5 Practical img2img Use Cases in ComfyUI

Introduction

So you’ve generated an image and it’s… almost there. Maybe the face is slightly off, or there’s a random floating hand in the background, or the composition cuts off right where it shouldn’t. We’ve all been there.

That’s where img2img comes in. Instead of rolling the dice on a brand new generation, you take what you already have and refine it. I’ve been experimenting with img2img workflows in ComfyUI for a while now, and I’ve landed on five techniques that I keep coming back to. They’re not fancy — most are surprisingly simple to set up — but they consistently save me from the “generate, sigh, regenerate” loop.

Let’s walk through each one and see what they can actually do.

1. The Chase Sampler (img2img Re-Sampling)

This one is almost embarrassingly simple. You load your generated image back into a KSampler and run it through another round of sampling. That’s it. I call it the “chase sampler” because you’re essentially chasing down a better version of what you already have.

How it works: Load your image, encode it through the VAE, and feed the latent into a KSampler. Set the denoise to around 0.45 with about 25 steps.

The key decision: with or without prompts?

Here’s where it gets interesting. You can run the chase sampler two ways:

• Without prompts or LoRAs — The sampler works purely from the image data. It smooths things out and adds detail, but colors can flatten and things like eye color might shift unexpectedly.
• With the original prompts + LoRAs — Feeding in the same text and LoRA configuration from the original generation keeps the character anchored. Eye color stays consistent, skin tones hold, and the overall look stays truer to what you intended.

My recommendation? If you want to preserve the character’s identity, always include the prompt information. If you’re going for a more artistic reinterpretation — maybe shifting the art style or mood — going promptless can produce some happy accidents.

When you’d use it: Anytime a generation looks “80% there” but needs that extra polish. It’s my most-used img2img technique by far.

2. Inpainting — Erasing Things That Shouldn’t Be There

You know those random artifacts that show up? A weird extra finger, a floating accessory that doesn’t belong, some bizarre background element the model hallucinated? Inpainting handles this.

How it works:

1. Load your image and right-click to open the Mask Editor
2. Paint over the area you want to fix
3. Connect the masked image to a VAE Encode (Inpaint) node, then feed that into a KSampler
4. Set denoise to 1.0 — you want full regeneration in the masked area

The trick I learned the hard way: After inpainting, the fixed area often looks obviously patched — like a bad Photoshop job. The solution? Follow up with a chase sampler (technique #1) at a low denoise of about 0.25. This blends everything together so the edit looks seamless.

Having prompt information makes a real difference here too. The mask affects the surrounding area slightly, and without prompts to anchor things, you can get unexpected color shifts bleeding outward from the repair.

When you’d use it: Removing artifacts, fixing weird hands, cleaning up backgrounds, erasing unwanted elements. It’s your digital eraser with AI-powered fill.

3. Outpainting — Extending the Canvas

AI-generated images love cutting things off. Arms cropped at the edge, compositions that feel cramped, backgrounds that end abruptly. Outpainting lets you extend the canvas and have the AI fill in what’s missing.

How it works: Use ComfyUI’s built-in “Pad Image for Outpainting” node to add pixels to any edge — left, right, top, bottom. This padded area becomes an automatic inpaint mask. Connect it to VAE Encode (Inpaint) and run through a KSampler.

Fair warning: this one’s a bit of a wild horse. The AI has to guess what belongs in the extended area, and it doesn’t always guess right. A few things I’ve noticed:

• White backgrounds can confuse it — the AI might add random background elements where you wanted blank space
• Adding “simple white background” to the prompt can fix the background but distort body anatomy
• Best results come from running the outpaint first, then cleaning up with a chase sampler, and removing any unwanted background afterward

I suspect combining this with IPAdapter (feeding it a reference of what the extended area should look like) would improve accuracy significantly. That’s on my list to test.

When you’d use it: Whenever a composition feels too tight or body parts are awkwardly cropped. Especially useful for turning portrait crops into half-body or full-body shots.

4. Face Detailer — Targeted Facial Refinement

Sometimes the body looks great but the face is just… not there yet. The FaceDetailer node is purpose-built for this — it automatically detects faces in your image and re-generates just that region.

How it works: Feed your image into the FaceDetailer node. It uses a detection model (like a face recognition network) to find faces, crops them out, regenerates them at higher quality, and composites them back in. You can adjust denoise and add face-specific prompts to steer the result.

Honestly, this one surprised me with how effective it is. I was skeptical at first — how much difference can re-doing just the face make? Turns out: a lot. It’s my “last line of defense” when everything else looks good but the face is letting the image down. The quality bump is immediately noticeable, especially for close-up and mid-range shots.

When you’d use it: Final polish pass on any image where the face needs to be sharper, more detailed, or more expressive. Works great in combination with the other techniques — generate, inpaint any artifacts, then run FaceDetailer for that final 10% quality boost. If you need precise control over your character’s pose before running these refinement techniques, ControlNet lets you define exactly how the character is positioned.

5. Hires Fix (img2img Style)

This is classic hires fix, but applied as an img2img operation rather than as part of the txt2img pipeline. The image gets upscaled first, then re-sampled — so you get both higher resolution and more detail in a single pass.

How it works: Upscale the image (using any upscale model), then run it through a KSampler at denoise 0.35 or so.

With prompts vs. without:

This is where prompt information really matters. At the upscaled resolution, the model has more pixels to work with and more room to “interpret.” Without prompts, it starts making assumptions — eye color might shift, hair color might drift, facial features can change. If you have the original prompt and LoRA information, include it. At minimum, include descriptive details about the character’s appearance (eye color, hair color, etc.) to keep things anchored.

The quality improvement over the basic chase sampler is significant because you’re working at higher resolution. Details that were blurry or undefined at the original size suddenly become crisp and clear. For a deeper dive into upscaling techniques and choosing the right upscaler model, see our post-upscaling detail enhancement guide.

When you’d use it: When you need a high-resolution final output and the base generation is solid but lacks fine detail. It’s the “premium” version of the chase sampler.

How These Techniques Power AI Companion Platforms

If you’ve ever wondered how AI companion platforms keep their characters looking polished and consistent, these img2img techniques are a big part of the answer.

Avatar refinement: When platforms like Candy.AI or YUKIKO.AI generate character images, the raw output from txt2img is rarely the final product. Chase samplers and hires fix passes clean up artifacts and boost detail quality — the same techniques you’d use in your own workflow.

Expression variations: Generating the same character in different emotional states (happy, thoughtful, surprised) often requires inpainting or FaceDetailer passes to get expressions looking natural. The face is the hardest part to get right, and these targeted refinement tools are what make it practical at scale.

Outfit and scene changes: When you change your companion’s outfit or scene, platforms often use img2img techniques to blend the new elements with the existing character identity. Inpainting handles the clothing swap while preserving the character’s face and pose.

Composition expansion: Outpainting is how some platforms generate full-body shots from originally cropped compositions, or extend scene backgrounds for more immersive visuals.

Understanding these techniques gives you insight into the production pipeline behind the companions you interact with — and if you’re creating your own characters, these five methods will immediately level up your results. For a deeper look at the model types involved, check out our guide to ComfyUI model types.

Credits & Source

This guide is adapted from nobin’s original Japanese article on note.com/nobinlog. We’ve translated and recreated his workflows to share these techniques with an English-speaking audience. nobin has a knack for finding the practical, everyday uses of these tools — the stuff you’ll actually reach for during a session rather than techniques that only work in theory.

Check out his original post for the full visual examples and workflow screenshots: ComfyUIでimg2imgを活かす活用アイデア５選.