GRANDI-AI Tutorial Guide

Introduction

When first using this site, many users may find that some images produce excellent results while others fall short of expectations or vary in quality, unlike the best examples shown on the site. This is not necessarily an operating error or system defect. Image upscaling (super-resolution reconstruction) is inherently complex and highly variable: results depend closely on the source and quality of the input image as well as the user's specific goals. There is therefore no universal one-click solution for every scenario. Even with today's advanced AI, it remains difficult to quantify precisely an input image's quality, type of damage, scene complexity, and the user's desired balance of creativity and resemblance. To get the most from this site, users need a basic understanding of the concepts and techniques described below.

Before using the site for the first time, we strongly recommend reading this guide carefully and watching the official case tutorials.

Feature Overview

This site provides the following main features:

• Core Feature 1 — Image Upscaling: Repairs and enlarges images while adding rich detail to your work.
• Core Feature 2 — Image Generation: Generates new images from prompts and reference images.
• Core Feature 3 — Image Editing: Provides several editing modes for global or local image changes.
• Core Feature 4 — Advanced Task-Card Tools: Provides follow-up tools including local upscaling, cutout, mask compositing, color matching, and image annotation.
• Supporting Features — Task Management and Toolbox: Provides practical tools for workspace management, splitting and merging images, and comparing results.

Note: Uploading images and saving results from browser-only tools are free. Tasks that use backend or GPU compute are billed at the price shown on the page (click here to view the billing rules).

Case Tutorials

These case tutorials demonstrate complete workflows: first diagnose the source image, then combine image editing, image upscaling, local processing, or multiple iterations as needed. Click a thumbnail to open the full comparison image; click “View tutorial” to open the corresponding video.

Image Upscaling

Input Image

• The system supports three upload methods: click, drag and drop, or paste.
• Tip: In Photoshop, stamp visible layers, select all, and copy; alternatively, click the Copy button at the upper-right of a task card on this site. You can then paste directly into the input component for fast image transfer.

Preset Parameters

• Click to enter the preset parameter interface. Click the "Apply Parameters" button to quickly load preset parameters configured for common scenarios.
• Click the thumbnail directly to view the case renderings processed by this parameter. Zooming and dragging the canvas are supported during viewing.
• For complex images that cannot be solved by preset parameters alone, it is recommended to learn the complete processing process in conjunction with the "Case Tutorial" above.

Important: Presets are only a baseline starting point. Upscaling requirements vary too widely for any preset to solve every problem in one click. Build experience and learn what each parameter means, so that you can quickly choose suitable settings for a new image and accurately diagnose and adjust them when a result is unsatisfactory.

Scale Factor

On the surface, this parameter only determines the size of the output image, but in fact it is extremely critical and closely related to other parameters and specific scenarios.

• Image clarity ≠ resolution: Resolution only limits the upper limit of definition. For example, if a high-definition original image taken by a professional photographer is compressed to 720P, its clarity may still be higher than the 2K/4K image produced directly by some AI-generated image models. Merely forcing a resolution increase in the software will not increase the actual clarity of the picture. Experience shows that the actual clarity of 4K images (resolution close to 5K) produced by the current mainstream excellent AI imaging models (such as Nano Banana, Ji Meng, etc.) may only be equivalent to high-quality real-shot or 3D-rendered 1080P images, and the details may be even scarcer. Therefore, even if you do not need extremely high resolution (for example, making an AI short film only requires a 1080P single frame), using GRANDI-AI for upscaling processing can still significantly improve the picture quality.
• Complex scenes require higher resolution: At the same resolution, AI-generated close-ups of faces are usually not prone to collapse. However, when generating grand scenes containing hundreds or thousands of people (such as battlefields and bustling streets), the faces of distant characters often collapse because they occupy very few pixels. The upscaling function of this site supports up to 32MP (8K) output, but for extremely complex scenes, this resolution may still be insufficient.
  Advanced Upscaling Tip 1: Use Image Split & Merge in the toolbox to divide the image into tiles, upscale each tile to 32 MP, and then merge them.
• Resolution affects output resemblance: Under the premise that other parameters remain unchanged, the higher the output resolution, the smaller the picture changes (the creativity is reduced, the similarity is improved); conversely, the lower the output resolution, the more drastic the picture redrawing.
• When the output resolution is high, the "similarity" parameter can be appropriately reduced and the "creativity" parameter can be increased to obtain richer details.
• When a strong change in the picture is required (such as converting a hand-painted image into a realistic style, or using strong changes to cover the poor original picture quality), a lower output resolution must be used.
• Advanced Upscaling Tip 2: For extremely poor sources, do not try to reach very high quality in a single pass. Use a staged, multi-pass upscaling strategy: repair image quality first, then add detail.

