Prompts LLaDA-Image
Prompts LLaDA-Image — Structures pratiques
Steal structure, not fluff: subject, lighting, bilingual text strings, camera notes, and edit instructions that map cleanly to the online workspace.
Ce site est un espace de travail tiers indépendant pour LLaDA-Image. Il n’est pas affilié à inclusionAI. Le code et les poids officiels restent sur GitHub, Hugging Face et ModelScope.
Prompt patterns that match the model
Good LLaDA-Image prompts are specific about scene and on-image text. For edits, separate “what stays” from “what changes”. These examples are starting points for this independent workspace — tune them against official sampling guidance.
Three prompt templates
Copy the structure, then replace brackets with your brief.
Photoreal generation
“Photoreal [subject] in [place], natural window light, shallow depth of field, 35mm look, no watermark.”
Bilingual poster
“Clean poster layout, headline ‘[EN text]’, subhead ‘[ZH text]’, high-contrast typography, generous margins.”
Instruction edit
“Keep the same person and pose. Change only the background to [scene]. Preserve facial identity and clothing.”
Prompt hygiene
Avoid stuffing the same keyword twenty times. Prefer concrete nouns, readable quoted text for posters, and one primary change per edit instruction. Use Turbo to explore variants, then Base for the keeper.
- Quote exact poster strings instead of paraphrasing.
- State what must remain unchanged in edits.
- Log prompts that worked for your brand kit.
Pages associées
Prompts FAQ
Do prompts differ for Turbo?
Usually the same prompt works; Turbo may need cleaner instructions because it uses fewer steps. Compare both before judging the model.
Can I reuse Midjourney-style stacks?
Some style words transfer, but prioritize plain descriptive English/Chinese and exact text strings over proprietary parameter soup.
Where else can I learn?
Read the Best Prompts blog post on this site and the official GitHub examples for grounding.
Are these “best prompts” ranked scientifically?
No. They are practical patterns for this workspace, not a forged leaderboard.
Run a prompt now
Pick one template, open the generator, and save both Base and Turbo outputs for comparison.