WAN 2.2 Batch Sampler (CRT)

A proper Wan 2.2 workflow is more nodes than it has any right to be: two KSamplers (one per expert model, switching over at a sigma boundary), a seed loop if you want to compare a few generations, a VAE decode, maybe a grid compositor to lay the results out side by side, and a save node with its own naming scheme. This node collapses all of that into one box. It's the biggest single node in the CRT-Nodes pack's sampling category, and it earns the size - it's doing the work of six or seven stock nodes.

Why Wan needs two models in the first place

Wan 2.2 is a mixture-of-experts model: a high-noise expert handles early structure and motion, a low-noise expert takes over for detail, and the switch between them happens at a sigma value the community calls the boundary. This split is also why speed LoRAs behave weirdly on Wan 2.2 - the community's hard-learned lesson (from a landmark PSA thread that's still the standard advice) is to put your lightx2v or CausVid speed LoRA on the low-noise pass only and leave the high-noise pass clean, because stacking it on both destroys motion, lighting and emotion. This node's model_high_noise / model_low_noise split and separate cfg_high_noise / cfg_low_noise sliders are built directly around that architecture, and around that community practice.

The inputs and outputs that matter

There's a lot here, but a handful of fields do most of the work:

• model_high_noise / model_low_noise - the two Wan 2.2 expert checkpoints. This is the one thing people get wrong first: Wan 2.2 needs two models, not one, unlike Wan 2.1.
• boundary (default 0.5) and sigma_shift (default 1) - where the handoff between experts happens and how the noise schedule is shifted. Community defaults for full-quality generation sit around 20-30 steps at CFG 3.5+ on the high-noise pass; this node's defaults (cfg_high_noise/cfg_low_noise both at 1, steps at 8) are tuned for the fast, distilled-LoRA path instead, so if you're chasing maximum quality rather than speed, expect to raise steps and high-noise CFG yourself.
• positive (required) and optional negative - if you skip negative, an empty one is auto-generated.
• start_image (optional) - plug in a reference image and the node switches to image-to-video automatically; leave it empty for text-to-video.
• batch_count and increment_seed - generate several variations from one graph, each on seed+N if increment is on.
• processing_mode - Parallel (fast, high VRAM) versus Sequential (slow, low VRAM); offload_conditioning forces sequential mode and moves text conditioning to CPU between items for a bit more headroom.
• create_comparison_grid and save_videos_images with save_folder_path/save_subfolder_name/save_filename_prefix - the node will lay your batch out as a grid and write everything to disk itself, no separate save node needed.

Outputs cover the whole pipeline: high_noise_latent_batch, final_latent_batch, final_images_batch, comparison_grid, and a settings_string for logging exactly what you ran.

How to install it

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

Common issues & troubleshooting

One model where two are expected. If you wire the same checkpoint into both model_high_noise and model_low_noise, it'll run, but you're not getting Wan 2.2's actual architecture - make sure you've downloaded both experts.

Output looks flat or over-plasticky. That's the classic symptom of a speed LoRA applied to both passes. Check lora_batch_config in the accompanying LoRA-select node (if you're using one) and keep the high-noise pass clean.

Parallel mode runs out of VRAM. Switch processing_mode to Sequential and, if it's still tight, enable offload_conditioning.

Frame count beyond 81 looks strange. Wan's native context window is 81 frames - pushing well past that without a dedicated extension technique tends to loop or degrade rather than smoothly continue.