# Improving output quality (/docs/troubleshooting/quality)



A generation that *finishes* but looks wrong is almost never a bug; it is the model, the variant, or the prompt. Work through the matching symptom below.

## Everything is blurry / noisy / "unfinished" [#everything-is-blurry--noisy--unfinished]

Most often this is the **model variant**. GGUF variants are compressed to fit smaller GPUs and trade away quality; heavily quantized ones can look genuinely bad.

**Fix:** open the preset card and switch to the **fp8** (or higher) variant if your VRAM allows it. If you must stay on GGUF, raise the step count in Advanced Controls and prefer higher resolution tiers; 480p upscaled will never look like native 720p.

## Faces or bodies are deformed (extra limbs, warped hands, horror stuff) [#faces-or-bodies-are-deformed-extra-limbs-warped-hands-horror-stuff]

**Fix:** put a proper recipe in the negative prompt ("what to avoid") field instead of hoping the model behaves. A good starting set:

```
deformed, extra limbs, extra fingers, bad anatomy, malformed hands,
fused fingers, mutated, disfigured, blurry, low quality
```

Keep the positive prompt focused. See [Negative prompts](/docs/generating/negative-prompts) for how these are applied per model.

## Output looks cartoonish / plastic / "Disney" when you wanted photorealism [#output-looks-cartoonish--plastic--disney-when-you-wanted-photorealism]

**Fix:**

* Don't stack quality keywords ("hyperrealistic, 8k, ultra detailed, unreal engine..."). On modern models these often push output *toward* CGI looks rather than realism.
* Add cartoon-blockers to the negative prompt: `cartoon, anime, 3d render, cgi, plastic skin, doll`.
* Check you're on a photorealistic preset in the first place; each preset's card says what it is tuned for.

## The model ignores parts of your prompt [#the-model-ignores-parts-of-your-prompt]

Prompt adherence differs per model family. For video, **Wan follows prompt instructions noticeably better than LTX**; LTX wins on other axes (speed, audio). If precise staging matters, try the same prompt on a Wan preset.

**Fix:**

* Simplify. Very long prompts dilute; models latch onto the first, concrete instructions.
* Use positional language for camera and movement ("walks from frame-left to frame-right" rather than "walks to the right").
* For a specific composition: generate a still image first, iterate until the framing is right, then feed it to an image-to-video preset. This is the single most reliable quality workflow.

## Faces drift / person morphs during a video [#faces-drift--person-morphs-during-a-video]

Identity drift over a clip's duration is a known limitation of current i2v models, not a settings problem.

**Fix:** keep clips short (5s), keep the subject prominent in the start image, and prefer generating several short clips over one long one. For a consistent character across many generations, use the character pipeline: base image -> image-edit preset for poses -> i2v ([Characters](/docs/characters)).

## Text in the video is gibberish [#text-in-the-video-is-gibberish]

Legible text is hit-or-miss for video models. Re-rolling with a new seed sometimes fixes it; Wan renders text more reliably than LTX. Otherwise: add the text in an editor afterward.

## Audio is robotic / distorted [#audio-is-robotic--distorted]

LTX's generated audio is a model limitation. For voice that matters, generate the video silent and use an external TTS plus the Lip-Sync preset.

## Still bad? [#still-bad]

* Check the preset for an **update** in the Preset Store; several quality issues have been fixed by pack updates.
* Compare against the preset's preview examples: if your output is far below them at similar settings, export logs (**Settings -> Troubleshooting -> Export Logs**), then post in the Discord issues forum (see [Getting help](/docs/troubleshooting/getting-help)).