Start with a basic text-to-image workflow: load a compatible checkpoint, encode a positive and negative prompt, sample a latent image, decode it, and save the result.

Watch on YouTube · Video published 2026-01-28 · Companion reviewed 2026-09-23

What you need

A working ComfyUI installation, one checkpoint you are licensed to use, and enough available memory for that model. This lesson demonstrates a Stable Diffusion 1.5-style checkpoint workflow. FLUX, Z-Image and SDXL have different model requirements; do not mix their components.

New installation? Watch the installation lesson or use the cloud setup guide. Start with one model and one image before adding custom nodes.

Build the first image

  1. Load Checkpoint selects the model file. Put a compatible single-file checkpoint in models/checkpoints, refresh the model list, and choose it in the node. Its MODEL, CLIP and VAE outputs serve different parts of the graph.
  2. Connect CLIP to two CLIP Text Encode nodes. One describes what you want; the other supplies negative conditioning when the selected model and sampler support it. Connect both conditioning outputs to KSampler.
  3. Connect MODEL to KSampler and an Empty Latent Image to its latent input. For an initial SD1.5 test, 512 × 512 and batch size 1 are a modest starting point. Follow the checkpoint author’s guidance when choosing other sizes.
  4. Use the checkpoint’s recommended sampler, steps and CFG. Fix the seed while comparing changes. Increasing steps or CFG is not an automatic quality upgrade.
  5. Connect KSampler’s latent output to VAE Decode, connect the checkpoint’s VAE, then connect the decoded image to Save Image. Run once and inspect the saved output.
  6. Save your workflow JSON and note the exact checkpoint filename. A reusable graph without its model/version notes is difficult to reproduce.

Try a small, readable prompt

These are new practice examples, not prompts claimed to reproduce the video’s exact output. They target a conventional SD1.5 workflow with supported negative conditioning.

Positive

a ceramic coffee mug on a wooden desk, morning window light, close-up product photograph, shallow depth of field

Negative

blurry, distorted handle, duplicate mug, watermark

Keep one subject and one lighting idea. If the image misses the subject, simplify before adding more adjectives. Commas separate readable concepts here; they are not a universal “better prompt” setting. Weighting such as (morning light:1.2) depends on the prompt parser and workflow.

Change one thing at a time

Keep the seed, checkpoint, resolution and sampler fixed. Change only the lighting phrase, generate again, and compare the two images. Then test a different seed. A fixed seed helps controlled comparisons within a compatible environment; it is not a character-identity lock across models or software changes.

If the first run fails

About the older interface in the video

The video was published in January 2026. Node menus and frontend features may look different now. Its beta-interface warning is historical, and its broad sampling ranges should not replace the settings recommended for your specific checkpoint. The model → conditioning → sampling → decoding sequence remains the concept to learn.

Sources and downloads

Related lessons