Sexiam - Img2img with Upscale

Img2Img Workflow (ComfyUI)

This workflow lets you take a base image and combine it with a secondary image (like a nebula, fractal, or other abstract texture) to generate new compositions that text-to-image alone wouldn’t produce. Instead of only relying on prompts, this setup uses the “driving image” to push the model into unexpected poses, angles, and layouts, while still respecting your text prompt and style settings.

This is especially useful when you want:

More dynamic posing (without manually inpainting or posing tools).
Unique layouts and camera angles.
A way to push your characters or scenes into new territory, while staying consistent.

What It’s For

Generating new compositions from an existing base image.
Keeping your characters consistent while experimenting with different layouts.
Using abstract images (nebulae, textures, fractals, clouds, etc.) as latent drivers for creativity.

High-Res Fix and Upscaling are built-in — this ensures the outputs stay sharp, detailed, and production-ready.

Step-by-Step Breakdown

1. Load Your Driver Image

Start by loading the image you want to use as a guide — this could be a nebula, a fractal, a painting, or anything with interesting shape and color flow. This image will shape how your final result is composed.

Nodes will automatically take the width and height of your input image and apply it to the generation.

For best results, a standard SDXL size is recommended for input images with img2img. Be sure to crop your image to around 1 million pixels. The following sizes can be used as a guide:

1:1 - 1024×1024 - Square

3:2 - 1216×832 - Landscape

4:3 - 1152×896 - Landscape

16:9 - 1344×768 - Widescreen

9:16 - 768×1344 - Vertical

2:3 - 832×1216 - Portrait

3. Add Your Prompt

Now it’s time to describe what you want to see — your subject, the style, the lighting, etc. You can also write a negative prompt to avoid things like blurry faces, extra limbs, or text.

The SDXL base model and VAE are loaded, and LoRAs can be added if you want to steer the style further.

4. Sampling and Upscaling – Build and Refine the Image

This step handles both the base composition and the detail refinement. It’s made up of two sampling passes and a built-in upscale between them.

In the first sampling pass, the KSampler uses your latent input and prompt to create a strong composition. A denoise setting around 0.6–0.7 is used here — that’s strong enough to change the image significantly, but still preserve the flow and layout.

Next, the image is upscaled by 50%. This isn’t just a resize — the model intelligently adds detail using High-Res Fix. It sharpens the image and gives it more pixels to work with.

Finally, a second KSampler refines the image using a lower denoise value (around 0.4). This keeps the structure but adds texture and small fixes.

Labels and Prompt Template

This workflow is fully labeled for ease of use. Each section includes notes explaining what the node does, how it works, or what settings to change. This makes it easier to troubleshoot or customize without needing to reverse-engineer anything.

A prompt template is also included in the workflow to help guide your text inputs. It covers subject, mood, lighting, and style, giving you a strong starting point for consistent and well-structured prompts.