EZ img2img

EZ Image - shown in the menu as EZ img2img - is the node you reach for when you have a source image and want a new image based on it. One node covers both plain img2img and masked inpainting; the difference is whether you feed it a mask.

Three inputs:

• image (IMAGE) - the source image.
• image_weight (FLOAT, default 0, range 0–1, step 0.1) - the honesty dial. The tooltip says it best: "Higher number means it closely matches the image, lower number means more random." This is the node's denoise control under another name: the pack samples at denoise = 1 - image_weight.
• mask (MASK, optional) - supply one and only the masked region changes. The tooltip: "Only generates inside of the masked region."

The output is IMAGE (EZ_IMAGE_SETTINGS), the settings bundle that carries the source image + weight + mask into EZ Generate - usually via EZ Crop and/or EZ Detail on the way.

The three modes

This node's behavior is really three modes hiding behind a dial:

Pure regeneration (weight 0). The source image is ignored except as a size reference - EZ Generate builds an empty latent at the image's dimensions and runs full txt2img-style denoise. Good for "same size, different content."

img2img (weight > 0, no mask). The source is VAE-encoded and partially denoised. Weight 0.5 keeps roughly half the original structure; weight 0.8 keeps almost everything and only nudges details. This is the classic "change the vibe, keep the subject" territory.

Inpaint (weight > 0, with mask). The masked region gets regenerated while the unmasked part is kept bit-identical - the pack uses InpaintModelConditioning and composites the result so only the mask changes. The mask is a shape; image_weight is still your denoise dial for how much of the masked pixels survive.

What masked inpainting is still for

It's worth being honest about the landscape here: instruction-editing models (Flux Kontext, Qwen-Image-Edit, Flux 2 Klein) took over most of what people used to inpaint, and they're often easier. But mask-based inpainting still uniquely owns bit-identical unmasked pixels, the denoise dial (a number, not a yes/no), and targeted detail work - the "fix the eye without touching the nose" case where an edit model would happily move both. On SDXL-lineage checkpoints, EZ Image + EZ Crop is a solid regional-inpaint setup, and paired with EZ Detail it's the classic face-fix pass: crop tight, low weight, done.

Installing it

It ships with the Prompt Helpers pack. In ComfyUI Manager search "Prompt Helpers" (Pauan/comfyui-prompt-helpers), install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Pauan/comfyui-prompt-helpers

Restart ComfyUI. No model downloads; the pack needs PyYAML and desktop-notifier and ComfyUI 0.8.0+.

The classic gotcha: forgetting the mask. People wire an image into EZ Image expecting inpainting, leave the mask socket empty, and get a full img2img rewrite of the whole frame. If "only the region changed" didn't happen, check that the mask is actually connected - and remember the mask does the work only when combined with EZ Generate's composite step, which it always is in the normal chain.