AI images: a traceable experiment
Learn workflows and model files, then compare prompts one variable at a time.
Models generate and reason. Agents use tools to work on tasks. A harness manages an agent’s environment. MCP connects tools and data. Skills are reusable task guidance.
Current task
Deliver 2 A/B images, 1 reloadable JSON, a model/license/settings sheet and 3 visual review notes; the same seed does not guarantee identical pixels across versions.
Prerequisites
Basic file skills; official ComfyUI, a specific image model supported by your hardware, and authorized references. Record license, size and hardware requirements before downloading.
Estimated time
Plan 3–5 sessions of 30–60 minutes; downloads and rendering are additional, not measured.
Limits & boundaries
Wrong model directory, mismatched encoder/VAE, unfixed seed or VRAM exhaustion; reduce size and batch first, without guaranteeing speed from GPU names.
Validation
Deliver 2 A/B images, 1 reloadable JSON, a model/license/settings sheet and 3 visual review notes; the same seed does not guarantee identical pixels across versions.
Next steps
Pass the route’s acceptance checks before adding one variable; retain failed samples, versions, costs and rights records.
Cost
Not verified
Amounts unverified. Local software can use existing hardware; model licensing, VRAM, storage, electricity and labor are separate. Cloud, APIs or subscriptions may charge.
Implementation steps
- 01Build a minimal workflow and asset list
- 02Change one variable
- 03Reopen and review publication files
Implementation steps
- Build a minimal workflow and asset list
- Change one variable
- Reopen and review publication files
Original sources
- ComfyUI first image tutorial ↗
docs.comfy.org · 2026-10-16 · Official documentation