XY Inputs: Diffusion Model //EasyUse

Choosing between five Anima fine-tunes or the Raw-vs-Turbo Krea 2 split by generating each one separately and eyeballing them side by side is the slowest part of the new-model era. This node exists to kill that. easy XYInputs: DiffusionModel (display name "XY Inputs: Diffusion Model //EasyUse") is one of the axis-generator nodes that feed Easy-Use's XY plot - the X or Y value being "which diffusion model," so one run renders a labeled grid of however many Anima/Krea 2 files you want to compare on the same prompt and settings.

Older SDXL checkpoints were self-contained, so model comparisons ran on the Checkpoint axis. Anima and Krea 2 split into a diffusion model plus a separate text encoder and VAE, so Easy-Use added this dedicated DiffusionModel axis that carries those two extra choices per slot. It landed in the same update as easy diffusionModelLoader ("Add Anima and Krea2 diffusion model support"), and the two are meant to work together.

How it works

The node builds a small dataset that says, essentially, "axis = DiffusionModel, and here are N entries of model,clip,vae." Each slot you fill becomes one grid value. When Easy-Use's XY plot engine walks the grid, it loads the entry for the current cell from models/diffusion_models, reads which family the weights belong to, pulls in that slot's text encoder and VAE, re-encodes your prompt, and stamps a fresh latent at the right shape - then samples that cell. Result: the same seed, prompt, sampler and steps run against every model in the list, with each cell's label showing which model produced it. The backend handling only knows two families right now - anima and krea2 - which is exactly who this node is for.

The inputs that matter

The interesting wrinkle is the defaults. Each of the up-to-ten slots takes a triplet:

• model_name_N - your pick from models/diffusion_models. Default "None" means that slot is skipped, so set only as many as you care about.
• clip_name_N - text encoder, default "Auto". This is the smart part: leave it on Auto and the node picks the family's expected encoder for you - Anima expects a Qwen3-0.6B Base encoder, Krea 2 expects the Qwen3-VL build - by looking at what model you loaded.
• vae_name_N - default "Auto", which resolves to the Qwen-Image VAE for both families.

And at the top, model_count (1–10, default 3) decides how many of the slots actually run.

The single output, X or Y (type X_Y), plugs straight into the X or Y socket of Easy-Use's XY plot node (easy XYPlotAdvanced, or the pipe-based XY Plot), alongside whatever axis you want on the other dimension - sampler, CFG, seed, prompts. One slot's "Auto" does the family lookup at load time, so you can mix an Anima and a Krea 2 file in the same list and each gets the right encoder.

Install

Same pack as all the other Easy-Use nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

or search "ComfyUI Easy Use" in ComfyUI Manager. Restart ComfyUI and run install.bat (Windows) / ./install.sh (Linux/Mac) once to install the pack's requirements - though this node itself leans on nothing from that heavy list.

The model files are the real install. For every model you grid you need its weights in models/diffusion_models/, its text encoder in models/text_encoders/, and the Qwen-Image VAE in models/vae/. The gotcha with "Auto" is that it resolves to a specific filename - Anima's auto encoder is qwen_3_06b_base.safetensors, Krea 2's is the Huihui fp8-scaled Qwen3-VL build. If you don't own that exact file, that slot errors at runtime. Set the dropdown to your actual file instead of Auto and the node happily uses it.

Where people get burned

• Expecting it to handle SDXL checkpoints. It won't - for classic self-contained checkpoints use easy XYInputs: Checkpoint, a different axis for a different file type.
• The "Auto" text encoder isn't installed. The most common red node in practice. Download the expected encoder file or override the slot.
• Dropping model files in the wrong folder. They won't appear in model_name_N unless they're in models/diffusion_models.
• Gridding ten models when you meant three. model_count defaults to 3 but the slots go to 10 - the plot runs whatever you set the count to, so check it before you queue a long anime comparison and walk away.

A practical note from the model side: Anima comparisons are slow - it's a 2B model that wants 30–50 steps, roughly 10s per image on a 5090 and over a minute on a 4060. A 3×3 grid is a coffee break, not a minute, which is exactly why you want one unattended run instead of nine babysat generations. When a winner emerges, save its triplet - easy diffusionModelLoader will reproduce that exact model + text encoder + VAE combo for your real workflow.