Build a Basic ComfyUI txt2img Workflow Without Guesswork
Academy is temporarily unavailable while it is being rebuilt.
A common beginner problem in ComfyUI is staring at a blank canvas and not knowing which nodes are required. For a basic text-to-image workflow, you need six parts: a checkpoint loader, text encoders, an empty latent image, a sampler, a VAE decoder, and an output node.
1. Add Load Checkpoint. This node provides three outputs: MODEL, CLIP, and VAE. Select a checkpoint from the dropdown. If the list is empty, check that your model file is in the checkpoints folder used by your ComfyUI installation.
2. Add two CLIP Text Encode (Prompt) nodes. Use one for the positive prompt, such as “a quiet mountain lake at sunrise,” and one for terms you want the model to reduce or avoid. Connect the checkpoint’s CLIP output to both encoder inputs. Check success by socket type: the CLIP wire should connect cleanly, while incompatible socket types will not connect.
3. Add Empty Latent Image. This sets the starting canvas before sampling. Choose dimensions that make sense for the model family you selected rather than treating one resolution as universal. Keep batch size at 1 while testing so errors are easier to isolate.
4. Add KSampler. Connect checkpoint MODEL to `model`, the positive encoder to `positive`, the negative encoder to `negative`, and Empty Latent Image to `latent_image`. Steps, CFG, sampler, scheduler, seed, and denoise all affect the result, but their best values depend on the model and workflow. To compare settings, change only one control at a time and observe the output.
5. Decode and output the image. Connect KSampler’s latent output to VAE Decode, connect checkpoint VAE to the decoder, then send the decoded IMAGE to Save Image or Preview Image.
The workflow is working when required inputs are connected, the prompt queues without node errors, and a visible image appears. Save the graph as JSON so you can reload it and build on it later.