Free lesson

Define a modest local image experiment

Define a modest local image experiment — Local image workflows with ComfyUI. Write an observable visual brief; Distinguish disk, RAM and VRAM; Choose local software instead of cloud nodes. Practical assignment: Lab evidence.

Demo edition. The materials are undergoing editorial review.

Topics covered

  • Write an observable visual brief
  • Distinguish disk, RAM and VRAM
  • Choose local software instead of cloud nodes

Write an observable visual brief

Use an invented still-life brief: one matte blue ceramic bowl on a plain light table, no people, logos or readable text. This provides visible acceptance criteria without impersonating a person or counterfeiting a branded product. The result is a concept illustration, not proof that a real item exists or matches a commercial product. Write a checklist before generation: one bowl, blue surface, full rim visible, simple background and no extra utensils. An AI assistant may help tighten the brief, but do not let it add unsupported claims such as handmade by a named company. Keep a neutral description of generated content for any later handoff. No real reference photograph is needed in this course. The first artifact is a saved brief and rejection criteria, so later selections can be explained instead of choosing whichever image looks impressive.

The requested blue is a visual direction, not a calibrated manufacturing color. Do not assign a paint code or exact product specification from generated pixels; retain the concept label when presenting the image outside the practice folder.

Distinguish disk, RAM and VRAM

Record the practice machine’s operating system, architecture, free disk, system RAM and GPU model with dedicated VRAM if present. A 2.13GB checkpoint download does not mean the workflow needs only 2.13GB of VRAM: text encoding, latent processing, decoding and other allocations consume additional memory. Apple unified memory is also not interchangeable with dedicated GPU VRAM, though this walkthrough uses Windows. Start with a single 512-by-512 latent, batch size one and a small fixed step count. Doubling both dimensions quadruples pixel area, but actual memory growth depends on model operations rather than a universal simple formula. CPU mode is a slower fallback only when the workload fits available RAM and the runtime supports it. If resources are inadequate, produce a prepared workflow and compatibility report without renting hardware automatically.

Budget space for the extracted package, checkpoint and output copies simultaneously. Deleting the only known-good checkpoint to make room for an unverified replacement removes rollback; inspect storage first and keep the experiment prepared-only when space is insufficient.

# Planning record, not a performance guarantee
image_width = 512
image_height = 512
batch_size = 1
queued_runs = 1
execution_device = documented compatible GPU or supported CPU mode
measured_runtime = not_measured

Choose local software instead of cloud nodes

This course uses the self-hosted Windows Portable package and core local nodes. Comfy Cloud, partner API nodes and optional online agents are different execution paths and may involve accounts, uploads or billing. Do not assume the word local in an application title applies to every node in a downloaded workflow. Before installation record the official release source and the hardware-specific package selected. Keep the service on loopback and do not add a public listen address to make setup easier. A browser opening 127.0.0.1:8188 proves access to a local interface, not that a later workflow cannot call a remote provider. The planned seven-node graph has no network service node or custom extension. Your outcome is a simple boundary diagram naming local browser, local ComfyUI process, local checkpoint files and local output directory.

Keep a list of nodes and any account prompts encountered during setup. An unexpected login or credit request means inspect the selected feature, because the local core graph does not need a partner-provider account or paid inference credential.

Practical assignment

Lab evidence

Continue to learning

Course outline

  1. Define a modest local image experiment
  2. Verify software, checkpoint and installation sources
  3. Build the complete seven-node graph
  4. Set a bounded run and compare one variable
  5. Diagnose failures and audit without generating
  6. Package evidence, inspect output and preserve recovery