Style Preference

Realism is the main style of this site, and GRANDI-AI is very good at adding realistic details to images. At the same time, in order to meet the diverse artistic needs of users, the system provides basic support for 3D cartoons (common in domestic 3D animation), ink, watercolor, oil painting and other styles. For other styles that have not been specially adapted (such as two-dimensional animation, etc.), although it can also be processed in the "Adaptive-Low Creative" mode, the effect is not absolutely guaranteed.

Optimization Mode

Includes two upscale-only modes, four upscaling-and-restoration modes, and one Text Poster mode. Click ⍰ after the optimization-mode setting to open a comparison table of all modes.

• Sharpen & Denoise: Pure upscaling mode (no repair effect). Suitable for images with good image quality that do not need to be repaired but contain noise (such as 3D images that are not rendered cleanly, movie screenshots, photos, realistic storyboards generated by N-Banana). The "supports complex pictures" mentioned in the comparison table means that it has no limit on the complexity of the pictures, rather than "can only handle complex pictures". Using this mode incorrectly with a clean, noise-free image may result in smearing or loss of detail.
• Standard Enhancement: Pure upscaling mode (no repair effect). Suitable for images with good image quality and clean images but lack of details (commonly found in images generated by AI models such as Jimeng, Z-image, QWen-image, Flux, etc. Non-AI images that meet this characteristic can also be used). If used for images that contain noise or are not pure enough, it may cause dirty spots on the screen.
• Deblur: Upscaling + restoration mode. Suitable for images with simple and extremely blurry content. This mode is not good at handling overly complex scenes (the more complex the content, the easier it is for details to collapse), but it has a very strong deblurring effect on simple pictures.
• Powerful Repair: Upscaling + restoration mode. It is suitable for images with simple picture content and certain image quality collapse and defects (such as limited pixel collapse caused by local cropping and enlargement). It is also not recommended for complex scenarios. If the image does not meet this characteristic but is used forcefully, it may cause excessive changes in the content of the screen.
• Faithful Repair: upscaling + restoration mode. Suitable for processing images whose quality is damaged due to network compression or resolution reduction (focusing on information recovery rather than redrawing). If there is no underlying information that can be restored in the original image, the image will remain almost unchanged after using this mode incorrectly.
• Smart Repair: upscaling + restoration mode. It has certain repair effects on various types of images, and is the best choice for images whose details are completely damaged, difficult to identify, and require complete redrawing. The advantage is that it supports pictures of any complexity; the disadvantage is that the color of the repaired picture may shift slightly.
• Text Poster: Best suited to images containing a large amount of text. It can repair blurred text and incorrect character strokes.

Creative Engine

Determines which creative algorithm version is used for image upscaling.

• 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

is used to repair the collapsed structure in the picture (such as the logical error image generated by MidJourney, please refer to the preset case for specific characteristics). If there is no obvious structural collapse in the picture, it is not recommended to turn it on, otherwise details will be lost and the structure will be simplified. It is recommended to start with the value 2 when testing.

*Note: This function consumes a lot of computing power and the processing time may be doubled, so it is only open to VIP users, and the total task cost is calculated as 1.2 times.

