoRA at a time to isolate the problem source.
- Compare results side by side so you can see which change actually fixed it.
This loop is fast on PixAI since you adjust weight, swap models, and edit prompts online, then regenerate in seconds. If you're still learning where these controls live, the how to use PixAI guide walks through the generation panel.
Quick reference checklist
- Do I know if this is a character, style, outfit, or element LoRA?
- Is the trigger still attached to my prompt, and spelled correctly if typed manually?
- Is the weight high enough to show, but not so high it overrides the prompt?
- Does the LoRA's architecture match my base model (DiT with DiT, SDXL with SDXL)?
- Is my prompt asking for something the LoRA is trying to override?
- Have I stacked too many LoRAs, and can I test them one at a time?
- Am I changing one variable at a time so I can tell what actually fixed it?
Wrapping up
A LoRA that looks broken usually isn't a broken file. It's a setting that needs adjusting, a base model mismatch, or a prompt pulling against it. Once you can tell those apart, most problems take only a few test generations to sort out.
Have a LoRA that's not cooperating? Start testing on PixAI for free and work through it one setting at a time.
Top comments (0)