WAN 2.2 LoRA Compare Sampler (CRT)

There's a specific mistake that shows up constantly in AI-video and AI-image circles: someone posts a glowing verdict on a new model or LoRA, and the actual explanation turns out to be that they never ran a real side-by-side. Same prompt, same seed, different settings, laid out next to each other - it sounds obvious, and almost nobody does it, because building that comparison by hand means running the same generation over and over and manually assembling the results. This node automates exactly that discipline for Wan 2.2 LoRAs.

What it does

You give it your two Wan 2.2 expert models, a prompt, and a lora_batch_config - a text block listing multiple LoRA configurations to test. It runs the same seed and prompt through each one, generates a labeled comparison grid automatically (add_labels, custom_labels, label_font_size control exactly how), and hands you back the individual results alongside the grid. This is a bigger, more specialized sibling of the pack's general-purpose WAN 2.2 Batch Sampler - where that node batches seeds, this one batches LoRA configurations for direct comparison.

The default settings tell you what it's tuned for: 8 steps and a 0.875 boundary (versus the more conservative 0.5 default elsewhere in the pack) point toward the fast, distilled-LoRA workflow rather than full-quality 20-30 step generation - reasonable, since you're likely running this multiple times per comparison and speed matters. Frame count defaults to 81, which is Wan's native context ceiling; push meaningfully past it and you're relying on an extension technique this node doesn't provide.

The inputs and outputs that matter

• model_high_noise / model_low_noise, positive / negative, width / height / frame_count, steps, sampler_name / scheduler - the standard Wan 2.2 sampling set.
• lora_batch_config (multiline STRING) - the LoRA configs under test, one comparison run per entry.
• boundary (default 0.875) and sigma_shift (default 8) - both noticeably different from the pack's other Wan sampler, worth tuning if your results look off from what you'd expect at more conservative values.
• create_comparison_grid, add_labels, custom_labels, label_font_size - the labeling controls that make the grid actually legible as a comparison rather than just a strip of unlabeled images.
• Optional i2v_image for image-to-video mode, optional vae and clip.

Outputs cover both the process and the result: high_noise_latent_batch, final_latent_batch, final_images_batch, comparison_grid, settings_string, and - notably - model_high_noise_out / model_low_noise_out, passing the LoRA-patched models back out so you can chain straight into further nodes without reloading.

How to install it

Part of CRT-Nodes. Via ComfyUI Manager, search CRT-Nodes and install. Manually: git clone https://github.com/PGCRT/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart.

Common issues & troubleshooting

Run takes much longer than a single generation. That's expected - you're generating once per entry in lora_batch_config, not once total. If it's too slow, trim the config down to the LoRAs you actually need to compare rather than testing everything at once.

Grid labels overlap or are illegible. Adjust label_font_size down for a large grid, or check custom_labels isn't mismatched in count against the number of configs in lora_batch_config.

Results at boundary 0.875 look different from other Wan 2.2 workflows you've run. That's the point - this node's defaults lean toward the fast/distilled path, not full-quality settings. If you want a comparison closer to a full-quality baseline, move boundary back toward 0.5 and raise steps accordingly.