Creativity

The core parameters of this site’s AI upscaling function are used to inject creative details into images. The higher the value, the greater the picture change, the richer the texture and details, and the similarity will be reduced accordingly. The parameter range is [-10, 10].

• [-10]: Maximizes resemblance by disabling the creative module during AI upscaling. The Resemblance and Fractal parameters below are then disabled and grayed out automatically.
• [-4 to -3]: Suitable for higher-quality original images to further improve similarity.
• [0 to 2]: Standard upscaling that adds detail steadily; values above 2 are generally unnecessary.
• [10]: Suitable for scenes that require drastic transformation or a new overlay and redrawing of the picture (such as converting from illustrations to real photos).

Resemblance

controls the parameters of the creative module to retain the characteristics of the original image during the enlargement process. The higher the value, the closer the enlarged image will be to the original image, but the added detail will be reduced. If you need a higher degree of similarity, it is recommended to start with 3 or 4 . Values ​​that are too high will gradually affect the final image quality.

* Reminder again: When the output resolution is high, the similarity parameters can be appropriately reduced and the creative parameters can be increased to obtain better details.

Fractal

The AI ​​creative module adds details by focusing on local blocks in the picture, and the "fractal" parameter controls the division size of these blocks. The lower the value, the fewer the divided blocks and the larger the area of ​​a single block. Generally, keep the default value 0 . When encountering the following situations, it is recommended to adjust this parameter to a negative value (it is recommended to try from -1 , setting it too low will affect the image quality):

• There is a large area of ​​solid color in the picture (such as a pure blue sky, dark environment).
• There are fewer screen elements, but the target resolution is higher.
• The target resolution of generated images is extremely high.
• The enlarged result shows obvious traces of block splicing, or AI illusion (such as mountains growing out of the sky) and dirty spots in solid color areas (you need to rule out whether the wrong optimization mode was selected first).

*Supplement: If the target resolution is extremely low, appropriately increasing this parameter may help improve the image quality.

Detail Refinement

Repairs high-frequency image details and improves visual texture. The range is 0–3; 0 disables the feature. Higher values apply stronger detail refinement but reduce similarity to the source. For large scenes where exact local-detail fidelity is not important, you can set it directly to 3 for more complete detail refinement and better overall image quality.

Levels 1–3 all multiply both credits and estimated processing time by 1.1.

Local Protection (Reset in Version 2.2)

This is a very powerful but easily overlooked feature. It can significantly improve the quality and similarity of small-sized elements with high-frequency details in the picture (such as objects/people in large scenes, faces in full-body photos, etc.).

• How to use: After enabling, you need to enter the mask editor, use brushes of different colors to paint the areas that need to be protected (up to 6 are supported), and enter a short prompt. The same color mask only supports a single continuous area. If there are multiple scattered areas of the same color, the system will only recognize the largest one.
• Parameter behavior: When the value is 0 or the mask is not drawn, this function is not activated (the slider is grayed out).1 prefers to increase creative details;2 prefers to maintain similarity;3 forces to maintain extremely high similarity. For example, when producing a live-action AI series, you can use this feature to lock the facial consistency of the protagonist and only add creative details to other areas.
• Parameter reuse: Click to enlarge the parameter reuse function in the upper right corner of the task card to synchronize historical records to the left panel. Since the mask is usually highly bound to the original image, for security reasons, after reusing the data containing the mask, this parameter is displayed as enabled but is grayed out and inactive. You need to enter the editing interface and click "Restore Mask" to reactivate.
• Billing: This function requires additional computing power. By default, each color mask consumes 1 point; if "Structure Repair" is also turned on, each color mask consumes 2 points.

Auxiliary Instructions

Normally, leave this field empty. The AI ​​model will automatically identify the screen content and assign appropriate details. Only under the following specific circumstances, you need to enter short instructions to assist AI recognition (therefore called "auxiliary instructions" instead of "prompts" for generating images):

• Correct recognition errors: If the elements of the original image are blurred, the AI ​​may misjudge (such as mistaking the statue for a lion). You can enter "The statue in the picture is a unicorn" for guidance and correction.
• Prevent unwanted details: Prevent AI from overplaying through negative descriptions, such as "No freckles on characters' faces", "Keep the walls clean without cracks and stains".
• Correct attribute drift: Prevent sudden changes in character characteristics (such as Chinese people becoming foreigners) or confusion of era characteristics (modern items appear in ancient scenes).

