Prompting

This page covers prompt-related features in Local Dream.

Generation Modes

• txt2img — Generate images from text prompts.
• img2img — Transform an existing image with a prompt and a denoise strength.
• inpaint — Redraw a selected area of an image while keeping the rest intact.

For a beginner-friendly explanation of prompt, negative prompt, CFG, denoise strength, and seed, see Generation Parameters.

Prompt Weights

Emphasize or de-emphasize words in a prompt using the same syntax as AUTOMATIC1111:

(masterpiece:1.5), (low quality:0.6), 1girl

Parentheses () are treated as prompt-weight syntax by default. If you want literal parentheses to appear in the prompt text itself, escape them as \( and \).

Example:

kafuu chino \(gochiusa\)

Prompt Token Counter

Local Dream shows a prompt token counter with the current token count and the model's maximum token budget (77 tokens for SD1.5 / SDXL CLIP, including the BOS/EOS markers).

This is useful because Stable Diffusion text encoders do not treat extremely long prompts as unlimited free text. If a prompt becomes too long, later content may become less useful or may be truncated by the model pipeline. Textual inversion embeddings also expand into multiple tokens, and that expansion is reflected in the counter.

Practical advice:

• keep the main subject and important details early in the prompt
• do not assume a very long prompt is always better
• if the counter is near the limit, simplify before adding even more tags

Embeddings

Custom textual inversion embeddings (e.g. EasyNegative) are supported. You can import embeddings from the app's settings.

Format notes:

• Many older SD1.5 embeddings are distributed as .pt
• Almost all SDXL embeddings are already distributed as .safetensors
• Local Dream imports embeddings in SafeTensors format

If your embedding is a .pt file, convert it to .safetensors before importing. As one reference, you can use this user-provided conversion script: pt2sf.py.

Important: importing an embedding does not activate it automatically. The embedding is only used if you put its trigger word into the prompt.

In Local Dream, the trigger word is the embedding's filename without the extension. For example:

• File: 12345.safetensors
• Trigger word in prompt: 12345

So if you import 12345.safetensors but never write 12345 in your prompt, the embedding will not be applied.

You no longer need to memorize the filename to type it — imported embeddings now appear at the top of the prompt autocomplete popup with their own badge. See Tag Autocomplete → Embedding Suggestions for details.

Inpaint Canvas Zoom

Local Dream supports zooming while using the inpaint canvas.

This is especially useful for:

• painting masks on small details
• checking edges more carefully
• editing on smaller phone screens

Inpaint — Fix Local Details on a Generated Image

Use case: fix local details on a generated image — clean up a hand, swap an outfit, redraw a face — without redoing the whole image. The masked region is regenerated; everything outside the mask is preserved exactly.

• At mask edges: the generated region is composited with the original using Laplacian Pyramid blending, producing a smoother seam than simple alpha blending.
• At save time: when you save from the result screen, Local Dream automatically reads the uncropped original image and composites the inpaint result back into it, saving the full image. The generation history only stores the cropped image.

This makes inpaint a practical refinement step on top of any txt2img output (including fast DMD2 generations), without needing an external editor.