Reading note
Generating or modifying images with ChatGPT, Gemini, Doubao and similar tools through text requests, reference images and ongoing feedback is a convenient form of image creation with a lower technical barrier. It is outside this chapter's scope.
Authoritative Midjourney tutorials
Midjourney's official documentation includes illustrated instructions and short videos covering the main operations, from a first generation to reference-image controls and local edits.
| Official resource | Main content |
|---|---|
| Getting Started Guide | Web generation, settings, image organization and entry points for further editing |
| Prompt Basics | Describing subjects, media, environments, lighting, colors and composition |
| Image Prompts | Using images to influence content, composition, colors and other visual features |
| Style Reference | Borrowing a reference image's palette, materials, medium and lighting style |
| Edit Model | Instruction-based editing, multiple references, local edits and outpainting |
| Modifying Your Creations | Variations, upscaling, Remix, panning and extending an image |
Also consult the version and feature compatibility table. For example, the current documentation distinguishes V8-series reference-image editing from the entry points and compatibility of older Character Reference and Omni Reference features. Older tutorials may point to controls that are no longer available in the same place.
Authoritative Stable Diffusion tutorials and resources
Stable Diffusion (SD) is a family of image generation models; SDXL and SD 3.5 are different series within it. To create images, you load a model through an application or code. Distinguish the following names before using these resources:
| Name | What it is | Practical use |
|---|---|---|
| Model card | Documentation on a model's release page | Check its publisher, intended uses, files, instructions, requirements, limitations and license; primarily a source for verifying model information |
| ComfyUI | Software for running models and organizing generation steps | Represent model loading, requests, generation, editing and saving as connected nodes in a workflow; supports SD and other models |
| WebUI | Software for operating a model through a web interface; here, AUTOMATIC1111 Stable Diffusion WebUI | Configure and run tasks with text fields, sliders and buttons in text-to-image, image-to-image and other panels |
| Diffusers | A programming library maintained by Hugging Face | Load models and generate or edit images in Python; its documentation also explains parameters and compares outputs |
Their relationship is straightforward: operate a model through ComfyUI or WebUI, learn about it through its model card, and use Diffusers when you need to run it in code. Installing an application is only part of preparation: you also need the model files and supporting components. Beginners can start with a visual interface and consult model cards and parameter explanations when needed; Python is not a prerequisite.
| Resource | Provider and purpose |
|---|---|
| Official SDXL model card, official SD 3.5 Large model card | Stability AI: check model files, usage, supporting components and limitations |
| Official ComfyUI text-to-image tutorial in Chinese | A practical introduction covering model loading, requests, execution and saving; its basic example uses SD 1.5 |
| ComfyUI maintainer's examples | Complete image-to-image, inpainting, LoRA, ControlNet, upscaling and SDXL workflows; load the example images in ComfyUI to inspect the workflow |
| AUTOMATIC1111 feature documentation | Instructions for this WebUI, including inpainting, parameter comparisons, batch processing and image information; check your version for specific features |
| Hugging Face Diffusers practical documentation | Code-oriented, but useful parameter explanations and comparison images help establish what a control actually does |
In particular, ComfyUI's official Chinese tutorials and example library are the best starting points for following along; Diffusers is especially useful for checking technical details.
The following sections introduce these techniques through practical creation tasks.
1. Text-to-image: turn visual requirements into comparable candidates
The main elements to describe are:
- Subject: what to depict, how many subjects, and their appearance.
- Actions and relationships: what each subject does and how objects relate.
- Composition: wide or close view, subject placement, viewpoint and empty space.
- Medium: photography, watercolor, line drawing, 3D rendering and so on.
- Lighting and color: light direction, brightness, temperature and main palette.
- Aspect ratio: landscape or portrait, and where the image will be used.
For example, this request can be used for practice:
An orange fox sitting on a forest path after rain, full body in side view, subject on the right with space for a title on the left, watercolor illustration, muted green background, soft morning light, landscape format.