* Strongly recommended: Do not enter long picture descriptions here, which have no practical meaning; do not enter editing instructions (such as "Adjust light and shadow", "Change tone", etc.). For editing tasks, use Image Editing.

Post-processing

• HDR: Automatically perform color mapping to eliminate the gray look and make the picture more transparent.
• Noise: Add slight film grain to improve the texture and realism of the image.
• Chromatic aberration: simulates the lens dispersion effect, producing red-blue/green-purple shifts at the edges to increase photography realism.
• Depth of field: Zooming in may cause the original blurry distant view to become clearer. This function can simulate the lens aperture to blur the background and restore the depth of field relationship.
• Fog: simulates the atmospheric perspective effect, adding a hazy atmosphere to the distant view, and is recommended for large scenes.

Random Seed

controls the randomness of screen generation. In the default (unlocked) state, a new seed is used every time a task is submitted; click the lock icon to fix the current seed, making it easy to control variables and adjust other parameters to compare effects.

Image Generation

Image Generation is separate from Image Editing and is intended for creating a new image from a prompt or reference image. It does not modify the source image in place. Use Image Editing when you need to preserve the source structure and change only part of it; use Image Generation when you want a new composition, character design, or concept.

Reference Images and Prompts

• Reference images: supports up to 8 pictures. Multiple reference images will enter the model context in numerical order in the upper left corner. You can directly reference objects such as "Figure 1" and "Figure 2" by entering @ in the prompt.
• Quick image transfer: The copy image button in the upper right corner of the task card can be dragged directly to the image upload area; the reference image thumbnail in the image generation history card can also be dragged to the image upload area or reference image area for reuse. This is more efficient than re-downloading and uploading when you need to continuously iterate the same set of materials.
• Prompt: The content, style, and composition of the target image should be described, as well as which features of the reference image need to be inherited. Don’t just write “generate a picture from a reference picture”, otherwise it will be difficult for the model to tell whether you want to preserve characters, materials, colors, or composition.
• Prompt preset: suitable for quickly calling common generation tasks (such as character disassembly, world view storyboarding, etc.). The default will automatically write or replace the corresponding complete prompt segment. Switching back to "Custom" will remove the complete matching default content and retain your own added text.

Model, Aspect Ratio, and Quantity

• Generation model: Different models have differences in image quality, semantic understanding, reference image inheritance capabilities and available proportions.
• Aspect ratio: When "Auto" is selected, the model will make its own judgment based on the prompts and reference images(Warning: G-Image2 model can only generate 160w pixel SD images when Auto is selected!); when a fixed ratio is selected, the system will try to generate the image at the specified ratio. Some models do not support extreme landscape or portrait ratios, and the page will automatically be constrained or replaced with available ratios.
• Number of images: Generating multiple images at one time is suitable for project divergence, but it will also increase the point consumption. If the goal is very clear, it is recommended to generate a small number of results first, and then continue iterating after confirming the direction.
• Random seed: Similar to image upscale, each generation will change when unlocked; after locking, the prompts or reference images can be adjusted under the premise of fixed randomness to facilitate comparison of variables.

Image Editing

This function includes four modes: removal, generation, editing, and optimization (of which "edit" and "optimize" are more widely used in actual scenarios).

Core Parameter: Context Area

Adjusting this parameter will affect the size of the surrounding area that the AI ​​refers to when processing tasks. Enabling the global button will use the entire image as a reference. A real-time preview box will appear when sliding the slider: green means that the current resolution is within the safe processing range of AI; red means that the image quality in this area will be degraded after processing.

