📑 CR Load Scheduled Models

Same idea as CR Load Scheduled LoRAs, one level up: instead of swapping which LoRA is active at a given frame, this swaps the entire checkpoint - model, CLIP, and VAE together - based on where you are in an animation. Use one checkpoint for the first third of a sequence and a different one for the rest, driven by a schedule instead of manually re-wiring loaders.

How it works

Two modes here (no Off - you always need some model to sample with): Load default Model just loads default_model and ignores any schedule, functioning like a normal checkpoint loader. Schedule reads current_frame against a connected schedule and switches to whichever checkpoint that frame maps to, using schedule_alias to identify the right track and schedule_format (CR or Deforum) to parse the keyframe syntax correctly.

The list of candidate checkpoints and their aliases comes from an optional model_list input - built with CR Model List, the pack's node for naming and aliasing a handful of checkpoints so the scheduler can reference them by name instead of by full path.

Inputs and outputs that matter

• mode - Load default Model / Schedule.
• current_frame - where you are in the animation (0–9999).
• schedule_alias - which named track in the schedule this node follows.
• default_model - the fallback (and the only checkpoint used) when mode is set to "Load default Model."
• schedule_format - CR or Deforum, needs to match how the schedule string was written.
• Optional: model_list (MODEL_LIST), schedule (SCHEDULE).
• Outputs: MODEL, CLIP, VAE, show_help - a full checkpoint's worth of outputs, same as a standard checkpoint loader.

How to install it

ComfyUI Manager: search "Comfyroll Studio", install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes.git, restart. Your checkpoints need to already be sitting in ComfyUI/models/checkpoints - this node just selects between ones you already have.

Common issues

Swapping full checkpoints mid-sequence is heavier than swapping a LoRA - each switch means ComfyUI has to load a new checkpoint into memory, which is slow and, depending on your hardware, can stall a long animation run noticeably at each transition frame. If you're scheduling frequent checkpoint swaps and generation grinds to a crawl at certain frames, that's expected behavior, not a bug - checkpoints are gigabytes, LoRAs are megabytes.

Two checkpoints from genuinely different architectures (say, an SD1.5 checkpoint scheduled against an SDXL one) will also break anything downstream that assumes a fixed latent size or CLIP dimensionality - the schedule node itself won't stop you from setting this up, so keep your scheduled checkpoints architecture-consistent unless you're deliberately rebuilding your latent/conditioning nodes to handle the switch too.

As always: schedule_alias needs to exactly match an alias defined in your model_list, and if the node itself doesn't show up after installing, check your ComfyUI startup console for an import error elsewhere in this large pack before troubleshooting this node in isolation.