ComfyUI lesson 2: Diffusion, seeds, CFG, samplers, and schedulers
Learn what each setting changes by holding every other variable still.
Listen to this article
On-device voiceUses the voice built into your browser; no article text leaves this page.
Audio narration is not supported by this browser.
The controls stop feeling mystical when you separate what each one owns. Keep the Proteus workflow open and change one value per run.
Diffusion in one useful paragraph
Generation begins with noise in a compressed latent space. At each step the model predicts which noise should be removed given the text conditioning. The scheduler chooses the noise levels; the sampler chooses the numerical path through them. The VAE finally decodes the cleaned latent into pixels.
What the settings actually control
| Control | Practical meaning | Good experiment |
|---|---|---|
| Seed | Initial noise and therefore composition | Compare 42, 43, 44 |
| Steps | Number of denoising updates | Compare 12, 24, 40 |
| CFG | Pressure to follow text conditioning | Compare 3.5, 5, 7 |
| Sampler | Route through latent space | Compare Euler and DPM++ 2M SDE |
| Scheduler | Spacing of noise levels | Compare normal and Karras |
| Denoise | Fraction of the denoising journey | Leave at 1.0 for txt2img |
A fixed seed is an experimental control, not a promise across every GPU, PyTorch version, or node implementation. It is still the best way to see whether CFG or sampler choice caused the visible difference.
Proteus V0.4's published guidance is a useful starting region: CFG 4–6, 20 or more steps, DPM++ 2M SDE, Karras, and roughly 1024-square output. More steps are not automatically more art. Past convergence they mostly buy time.
A 12-image contact sheet
Hold the prompt and seed still. Render three CFG values across four step counts. Label every output. Pick the cheapest cell that already has the composition and detail you need. That small grid will teach more than copying a stranger's “best settings,” because it measures your model, prompt, and hardware.
Run the lesson, do not just read it
- Open the tested workflow in Comfy Studio
- Download the exact API-format JSON
- CLI:
app comfy download proteus-v04-text-to-image
The gallery records model sources, minimum VRAM, custom-node requirements, and verification state beside the graph. That is the portable unit for this series.
Series progress: lesson 2 of 10. Previous: Install ComfyUI cleanly and make your first image Next: Prompt with intent, not incantations