• If the parameters are too small: Although the image quality is absolutely safe, the AI ​​lacks sufficient contextual information, which may cause the generated part to be incoherent with the surrounding environment.
• If the parameters are too large: the preview box turns red and the mask area accounts for a very small proportion, which may cause the image quality of the generated area to be damaged.
• Recommendation: Control it to the minimum range where the AI ​​can just cover enough context, and the image quality will not suffer a serious decline even if it is slightly reddish.

Removal Mode

is designed to eliminate screen content. Usually works well without a cue word. If the content filled by AI after removal does not meet the needs, you can enter prompts for directional guidance.

Generation Mode

adds new elements to the scene. Supports text prompts and reference picture functions. When using a reference image, AI generates variations of the reference image instead of copying it. It is recommended that the background of the reference image be as clean as possible to avoid interference.

Editing Mode

uses natural language instructions to modify the entire or partial picture. The function is extremely powerful, but the prompt format has high requirements (if the format is not standardized, it will not be effective).

• Common instruction patterns: Add... to/to the scene; enhance/improve/adjust...; remove... in the screen; modify... in the screen to/become...; make the screen become....
• This mode supports precise migration of elements or styles of reference images. "Figure 1" and "Figure 2" must be clearly marked in the instruction, for example: "Change the style of Figure 1 to the same as Figure 2."

Optimization Mode

Repairs partially collapsed areas in the screen. The higher the optimization amplitude, the greater the picture change; turning on "Solid Shape" can maintain a certain structural similarity under high optimization amplitude.

Chinese Text Generation/Editing Mode

The "Character" button located in the lower right corner of the text box in generation and editing mode. When turned on, it supports redrawing of partial Chinese characters. It is recommended to use the "Edit Chinese Characters" mode, which can better retain the style and color of the original text; while the "Generate Chinese Characters" mode will ignore the original style and generate a new one.

Editing Model

Edit mode supports the selection of different models. Click ⍰after editing model parameters to open the comparison table to view the model characteristics. Model selection should not only look at "whether it is strong or not", but should look at the task type:

• E1 High-Quality General Model: The image quality and local integration are good, suitable for simple style conversion, local optimization, post-processing and other tasks. The disadvantage is that the semantic response is relatively conservative, and complex instructions may not change significantly.
• E2 Simple-Element Model: prompt response is more direct, suitable for editing Chinese characters, replacing simple elements, processing character faces and other low-complexity content. If used in complex areas, image quality may deteriorate.
• N-Banana 2 Model: It has stronger semantic understanding ability and is suitable for complex editing requirements that are difficult to complete with other models. The disadvantage is that each edit may affect the context area, and the user needs to check the degree of picture integration.

* Tip: Chinese character editing will automatically use a model more suitable for text tasks; when switching third-party API models, please pay attention to the privacy and price change tips on the page.

Editing Presets

The preset parameters in the editing panel are used to quickly load prompts, models and context configurations for common editing tasks. After selecting a preset, the system will append or replace the corresponding preset segment according to the current prompt content; switching back to "Custom" will remove the complete matching preset segment and retain the content you entered.

• Edit from annotation: It is suitable to use "image annotation" to draw the modification intention on the screen, and then let AI perform editing based on the annotation. The clearer the labeling, the shorter the prompt.
• Blend collage lighting: It is suitable for integrating collage, texture or synthetic materials into the original picture environment to make the light, shadow, tone and edge relationship more natural.
• Refine character face: used to repair the character's face, improve facial quality, and try to maintain the original character's characteristics. If the face is small or has been severely damaged, you can try multiple times with more suitable contextual areas.

