Image2Image Sampler v1 (Searge)

This is Searge SDXL's original all-in-one image-to-image node: hand it an image and it encodes it, runs SDXL base plus refiner over it at a denoise strength you set, optionally upscales, and hands you back a finished image. It rolled the encode, the two-stage sample, an optional hi-res upscale pass and the VAE decode into one box, which was a lot to have in a single node in 2023.

Read the category first, though - Searge/_deprecated_/Sampling. This is a v1 node, kept for backward compatibility with old workflows but not what the current EVOLVED workflow uses. There's also a v2 with the same interface. If you're building something new, the modern path is the Magic Box pipeline; this node is here for graphs that already depend on it.

How it works

Unlike the plain SDXL sampler, which takes a latent and returns a latent, this one is end-to-end on pixels. It takes an image and a vae, encodes the image to a latent, samples with the base then the refiner (split by base_ratio, default 0.8), and decodes back to an image for you. The amount it changes the input is set by denoise - the default 0.25 is a light touch that keeps most of the original composition, which is the usual img2img sweet spot for polishing rather than reinventing.

It also has an optional upscale stage built in: give it an upscale_model and target scaled_width/scaled_height, and it'll do a larger second pass - Searge's version of a hi-res fix, in the same node. One thing to keep in mind from how VAEs work: every encode/decode round trip is slightly lossy, so chaining several of these img2img passes at low denoise will gradually grind detail down. Do the work in as few passes as you can.

The inputs and outputs that matter

• image (IMAGE) and vae (VAE) - the source and the codec. Required.
• denoise (default 0.25) - how much to change the image. The knob you'll actually tune: low keeps the original, high reinvents it.
• base_model/refiner_model plus their four CONDITIONING inputs - the two-stage models.
• base_ratio (0.8), steps (20), cfg (7), sampler_name (ddim), scheduler (ddim_uniform) - sampling controls.
• Optional hi-res: upscale_model (UPSCALE_MODEL), scaled_width/scaled_height (default 1536), plus refiner_strength, softness and noise_offset for fine control of the second pass.

Output is a single IMAGE, already decoded.

How to install it

ComfyUI Manager: search SeargeSDXL, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

then restart. Manual installers need python -m pip install opencv-python run once. For the upscale pass, the README recommends ESRGAN models like 4x-UltraSharp and 4x_NMKD-Siax_200k in ComfyUI/models/upscale_models.

Common issues

Black output on the decode is the classic SDXL-in-fp16 failure - use the sdxl-vae-fp16-fix VAE (the README lists it exactly for this) on the vae input. If your img2img barely changes the picture, your denoise is too low; if it ignores the source entirely, it's too high - 0.25 to 0.5 is the useful band for most polishing. And remember it's deprecated: it works, but new workflows should sample in the Magic Box pipeline. If it throws after a ComfyUI update, update the pack first.