> ## Content Index
> Fetch the complete content index at: https://montagea.com/blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# How to Upscale Portrait Reference Images
- URL: https://montagea.com/blog/upscale-portrait-reference-images/
- Published: 2026-11-18T06:08:02.000Z
- Updated: 2026-11-18T06:08:02.000Z
- Description: Upscaling portrait reference images for animation and video production requires a careful balance. Learn how to configure upscaling parameters, manage the hallucination budget, and inspect outputs to maintain character consistency.
- Author: Tim
- Tags: ai creative tools, video production, ai filmmaking, comparisons

When upscaling portrait reference images using AI upscaling tools while preserving character identity, managing the model's "hallucination budget" is a key consideration. By pairing a descriptive text prompt with a low-to-moderate denoising strength, you can instruct the AI to inject realistic textures—such as skin pores and hair strands—without altering the subject's core facial geometry. This approach is designed to help keep upscaled portraits consistent for downstream animation pipelines, though it will not recover historical camera data that was never captured.

## Understanding the Hallucination Budget in Portrait Upscaling

AI upscaling tools are often used by visual creators. However, to use them effectively, you must understand a fundamental truth of AI-guided upscaling: the model does not recover lost historical camera data. If a source image is a blurry, low-resolution portrait, the original camera sensor data for individual eyelashes or skin pores does not exist. Instead, the upscaler generates entirely new, highly plausible textures based on patterns it learned during training.

This generation process operates on what we call a "hallucination budget." This budget represents the degree to which the model is allowed to invent new visual details. When upscaling landscapes or abstract art, a high hallucination budget is acceptable because newly generated leaves or stone textures only add to the realism. With human portraits, however, a high hallucination budget can be disastrous. If the model invents a slightly different eyelid fold, shifts a beauty mark, or alters the reflection in the iris, the subject's unique identity is lost. Managing this budget is the key to successful portrait upscaling.

## Configuring Parameters for Character Identity Preservation

Preserving identity typically requires fine-tuning the inputs sent to the upscaling API. The two most critical levers at your disposal are the text prompt and the denoising strength.

![A technical diagram illustrating how denoising strength and prompt conditioning interact to preserve facial identity during AI upscaling.](https://montagea.com/blog/content/images/2026/11/montagea_sec_83cd698a.png)

The prompt is not just for generating new images; in upscaling, it acts as a guide. By describing your character's key features in the prompt (e.g., "a close-up portrait of a 30-year-old woman with green eyes, subtle freckles, and dark brown hair"), you prevent the model from guessing incorrectly. Before running these high-resolution passes in your local build or cloud environment, ensure your pipeline is secure by following our guide on [How to Configure API Keys for Creative Models](https://montagea.com/blog/creative-model-api-key-configuration/).

The denoising strength parameter controls the balance between the original image structure and the upscaler's creative freedom. A denoising strength of 0.0 leaves the image unchanged, while 1.0 completely replaces it. For portraits, the sweet spot lies in a narrow band. Choosing the right upscaling target is also a balance of resource costs. For a broader look at balancing processing times against output quality, see our comparison of [Standard vs. Pro Image Model Tiers: Comparing Speed and Fidelity in Production](https://montagea.com/blog/nano-banana-2-vs-nano-banana-pro/).

When configuring these settings, creators generally balance conservative adjustments against more creative reconstructions depending on the source quality:

- **Conservative Settings:** Best for high-quality source portraits where you only need to add clean micro-textures (pores, hair strands) without altering facial structure.
- **Balanced Settings:** Best for standard generated portraits or mildly compressed reference shots that require clean-up and moderate detail enhancement.
- **Creative Settings:** Best for heavily pixelated or low-resolution sources where the model must reconstruct missing facial geometry and textures from scratch.

When implementing these configurations in your pipeline, visualize the denoising strength as a dial. Setting the dial too low results in a blurry upscale that merely interpolates pixels. Setting it too high pushes the model outside the "safe zone," causing it to invent entirely new facial features that deviate from the original character reference.

## Inspecting Invented Detail: A Portrait Quality Checklist

Once the upscaler delivers the rendered output, you must systematically inspect the results. AI upscalers often introduce subtle artifacts that can ruin downstream animations, particularly around high-risk facial zones.

Use the following checklist to evaluate your upscaled portrait before moving it into your creative workflow:

- **Eyes and Pupils:** Check for symmetry. The upscaler should not alter the natural shape of the iris or introduce mismatched pupil sizes. Check if catchlights (the reflections in the eyes) appear consistent in both eyes.
- **Teeth and Mouth:** If the subject's mouth is open, inspect the teeth. AI models frequently merge teeth together or create unnatural, jagged alignments when upscaling low-resolution mouth areas.
- **Hair-to-Skin Transitions:** Look closely at the hairline and eyebrows. The transition should feature distinct, fine hair strands rather than a solid block of color or a blurry smudge.
- **Skin Texture (The "Plastic" vs. "Grit" Check):** Check that the skin does not look overly smooth and plastic-like, which can occur when denoising is too low, or unnaturally gritty and hyper-detailed, which can happen when denoising is too high.

At Montagea, we look at upscaling as the foundational step of a broader creative pipeline. Getting the portrait reference right at this stage prevents costly errors when you begin animating the character. Once your portrait is cleanly upscaled, you can proceed with asset preparation using our walkthrough on [How to Prepare Assets and Inspect Lip Sync in Portrait Animation Models](https://montagea.com/blog/omnihuman-asset-prep-lip-sync-guide/).

Why do upscaling models sometimes change the eye color of a character?

If the denoising strength is set too high or the text prompt does not explicitly specify the eye color, the model may hallucinate details based on its training data. To prevent this, lower your denoising strength and write a descriptive prompt that explicitly states the character's eye color.

Can upscaling models be used on heavily compressed or pixelated AI-generated portraits?

Yes, but you may need to increase the denoising strength so the model has enough creative freedom to reconstruct the heavily damaged areas. Keep in mind that higher denoising strengths will result in more hallucinated details, which may alter the original character's identity.

How does prompt strength affect the fine details like jewelry or clothing textures in a portrait?

Higher prompt weights (guidance scales) force the model to adhere strictly to your text descriptions. If your prompt describes "intricate silver earrings" or "knit wool sweater textures," a higher guidance scale will make these details more defined, while a lower scale allows the upscaler to rely more on the original pixel patterns.