Prompt Enhancement

is located in the lower right corner of the text box (activated after entering the prompt). It is highly recommended for novices to use. This function will automatically optimize and generate high-quality prompts according to the mode:

• Generation mode: combined with the background, simple vocabulary will be expanded into a detailed and suitable description. For example, input "a mouse", and after enhancement it becomes "a furry vole, hiding in the dry grass, with fluffy and earthy hair, small and black eyes, casting a light shadow on the grass blades."
• Editing mode: Convert fuzzy instructions into standardized standard instructions. For example, enter "Replace", which is enhanced to "Refer to the style of Figure 2, replace the simple-drawn smiling flower in Figure 1 with the brightly colored and more childlike smiling flower pattern in Figure 2, maintaining the original circular composition and texture of the metal material."
• Chinese character mode: Comprehensive intention and base map reference, automatically typeset and specify fonts. For example, input "What is written on the screen", and it will be enhanced to standard multi-line text typesetting and "Font: brush calligraphy" and other required formats.

Advanced Task-Card Tools

After the task is completed, a set of follow-up processing tools will appear in the upper right corner of the card. They are not simply view buttons, but secondary creation entrances designed around "continue retouching, local enhancement, result synthesis, and intent annotation."

Local Upscaling

Local upscaling is designed for small areas within a large image, such as distant faces, small props, local text, or cropped details with too few pixels. It is not a replacement for upscaling the entire image; use Image Upscaling when the whole image needs improvement.

• Selection method: Rectangle is suitable for clear-bounded areas such as buildings, products, and partial blocks of pictures; circles/brushes are suitable for irregular areas such as faces, animals, and local objects. If a mask is already available, you can use "Reference Mask" to reduce repeated drawing.
• Composite back onto source: Synthesize the local upscaling result back to the original image, which is suitable for repairing a small area to directly obtain the complete picture.
• Output local result only: Only output local area results, suitable for separate inspection, downloading or continued use as material.

* Recommendation: The physical resolution of the selected part should not be too low. If the original selection itself has too little pixel information, even if it is enlarged and then pasted back to the original image, the improvement in clarity will be significantly limited.

Image Cutout

Image cutout is used to extract characters, objects or other specified content from the result image of the current task card. By default, a transparent base map is output, and a mask map or a result with a specified solid color background can also be output.

• Automatic cutout: Suitable for pictures with clear subjects. When no selection is drawn, the entire image will be processed; after selecting an area on the canvas, only the image within the selection will be sent to the cutout task. After the task is completed, the card will paste the result back to the original selection position for sliding comparison, but the downloaded result is still the partial result map actually output by the cutout task.
• Cutout strength: range is 0–5. After lowering this value, the cutout task will be more conservative and more content can be retained in the cutout results.
• Semantic cutout: Suitable for scenes with many subjects, inaccurate automatic recognition, or where only specific content needs to be extracted. You can enter prompts or add mark points directly on the canvas; both require at least one effective guide to submit.
• Guide points: Left-click to add green positive points, indicating content that needs to be retained; right-click to add red negative points, indicating content that needs to be excluded. Click on existing markers to delete them individually, and "Clear" will delete all markers. Negative points need to be used in conjunction with at least one positive point. After adding a marker point, the prompt will no longer take effect.
• Mask offset: The subject mask edge can be adjusted between -3 and 3 in semantic cutout mode. If the edges of the result contain too much background or are cropped too tightly, you can try making small adjustments.
• Billing: Each cutout costs a fixed 2 credits; credits are not deducted in Unlimited Mode.

Mask Compositing

Mask compositing is used to manually blend the results between the "base image" and the "top image". A common use is to erase the unsatisfactory parts after enlarging/editing back to the original image, or to retain the better parts between multiple result images. It's closer to a post-production compositing tool than an AI regeneration.

• Choose the base image: The default base image is usually the original image; if the current card has multiple result images, you can also select other results as the base image. The top image size is used as the composition canvas size, so please enter the mask composition with the final result image you want to keep first.
• Black/white relationship: The whiter the mask, the more it leans towards the top image; the darker it is, the more it leans towards the bottom image; gray is mixed in proportion. You can use fill black/white to quickly create an initial mask, and then use the brush to make local adjustments.
• Brushes and opacity: soft-edged circles are suitable for natural transitions; hard-edged circles are suitable for areas with clear boundaries; brushes are suitable for textures or irregular edges. The brush transparency controls the strength of each application, and the overall transparency controls the blending strength of the current composite result.
• Reference mask: You can reuse the mask that has been used for partial editing, local upscaling or partial protection, and is suitable for continued synthesis and repair in the same area.
• Save: The mask compositing result is saved to the cloud for free. After the saving is completed, you can continue to edit, connect or use it as input for subsequent tasks.

