Image Skill — Modelscope API Inference

Generate and edit images via Modelscope's API Inference, through the image-gen CLI.

If image-gen is not found on PATH, or config.yaml is missing/needs an API key, stop and read references/AGENTS_README.md first.

Commands

Every generate/edit command accepts: --negative-prompt, --steps, --cfg-scale, --width, --height, --seed, -o/--output (output directory, defaults to ./outputs relative to wherever the command runs), -p/--prefix (filename prefix).

edit defaults to Qwen/Qwen-Image-Edit when --model is omitted, since not all models support editing.

By default generate/edit block until Modelscope's async task finishes — typically 20s to 2min, occasionally longer for edit. If you need to keep working while it runs, don't background the whole image-gen process at the shell/tool level — a nested background job (a backgrounded command that itself launches another background job) can report "completed" as soon as the outer wrapper returns, before image-gen has actually finished and written its output. Instead, add --no-wait: it submits the job and returns a task id in well under a second, so no OS-level backgrounding is needed at all. Do other work, then check it with image-gen status <task_id> (single check, exits immediately either way) or image-gen status <task_id> --wait (blocks, then downloads) whenever it's convenient — that final status call only reports done when the task is genuinely done.

Writing the prompt

Apply before every generate/edit call — generation and editing need genuinely different prompt shapes, and getting this wrong is the main cause of bad output:

• Generate with Qwen-Image: order the prompt subject → style/medium → setting → lighting/mood → detail, 1–3 sentences. Any text that must render in the image goes in "double quotes" — this alone is the biggest lever on text accuracy.
• Generate with Z-Image-Turbo: write a long, detailed creative brief instead — this model rewards detail. It has no negative-prompt support at all; --negative-prompt is silently ignored, so fold every constraint into the main prompt as a positive statement ("sharp focus" not "not blurry").
• Edit with Qwen-Image-Edit: write an instruction describing the change, never a description of the end state. Bad: "A woman with short brown hair in a blue blouse". Good: "Change her hair to short brown and her blouse to blue". When multiple similar objects are in frame, disambiguate by position/attribute ("the second person from the left"). For in-image text edits, quote the exact replacement text — font/size/style carry over unless you say otherwise.
• For a multi-stage edit, run edit twice — feed the first result back in as the next image argument — rather than stacking unrelated changes into one instruction.

Full rationale, more examples, and per-model parameter recommendations (--cfg-scale/--steps) — references/prompt-guide.md.

Available Models (from config.yaml)

• Tongyi-MAI/Z-Image-Turbo — Fast, low cost, LoRA support (generate only)
• Qwen/Qwen-Image — High quality, editing support, LoRA support
• Qwen/Qwen-Image-Edit — Image editing, LoRA support

Further reference

• Full prompt-writing rationale, examples, and sources — references/prompt-guide.md.
• Installing image-gen as a uv tool, first-time config.yaml setup, adding new models, and how output_dir resolves — references/AGENTS_README.md.
• Calling ModelscopeClient directly from Python and passing loras (not exposed via the CLI) — references/python-api.md.