GRANDI-AI Studio

Mode \ Trait	Best for	Supported complexity	Strength	Possible drawback
Sharpen + denoise	Good-quality images that contain noise	Complex	Produces a cleaner image	Creates a smeared look
Standard enhancement	Clean, good-quality images	Complex	Adds fine creative detail	Makes the image look extremely dirty
Deblur	Very blurry images	Simple	Strong deblurring	High-frequency details break down
Powerful repair	Damaged images with obvious defects	Simple	Produces a strong repair effect	Changes the image too much
Fidelity repair	Images with compression damage	Complex	Ensures high similarity while repairing	Produces no change in the image
Smart repair	All image types, especially images whose details are completely destroyed	Complex	Can reconstruct hard-to-read images	Changes the image’s color and lighting
Text poster	Poster images containing substantial text	Complex	Repairs damaged text	Changes non-text regions too much

Model	Scene complexity	Prompt understanding	Image quality & local-edit blending	Recommended use	Common drawback
E1 high-quality general-purpose model	General	Moderate; requires concise, precise prompts	Extremely high. The edited region suffers almost no loss of image quality and, in the vast majority of cases, has no visible seam with its surroundings.	Simple editing needs such as style conversion, and high-quality post-processing	The image does not change at all after editing, or the result does not achieve the effect requested by the prompt.
E2 simple-element specialized model	Simple visual content only	Good; requires concise, precise prompts	Adaptive. The more complex the content in the edited region, the worse the image quality; under normal circumstances, there is no visible seam with the surrounding region.	Editing Chinese characters, removing or replacing low-complexity simple elements, and processing character faces	Image quality becomes worse after editing, and the result still does not achieve the effect requested by the prompt.
N-Banana2 model	General	Extremely high; understands instructions of varying complexity	Moderate. Every edit changes the image in the context area to some extent, making it look slightly different from the surrounding area.	Complex editing needs and tasks that other models cannot complete	Over-invents and makes unauthorized changes to parts of the image.
N-BananaPro model	Its overall characteristics are similar to the N-Banana2 model. In our testing, it showed better understanding of color and style. It performs poorly on annotation-guided image-editing tasks. It also does not support extreme aspect ratios.
G-Image2 model	Its overall characteristics are similar to the N-Banana family. In our testing, it followed prompts more closely and preserved the source image’s content better when converting or transferring styles. However, image quality is slightly lower, and extreme aspect ratios are not supported.

Credits: log in

Unlimited

Choose tool

Set the target scale or total pixel count.
• 1X–4X: fixed scaling
• MP modes: fixed target resolution, such as 8 MP ≈ 4K and 32 MP ≈ 8K
* Choose output resolution based on the image content and quality.
• Wider, more complex scenes require a higher output resolution.
• For poor-quality sources, use several gradual upscale passes instead of immediately targeting a very high resolution.

Specifies the stylistic direction for upscaling the image.
【Realism】The site’s primary style, suited to realistic imagery.
【3D Cartoon】Suited to more cartoon-like 3D animation renders.
【Ink / Watercolor / Oil Painting】Suited to artwork in the corresponding styles.
All of the styles above have been specially optimized and are strongly recommended. If the source image does not fit any of them, try 【Adaptive】.

【Sharpen + Denoise】A versatile mode, particularly suited to images that already contain enough information but look slightly blurry and noisy because of low resolution, such as high-quality real-world photographs.
【Standard Enhancement】Favors adding more detail. It is suited to clean, detail-poor conventional AI images (such as Qwen-Image and Jimeng), but can easily make an image look dirty, so use it with caution.
【Deblur】Designed for images that are especially blurry and not very complex, such as low-resolution web images with a single subject or screenshots from highly compressed online videos. Using it on an image that is not blurry enough may give the result a “burned” look.
【Powerful Repair】Designed for images with extreme structural breakdown. For ordinary structural damage, use Structure Repair below instead; this mode is unnecessary and may reduce image quality.
【Fidelity Repair】Works well on many image types, especially low-quality images, but has limited ability to improve clarity.
【Smart Repair】Intelligently analyzes the image, balancing detail restoration with clarity enhancement, and is suited to a wide range of complex scenes.
【Text Poster】Designed for poster-style images containing a large amount of text. It prioritizes improving text edges, layout hierarchy, and paper or background texture.
* For images with severe information loss (such as damaged old photographs or extremely low-quality night scenes), try entering a prompt such as “repair the image and improve clarity” in Edit mode as a fallback repair approach.

V1.0
The site’s long-standing classic creative-algorithm engine, retained only for compatibility and to reproduce previous processing results. It is generally not recommended for new image-processing tasks.

V1.2
An improved version based on V1.0. It fixes some known issues and adjusts the algorithm pipeline. Compared with V1.0, it improves both overall image quality and similarity to the source.

V2.0
Uses an entirely new creative-algorithm pipeline. Compared with the V1.X series, details are softer, with better consistency across image regions.