Color Matching

Color Matching is intended for results whose content is essentially correct but whose color relationships still need work. It offers two approaches. Color Correction restores color drift introduced by AI editing, upscaling, or other processing, so the reference and target should normally depict the same image. Color Transfer applies the reference image's palette, lighting character, and overall atmosphere to a different target image. If the two scenes differ too greatly—for example, a sunny blue-sky scene versus a seaside sunset—Color Matching alone may be unstable; use the Migrate Color Style preset in Image Editing instead.

• Color correction: Mainly used to deal with the situation where "the same picture has a color cast in subsequent processes". For example, after a picture has been edited, enlarged, or processed through multiple rounds of AI, the overall color has shifted compared with the original picture. In this case, the original picture or the same source reference can be used to correct the color relationship back. Generally speaking, the reference picture and the target picture should correspond to the same picture.
• Color transfer: can handle the situation where the reference image and the target image are not the same image. Its function is not to correct the color back to its original color, but to transfer the color tendency of the reference image to another target image.
• Reference image: Determine the target direction of color matching with reference to the image. If the current task card has an original image, you can use the original image by default if you do not upload a reference image; if the current card does not have an original image, you need to manually upload the reference image before submitting it.
• Mask: Supports drawing masks like edit mode. After drawing the mask, color matching will take priority on the mask area, which is suitable for adjusting only the sky, clothing, skin, ground or local light and shadow without affecting other parts of the entire image.
• Preset parameters: Two presets are provided in the color transfer mode: "Natural" and "Strong". The former is more restrained, while the latter changes more obviously; if you continue to manually modify relevant parameters after calling the preset, the status will automatically change to "Customized".
• Algorithm weights: Color transfer is not simply a matter of choosing one of the three algorithms. It can be understood as two steps: first do "natural migration", then see whether to continue to superimpose highlight/shadow dyeing on this result, and finally mix it with "strong migration" in proportion.Highlight/shadow tinting determines how much dyeing is added to natural migration;Strong-transfer weight determines how much strong migration accounts for the final result. The former affects the degree of dyeing within the natural migration, and the latter affects whether the final result is more natural or more intense.

Image Annotation

Image annotation is used to directly draw the modification intention on the picture, which is particularly suitable for editing needs that "cannot be explained with words alone". For example, let AI modify a certain position, adjust local structure, explain the relationship between collage materials, etc. The annotation itself will not automatically change the image. It usually needs to be used in conjunction with the "Change image by annotation" preset in image editing.

• Annotation types: Supports rectangles, circles, arrows, text, picture layers and graffiti. Arrows are suitable for pointing to the target, text is suitable for writing modification requirements, picture layers are suitable for pasting reference materials, and graffiti is suitable for free circle selection or drawing irregular intentions. When adding a picture layer, in addition to clicking the button to upload, you can also drag the image directly into the annotation canvas, or use the paste method to quickly place it into the screen; when importing multiple pictures at one time, the system will automatically stagger them slightly to avoid complete overlap.
• Edit selected objects: After the annotation object is created, it can be selected again and moved, scaled and color adjusted; the arrow can adjust the endpoint, and the text annotation can be re-edited by double-clicking. When the object is not selected, rolling the mouse wheel will adjust the line thickness, graffiti brush size, or text font size of the current tool; when an object is selected, the scroll wheel is used to adjust the line thickness, text font size, or graffiti stroke thickness of the object.
• Multi-select and connected strokes: Hold down Shift and click to select multiple annotation objects for unified movement, grouping or level adjustment. In graffiti mode, press and hold Shift to connect a line from the previous graffiti point, which is suitable for drawing straight lines or polyline annotations.
• Right-click menu: After selecting the annotation object, you can adjust the level, group/ungroup, lock or delete it by right-clicking; after multiple selection, you can group it, and multiple graffiti objects can be merged into graffiti. Right-clicking in graffiti mode is an erasing operation, not opening a menu.
• Color, fill, and clear: The color button not only affects new annotations, but also changes the color of selected shapes, text or graffiti; "Fill black/fill white" can quickly set the background color of auxiliary annotations, and "Clear" is used to clear the current annotation content and rearrange the screen.
• Save: Image annotations are saved to the cloud for free. After saving, you can continue editing or drawing connections, or use the annotated image as input for a later task.

