Image2Image Sampler v2 (Searge)

The 2 on the end tells you what this is: the second iteration of Searge's all-in-one image-to-image node. Same job as the v1 - take an image, encode it, run SDXL base plus refiner at a chosen denoise strength, optionally upscale, decode, done - with refinements the author made to the sampling internals in a later version of the pack.

From the outside the two are drop-in equivalent: they expose the identical set of inputs and outputs, so if you're swapping one for the other in a workflow, nothing needs rewiring. What changed lives under the hood, not in the interface. And like v1, this one sits in Searge/_deprecated_/Sampling - it's kept for backward compatibility, not what the current EVOLVED workflow reaches for. New builds should sample inside the Magic Box pipeline.

How it works

Same shape as v1. It takes a pixel image and a vae, encodes to a latent, samples with the base model then the refiner (split by base_ratio, default 0.8), and decodes back to an image - so it's end-to-end pixels-in, pixels-out, not the latent-in/latent-out pattern of the plain SDXL sampler. denoise (default 0.25) sets how far it moves from the source: low for polishing, high for reinvention.

The optional built-in hi-res pass is here too: feed an upscale_model and a target size and it runs a larger second stage. The VAE caveat carries over - each encode/decode cycle is slightly lossy, so stacking multiple low-denoise passes slowly softens detail. Fewer, more deliberate passes beat a long chain.

The inputs and outputs that matter

• image (IMAGE) + vae (VAE) - source and codec, required.
• denoise (default 0.25) - the change amount, and the main dial to tune.
• base_model/refiner_model and 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, scaled_width/scaled_height (default 1536), plus refiner_strength, softness, noise_offset.

Output is one decoded IMAGE.

Which one, v1 or v2?

If a workflow already uses one, leave it - they're interface-compatible and there's no reason to churn a working graph. If you're picking between them cold, take v2, since it's the later revision of the same node. If you're starting from scratch, honestly take neither: use the Magic Box pipeline the current workflow is built around, and keep these two in mind only for opening old Searge graphs.

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, put an ESRGAN model like 4x-UltraSharp in ComfyUI/models/upscale_models (the README's recommended set).

Common issues

Black images out of the decode are the SDXL fp16 issue - use sdxl-vae-fp16-fix on the vae input, which the README calls out specifically for this. Too-subtle or too-wild results are a denoise problem, not a bug: 0.25–0.5 is the workable range for most img2img. And it's deprecated, so treat it as legacy. If it errors after a ComfyUI update, update the pack before anything else.