FlowyAIPC
  • English

    • Image Generation

    Generate and edit images locally using Herdsman's text-to-image and image-to-image models.

    This feature is hardware-intensive. At least 6 GB of VRAM is required.

    Quick Start

    1. Type your image description in the chat box and click the Generate Image button.

      Generate Image entry

    2. Wait for the model to render the image. A progress bar appears below the chat box.

      Render time varies with hardware and prompt complexity. Please be patient.

      Progress bar

    3. Once complete, the result appears in the Generated Result panel on the right.

      Generated result

    Parameter Reference

    1. Size

    • Controls the aspect ratio and resolution of the output.
    • Presets: 1:1, 16:9, 9:16, 3:2, 4:3, or a custom resolution.
    • Higher resolution yields more detail at the cost of longer render times.

    2. Negative Prompt

    • Tells the model what not to include.
    • Example: low quality, blurry, deformed, extra fingers, text, watermark — to suppress common artifacts.
    • Particularly useful for keeping the image clean and improving anatomical accuracy.

    3. Steps

    • Number of diffusion iterations. Higher values produce more detail but take longer.
    • Default: 9.
    • Returns diminish after a certain point — more steps is not always better.

    4. CFG Scale (Classifier-Free Guidance)

    • Controls how strictly the model adheres to your prompt.
    • Higher values stick closer to the prompt but may look oversaturated or unnatural; lower values are looser and softer.
    • Default: 1.0.

    5. Sampler

    • The algorithm that turns noise into the final image step by step.
    • Default: euler — a classic, fast, stable sampler well-suited to photorealistic styles.
    • Alternatives such as DPM++, DDIM, and UniPC offer different speed/quality trade-offs.

    6. Scheduler

    • Controls the noise-decay schedule, affecting smoothness and detail.
    • Default: discrete — stable and reliable across most models.
    • Continuous schedulers (e.g., karras, exponential) can sharpen edges and improve texture.

    7. Seed

    • Controls the randomness of the result. With the same seed and parameters, the output is virtually identical.
    • -1 generates a new random seed each run.
    • To reproduce a specific image, record and fix its seed value.

    8. ZImage Model and LoRA

    • ZImage Model: the base generation model (e.g., SDXL, SD 1.5, Flux). The default is Default Model.
    • LoRA: lightweight fine-tunes that bias the output toward a particular style, character, or outfit.
      • "No LoRA imported" means no extensions have been loaded yet.

    Generation parameter panel

    Generation History

    The history panel stores previous prompts, making it easy to iterate or save high-quality prompts.

    1. Click a history entry to load its prompt back into the input box.

    FAQ

    Q: Why can't my computer generate images?

    • Confirm that a local text-to-image model is installed. If not, install one first.
    • Check your hardware against the model's requirements — image generation has higher demands than chat.
    • If hardware is sufficient but generation still fails, you may have too many models running and competing for resources. Stop the models you aren't using: Models page → review the running models → stop the ones you no longer need.

    Q: Why does the generated image look different from what I imagined?

    There are four common causes:

    1. The prompt is too vague — the model interprets the literal text and cannot read your mind. Compare "a cat" vs. "an orange tabby cat sleeping on a windowsill in afternoon light" — specificity matters.
    2. Defaults fill in the blanks — anything you don't specify (composition, lighting, style) is decided by the model, and its defaults rarely match what's in your head.
    3. Recognition vs. understanding — the model knows a hand has five fingers, but it does not understand how fingers naturally rest. Complex spatial relationships, perspective, and text are common weak points.
    4. Subtext is invisible — emotional qualities like "lonely," "dramatic," or "warm" must be stated explicitly; the model cannot infer them.

    Q: Why does generated text come out garbled?

    Text-to-image models do not actually read text — they treat letters as "shapes" to be drawn.

    Analogy: Asking the model to write the word "Hello" is like asking someone who doesn't speak English to copy unfamiliar strokes — the result is usually messy.

    Tips to improve text in generated images:

    • Minimize the amount of text in the image.
    • If text is essential, keep it short and use a simple font.
    • For best results, add text in a post-processing tool like Photoshop rather than asking the model to draw it.