*Suggestion: Annotations should be "few but accurate". Too many colors, arrows, and text will clutter the picture and reduce the AI's understanding of the core modification intent.

Save Preset

Save preset is used to save the parameters of the current task card for direct reuse in the future. This function currently only supports image upscaling, image editing and image generation task cards. Subsequent processing cards such as color matching, local upscaling, mask compositing, and image annotation do not support saving presets.

• What is saved: What saves the preset record is the parameter combination corresponding to the current function, rather than saving the picture itself. In other words, the preset saves the "processing method", not the "processing result".
• How to use: After saving, it can be called again in the default entrance of the corresponding function. In this way, when you encounter a similar task next time, you don't need to re-adjust the parameters from scratch, you only need to fine-tune based on the existing presets.
• Difference from parameter reuse: Parameter reuse is more interested in temporarily bringing the settings of a certain historical task card back to the panel for further modification; saving the preset is more inclined to precipitate a set of methods worthy of long-term retention into a template that can be called repeatedly. The difference between
• Notes: Presets can only be used as a starting point and cannot be guaranteed to be directly applicable to all images. Even for the same type of subject matter, the original image quality, resolution, content complexity and target style are different, and further fine-tuning is often required.

Workspace Management

The workspace is used to manage task records for different projects, clients, or creative directions. If you use this site for a long time, it is recommended to establish a workspace by project as soon as possible, otherwise it will be difficult to retrieve historical tasks as it increases.

Workspaces and tabs

• Workspace home: You can view all workspaces and create, rename, delete or view details. The default workspace cannot be deleted.
• Workspace tabs: After opening multiple workspaces, they can be switched like browser tabs. Right-click a tab to close the current tab, close other tabs, or close the right tab.
• Task assignment: Newly submitted tasks will enter the current workspace. If you frequently switch between multiple workspaces, please confirm the current workspace before submitting to avoid misclassification of tasks.

List View and Grid View

• List view: suitable for viewing complete parameters, continuing to edit, download, copy or perform card-level operations.
• Grid view: suitable for quickly browsing a large number of historical results. You can filter by task type, such as image upscaling, image editing, image generation, local upscaling, image cutout, mask compositing, color matching, image annotation, etc.
• Batch management: Grid view supports batch deletion after multiple selections; right-click task cards to move to other workspaces. Deleting a task will clean up the corresponding records and files. Please confirm that they are no longer needed before proceeding.

Toolbox

• Image Split & Merge: Applicable to the following two situations:
  1. Grand and complex scenes: When the content complexity exceeds the upper limit of the website's maximum 32MP (8K) capacity, it can be split and processed before merging.
  2. Ultra-large resolution requirements: For example, long rolls and other images that exceed the 8K limit should be split and enlarged before merging.

  *Note: Seam processing: If ghosting appears at the seams after merging, you can use the "local optimization" in the editing function to repair it.

• Image Comparison: Supports up to 3 groups of pictures for comparison on the same screen. It is very suitable for final effect testing after multiple rounds of iterations or splits and mergers.

Conclusion

GRANDI-AI is committed to becoming the most powerful AI image enhancement tool on the planet. If you are not satisfied with the processing results, you are welcome to click the "Not good results?" button on the task card to submit feedback, and we will usually respond within 3-5 working days. If you need immediate help, business consultation, or problem feedback, please contact customer service on WeChat:cganimitta.