GiusTex-ComfyUI-MoreEfficientSamplers/README.md

GTNodes for ComfyUI

A bundle of nodes with different purposes. More info below. Some of those nodes (SamplerCustomAdvancedEfficient, MultiLoraLoader, ...) exist as an alternative to comfyui's subgraphs, where sometimes the widget values are not passed correctly inside. Some nodes use parts of other auhors nodes, more info below/in the credits.

Updates:

• 10/11/2026: Converted H3 Tiled Sampler Custom Advanced into H3 Tiled Sampler Custom Advanced (Efficient).
• 8/11/2026: Added taeh3 preview method and H3 Tiled Sampler Custom Advanced node.
• 31/08/2026: Added MultiLoraLoader and MultiLoraLoaderAdvanced.
• 26/08/2026:
  • Updated SamplerCustomAdvancedEfficient: model input is now optional. Added audio decode option. Added option to disable cfg and negative inputs to slightly slim down the ui.
  • Linked ScheduledCfgGuiderClass to CFGGuider, and removed now redundant functions.
• Updated SamplerCustomAdvancedEfficient to work through comfyui Sampler Custom Advanced to support LTXV combined latent and future comfyui updates.
• Added ScheduledCfgGuider, CFGFloatListScheduler (taken from my Wan-TimeToMove) to support higher cfg values at first steps without needing Wan-TimeToMove.
• Removed redundant nodes.

Download

To install ComfyUI-GTNodes, follow these steps:

• Go in the ComfyUI custom_nodes folder, then download the repository or clone it here: git clone https://github.com/GiusTex/ComfyUI-GTNodes.
• Optional: install ComfyUI-SigmaSync if you want to use Multi Lora Loader Advanced.
• Optional: download kijai's taeh3 vae in models/vae_approx to use taeh3 single image preview method;
• Optional: install Kosinkadink/ComfyUI-VideoHelperSuite (+ kijai's taeh3 vae) to animate H3 video previews;
• Restart ComfyUI.

Nodes:

Multi Lora Loader

• A simple wrapper around comfyui's Load Lora node to load multiple LoRAs.
• Warning: changing number of loras resets the values, I suggest to start with 3-4 loras and leave None in the unused slots.
• As mentioned above, widgets with None are skipped.

Multi Lora Loader Advanced

• A wrapper around Multi Lora Loader and a simpler version of ComfyUI-SigmaSync-LoRA, to specify just when to start or end the use of each lora. As you may imagine, the node requires ComfyUI-SigmaSync.
• Leave end_at_step: -1 to apply the lora's strength until the last step.
• The sigmas required are the same ones you pass to SamplerCustomAdvanced.
• As for Multi Lora Loader, widgets with None are skipped, and changing number of loras resets the values. I suggest to start with number of loras: 3 and leave None in the unused slots.

How it works:

• Instead of specifying different strengths at different steps or setting manually their list manually, the node forces the explicit curve and builds the strengths_list with a list of zeros, then it applies the lora's strength at the specified steps.
• If no sigmas are provided the node falls back to the Multi Lora Loader node, otherwise it applies each strenght at specified steps to the corresponding loras.

Sampler Custom Advanced (Efficient):

An edit of efficiency-nodes's ksampler adv. (eff.) to let it accept custom samplers, schedulers (sigmas) and guiders through comfyui Sampler Custom Advanced and other native nodes. Supports:

• internal slicing of the sigmas, useful for multi-stage sampling;

• use of only positive, useful when sampling with cfg 1 and/or the negative isn't required;

• use of optional custom samplers, schedulers and guiders. Connect and use them, or bypass/leave their input slots empty to use the corresponding default widgets. Default guiders: CFG Guider when use_cfg is True, Basic Guider when False.

• Decoding of images/audio, when one or both are available (the node auto-checks what is available in the latent).

• Embedded preview for MiniMax H3, using taeh3 preview method. Max_preview_resolution is hardcoded at 1024 (instead of 512 like in default ComfyUI, anyway only greater inputs will be resized).

  Requires kijai's taeh3 vae in models/vae_approx.

  Requires an H3 latent dict (image/video + audio), and works both with a single image or a video. For the single image you can use P-H3-Image Conditioning node, that returns a latent dict of image + audio, then their H3 image vae or another custom image vae (link in "Suggested resources").

  For an animated video preview, Kosinkadink/ComfyUI-VideoHelperSuite is also required, then from ComfyUI's Settings window -> 🎥🅥🅗🅢: toggle "Display animated previews when sampling" ON. You can also change video preview's fps here.

  Decoding the preview stops the sampler for some seconds. Decode tests (on a rtx 3060 12gb):

  • video of 8 seconds, size of 448x704: the sampler stopped each step for 1.3 seconds;
  • video of 8 seconds, size of 576x832: the sampler stopped each step for 2.1 seconds.

H3 Tiled Sampler Custom Advanced (Efficient):

An edit of was-node-suite-comfyui's H3 Tiled Sampler, to let it accept custom sigmas, useful if you have lightx loras that require custom sigmas, like MiniMax H3 HyperFlow model/lora.

Requires an H3 latent dict (image/video + audio), and works both with a single image or a video, check Sampler Custom Advanced (Efficient) for more info.

To use taeh3 preview method, as said in the previous node, kijai's taeh3 vae is required in models/vae_approx; more info on this and turning on animated video preview above.

This tiled sampler is useful if you upscaled an image/video, for example with the ComfyUI's native MiniMax H3 Latent Upscale With Model node (example upscale model in "Suggested resources"), the sampler then returns a dict, so you have to use Separate AV Latent node to get the video.

Scheduled CFG Guider (with its CFG Float List Scheduler):

(Deprecated, although they still work; I advice to use comfyui's native CFG Override). Scheduled CFG Guider, anyway, let you decide what cfg to apply at each step through a list of floats connected to the cfg input. Based on kijai nodes.

Suggested resources:

• flowmatch scheduler: useful with video models using lightx loras;
• P-H3-Image-Comfy's image vae;
• custom MiniMax H3 image vae;
• minimax_h3_latent_upscaler_3d.

Credits:

• jags111/efficiency-nodes-comfyui, whose KSampler Adv (Efficient) was the base for the Sampler Custom Advanced (Efficient);
• WASasquatch/was-node-suite-comfyui, for his H3TiledSampler and his 3 modules (h3_low_vram.py, h3_tiles.py, preview.py) which still fills most of the H3 Tiled Sampler Custom Advanced;
• jlucasmcrell/ComfyUI-H3-Multishot, whose _build_decoder is used to build the taeh3 decoder;
• Kosinkadink/ComfyUI-VideoHelperSuite, whose decode_latent_to_preview_image and process_previews are patched from (link);
• Claude for helping creating and debugging the code.