Structure repair [×1.2] VIP? 0

Repairs broken image structures, such as the characteristically malformed results sometimes produced by MidJourney.
This is not recommended when the image has no obvious structural damage; otherwise, the image may lose detail and its structures may be simplified.
Enabling Structure Repair doubles the processing time, and the total cost is calculated at 1.2× the normal rate.

Creativity ? 0

Allows the AI to invent details. Higher values produce more texture and detail while reducing similarity. If similarity is critical, set it to −10 to turn this effect off. If the source image is high quality, consider lowering it to around −3 to improve similarity. For ordinary upscaling, it generally should not exceed 2; for drastic transformations, such as converting an illustration into a photograph, it can be raised all the way to 10.

Similarity ? 0

Higher values preserve the source more closely but reduce generated detail. Start around 3 or 4 when fidelity matters; very high values may reduce image quality.

Fractal ? 0

Lower this parameter to a negative value in the following situations. Lowering it too much can reduce image quality, so start at −1.
1. The image contains large areas of solid color, such as a clear blue sky or a dark environment.
2. The image contains relatively little content, but the target output resolution is high.
3. The target output resolution is exceptionally high.
4. The upscaled result shows obvious tiling, or unwanted content appears in large solid-color areas.
If the target output resolution is low, raising this parameter slightly can produce a modest improvement in image quality.

Detail Refinement [×1.1] ? 0

Repairs high-frequency image detail to improve visual texture. Higher values produce a stronger effect, but reduce similarity.
For scene images where consistency of individual elements is not critical, consider setting it directly to 3 for better image quality.
Enabling Detail Refinement increases task processing time, and the total cost is calculated at 1.1× the normal rate.

? 0

Improves the similarity and quality of small elements in an image, such as people or objects in a wide scene, or a face in a full-body portrait.
After enabling it, open the mask editor, use different brush colors to paint the regions that need attention, and enter a prompt for each region.
This feature is inactive when the parameter is 0 or the mask is empty (the parameter slider is gray). Level 1 favors creative detail, level 2 favors similarity, and level 3 preserves extremely high similarity.
If reused parameter data includes mask data, reopen the mask editor and click Restore Mask.
This feature consumes additional credits. When Structure Repair is 0, each colored mask costs 1 credit; when Structure Repair is enabled, each colored mask costs 2 credits.

Assistant prompt ?

AI upscaling requires the AI to “guess” the details in blurry areas. Assistant prompts help the AI understand the image content correctly, preventing misidentification or unwanted details:

• Positive description (what it is): Helps the AI identify blurry objects.
Example: If the source contains a blurry statue that the AI might mistake for a lion, write “The statue in the image is a qilin.”

• Negative description (what you do not want): Prevents the AI from adding unnecessary details.
Examples: “Do not add freckles to the person’s face” or “Keep the wall clean; do not add cracks or stains.”

✕ Editing instructions are not supported (such as “make the lighting more harmonious” or “adjust the color tone”). For editing requests, switch to Image Edit above, then use the editing tools on the task card.

Post processing

▶

HDR ? 5

Enhances the image’s dynamic range, improving contrast and detail in highlights and shadows for a clearer, more open-looking result.

Noise ? 5

Adds a film-grain effect to increase the image’s texture and realism.

Chromatic aberration ? 1

Simulates lens dispersion by creating red/blue or green/purple shifts along edges, increasing the image’s texture and realism.

Depth of field ? 0

Simulates a lens-aperture effect, blurring the background to increase the image’s texture and realism.

Fog ? 0

Adds a hazy atmosphere to the image, simulating atmospheric perspective. Recommended for wide scenes.

+Reference image 1 1

2Add next 2

3Add next 3

4Add next 4

5Add next 5

6Add next 6

7Add next 7

8Add next 8

Selecting a commonly used official preset automatically applies it to the current image-generation panel. If the prompt already contains text, the preset is appended as a separate section; if a complete preset section is already present, switching presets replaces that section. Returning to Custom removes an exactly matching preset section while preserving the text you wrote.

Choose the model for this image generation. GPT-Image2 supports custom ratios from 1:3 to 3:1, while the NanoBanana family uses fixed ratio options.

Multiple-reference-image limits:
NanoBanana2 supports up to 8 object images and 4 character images (style references are not supported).
NanoBananaPro supports up to 6 object images, 5 character images, and 3 style reference images.

Prompt ?

Describe the content, style, and composition you want to generate, along with the features that should be preserved from the reference images. Multiple reference images are used as context in the order of the numbers shown in their upper-left corners.

Image ratio ?

Choose the generated image’s aspect ratio. With Auto, the model decides based on the prompt and reference images. GPT-Image2 accepts positive custom width-to-height ratios with up to two decimal places, from 1:3 to 3:1, and automatically substitutes 3:1 and 1:3 for the 4:1 and 1:4 presets.

8:1 :

Choose how many images to generate in one request.

Use redemption code

Request electronic invoice