July 22, 2026·6 min read·app.nz

ComfyUI lesson 5: Build image-to-image and master denoise

Encode a source image and use denoise to choose between a retouch and a rebuild.

Listen to this article

On-device voice

Uses the voice built into your browser; no article text leaves this page.

Audio narration is not supported by this browser.

Image-to-image replaces the empty latent with a latent encoded from real pixels. Everything else becomes familiar, which is why denoise—not a mysterious new model— is the main control.

Rewire text-to-image

Add Load Image and VAE Encode. Feed the source pixels and the checkpoint VAE into VAE Encode, then connect its LATENT output to KSampler. Keep the text conditioning and decode/save tail.

The source image should use dimensions sensible for the checkpoint. Resize before encoding if you need predictable framing. A portrait forced into a square can be a bigger composition change than any sampler setting.

Denoise is distance from the source

Denoise	Typical use
0.15–0.30	Texture cleanup, light retouch
0.35–0.55	Restyle while preserving layout
0.60–0.80	Strong reinterpretation
0.90–1.00	Source contributes little beyond broad structure

These are starting bands, not laws. Run 0.30, 0.50, and 0.70 at the same seed. Choose the lowest value that accomplishes the edit: it preserves more intentional information and is usually easier to art-direct.

The lesson workflow maps source image, prompt, denoise, and seed. Hosted media inputs may be public URLs or app.nz artifact references; local ComfyUI expects a filename already present in its input directory.

If the output is nearly identical, increase denoise or strengthen the descriptive difference. If identity or geometry collapses, lower denoise or move to ControlNet, which is designed to preserve explicit structure.

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 realvisxl-img2img

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 5 of 10. Previous: Build text-to-image from a blank canvas Next: Use LoRAs without losing the base model

Rewire text-to-image

Add Load Image and VAE Encode. Feed the source pixels and the checkpoint VAE into VAE Encode, then connect its LATENT output to KSampler. Keep the text conditioning and decode/save tail.

Denoise is distance from the source

Denoise	Typical use
0.15–0.30	Texture cleanup, light retouch
0.35–0.55	Restyle while preserving layout
0.60–0.80	Strong reinterpretation
0.90–1.00	Source contributes little beyond broad structure

Run the lesson, do not just read it

The gallery records model sources, minimum VRAM, custom-node requirements, and verification state beside the graph. That is the portable unit for this series.