Compare the subject and composition first, then adjust style and details. Midjourney's official prompt guide also emphasizes clearly stating key features rather than piling on lengthy requirements. Prompt Basics
2. Reference images: specify what the model should use
Reference images serve different purposes. Keep them separate during operation:
| Reference purpose | Example | Corresponding operation |
|---|---|---|
| Content and composition | Room layout or object arrangement | Image prompts or structural controls |
| Style | Watercolor texture, palette or lighting | Midjourney Style Reference or an appropriate image-reference method |
| Subject | Preserve a person or product's features in a new scene | Midjourney editing references, IP-Adapter and similar tools |
| Modify an existing image | Keep the overall image while changing clothing or surroundings | Image-to-image or image editing |
A similar style and the same person are different goals. Midjourney states that Style Reference primarily transfers an overall visual style rather than copying people or objects from the reference. In the SD ecosystem, IP-Adapter guides generation with images alongside text and allows adjustment of the reference's influence. Style Reference, IP-Adapter guide
3. Image-to-image: control the amount of change
Image-to-image is useful for completing a sketch, changing a medium or making variations on an existing image.
An important control is denoising strength, often named strength or denoise:
- Lower values usually retain more of the original structure.
- Higher values usually permit greater changes.
- Too little change may leave the requested edit undone; too much may alter parts you wanted to keep.
Keep other settings fixed and compare strengths to observe the tradeoff between retaining the original and making the edit. Diffusers image-to-image guide
4. Inpainting and outpainting: define where changes happen
Inpainting uses a mask to select a region, followed by a description of what should appear there: replacing a hat, removing background clutter or modifying an object, for example.
Check the following during operation:
- Does the mask cover the whole object and the edges that need to blend?
- Is enough surrounding context available to the model?
- Do the new content's perspective, lighting and colors match its surroundings?
- Did the edit affect anything that should have remained unchanged?
Outpainting adds content beyond the original boundaries, such as extending a square illustration into a landscape cover. Inpainting guide, Midjourney editing documentation
5. ControlNet: control structure with poses, lines and depth
When a task requires specific positions, actions or spatial relationships, provide those conditions directly as control images.
| Control condition | Practical use |
|---|---|
| Edges or line art | Retain building outlines, object boundaries or sketch structure |
| Human pose | Specify standing or sitting poses and the broad movements of arms and body |
| Depth | Specify distance, spatial layers and approximate object volumes |
A typical workflow is: prepare a reference → extract the required control image → inspect it → load a matching control model → adjust control strength → generate and compare.
Inspect the control image itself first: a wrongly detected pose can constrain subsequent generation incorrectly. ComfyUI's examples also explain that ordinary photos need suitable preprocessing into the format expected by the control model. ComfyUI ControlNet examples, Diffusers ControlNet guide
6. LoRA: add customization for a style, person or concept
When using an existing LoRA, check:
- Which base model it matches.
- Whether it needs specific trigger words.
- Its recommended influence strength.
- How loading it changes the subject, style and other image content.
Compare an output without LoRA against one with a single LoRA before adjusting strength. Training your own LoRA is a further customization task: prepare suitable training images and descriptions, then check its performance across scenes. Diffusers LoRA documentation, ComfyUI LoRA examples
7. Parameters: learn to compare rather than memorize numbers
In common SD interfaces, these controls are especially useful:
| Parameter | Main practical effect |
|---|---|
| Model and version | Available content, styles and supporting components |
| Width, height and aspect ratio | Composition and usable output dimensions |
| Seed | Controls the random starting point to help compare settings |
| Steps | Generation time and output behavior; choose for the model |
| CFG / guidance scale | Text conditioning strength; excessive values can also harm results |
| Sampler / scheduler | Specific generation settings affecting speed and output behavior |
| Denoise / strength | How much image-to-image changes the original |
Start with the model or workflow's recommended settings. Hold other conditions fixed and compare a small number of variables at a time to understand the cause of a change. AUTOMATIC1111's X/Y/Z Plot is designed for such parameter comparisons. ComfyUI parameter documentation, X/Y/Z comparisons
A seed alone does not guarantee identical images across models, software versions or devices. Record the model, parameters and environment as well. Diffusers reproducibility guide
8. Upscaling and saving workflows: keep results reusable
After choosing an image, increase its dimensions with an upscaling model, then check text, faces, materials and edges for unwanted changes. ComfyUI provides complete examples loading ESRGAN and other upscalers. Primary technical resources also include the original Real-ESRGAN project. ComfyUI upscaling examples
Save the final image along with the model name, prompt, references, parameters and workflow. This lets you continue from existing work when changing dimensions or style, or extending the same series.
For example, three scene illustrations of one character can combine subject references, pose control, style references or LoRA, inpainting and workflow records. Practical skill is largely the ability to select the right control for a specific discrepancy and judge whether the edit succeeded.