Prompt guide
I am looking for clear documentation on prompt structure, supported editing instructions, best practices, constraints, and examples covering common and advanced use cases such as object addition/removal, style transfer, background modification, and fine-grained control. A deeper explanation of how the model interprets instructions, handles ambiguity, and prioritizes edits would be extremely valuable for achieving consistent and high-quality results. Such a guide would greatly help users like me better understand the model’s capabilities and limitations and use it more effectively in real-world workflows.
Another notable source of information is the official Qwen Image report from August
https://arxiv.org/pdf/2508.02324
Also Qwen Image repository contains sample script for rewriting edit style prompts and you can use instructions from there in any "chat" LLM as the system prompt for JSON style prompting:
https://github.com/QwenLM/Qwen-Image/blob/a76c8a3873c369a097aafd7ea229b7404659043c/src/examples/tools/prompt_utils.py#L183
it's worth keeping in mind that it uses a llm as clip. It is a LOT more flexible than all the various guides suggest. We're long past "THAT is how to do it", as we were a year or so ago.
so don't hesitate to just experiment. You can basically just talk to it and tell it what you want in natural language and get often very good results.
And you can push it way beyond simple editing jobs. For example "isolate the subjects in image 1 and image 2 and put them in a new scene" And then just describe like a normal image gen prompt. r do just that without any image input (need to connect the empty latent node to the ksampler).
It essentally replaces the normal Qwen image or Z-image (i'm not sure how good it is compared but it seems basically as capable.).
just use a 4step lora for quick tries and then a 8step or native settings (25-40 steps, cfg 3) for the high quality attempt.