generating-images

Template + patch Skill for text-to-image. No LLM emits graph JSON — compile.py applies a flat override map from the Visual Intent Object onto a bundled template and writes the result to stdout. Pipe into comfyui-execution's submit_prompt.py.

When this Skill applies

• The planner produced a step with {goal: "generate"}.
• A caller wants a one-shot txt2img from a natural-language prompt.

Templates

• templates/txt2img-sdxl.json — SDXL base, 1024², euler/normal, cfg 7.
• templates/txt2img-flux.json — FLUX.1-schnell variant.

Template selection: VIO model.checkpoint hint containing "flux" picks txt2img-flux.json; otherwise txt2img-sdxl.json.

Provenance in templates/PROVENANCE.md.

Intent overrides honored

VIO field	Node.input
prompt	6.inputs.text
negative_prompt	7.inputs.text
params.width / params.height	5.inputs.width / height
params.steps	3.inputs.steps
params.cfg	3.inputs.cfg
params.sampler_name	3.inputs.sampler_name
params.scheduler	3.inputs.scheduler
params.seed	3.inputs.seed (-1 → random via time)
model.checkpoint	4.inputs.ckpt_name

Scripts

• scripts/compile.py — stdin: VIO JSON; stdout: prompt JSON. stdlib only.

Errors

Bad VIO → {"error":"...","code":"..."} on stdout, exit 2.

References

• knowledge/workflows/ComfyUI_workflow.md §3 (prompt JSON shape)
• ../planning-visual-tasks/reference/intent-schema.md