EA 4-Way Image Compare

The 4-Way compare is the pack's answer to "how strong should this LoRA actually be?" Base, 0.5, 1.0, 1.5 - four panels, one image, and the sweet spot jumps out at you instead of being a memory exercise. It's the same compositor as the 2-way and 3-way siblings, just scaled to four slots, and for weight sweeps it's the one you'll actually reach for.

The inputs

Four image slots and four captions, same shape as the smaller versions:

• image_1 … image_4 - IMAGE tensors. Default captions are Base Model, LoRA 0.5, LoRA 1.0, LoRA 1.5, which is a reasonable first sweep for most character and style LoRAs. You'll relabel them for whatever your actual axis is - sampler, CFG, seed, denoise strength.
• scale (0.1–2.0) resizes all four proportionally. Four full-res images is a lot of pixels; this is the knob that keeps the composite viewable.
• font_size, caption_height, font_color, background_color, spacing - caption bar and gutter styling. font (optional) uses a TTF from comfyui-kjnodes/fonts if present, else PIL's default font.

Same trap as the rest of the family: all four images must be the same dimensions or you get a ValueError. If your sweep produced mixed resolutions, normalize upstream first - scale only shrinks proportionally, it can't reconcile two different aspect ratios.

What comes out

One comparison IMAGE: the four captioned panels tiled left to right with a few pixels of spacing. Preview it, save it, post it - the whole sweep lives in a single image, which is exactly what you want when you're iterating weights and comparing runs that are a hundred generations apart.

One practical note: because four panels get wide fast, you'll usually drop scale below 1.0 while comparing, then save the full-res version for records. The composite honors whatever scale you set, so pick a value that fits your monitor for viewing and your gallery for sharing - it's the same output.

Install

Ships in comfyui-ea-nodes. ComfyUI Manager → search comfyui-ea-nodes, or git clone https://github.com/ExoticArts/comfyui-ea-nodes into custom_nodes/, restart, hard-refresh. No extra Python deps - the pack's requirements.txt is empty.

If four panels is more than your typical test needs, the 3-way (with the base + two weights) is the more economical default; the 4-way earns its extra slot when you're chasing the boundary where a LoRA tips from "nice" to "too much."