ComfyUI-PlagueKind-Nodes Introduction
ComfyUI-PlagueKind-Nodes is a collection of ComfyUI nodes for image editing, reference handling, LoRA stacking, and MiniMax H3 image and video workflows. It brings several common tasks into fewer nodes: keeping an image and its mask aligned while resizing, drawing a crop directly on a node, organizing many media references, and enlarging a finished H3 clip.
You do not need to use every part of the collection. The image, crop, loader, and bundle nodes can be useful on their own. The MiniMax H3 nodes are for workflows using ComfyUI’s native H3 support. Optional features are designed to warn or fall back when their requirements are unavailable rather than prevent the whole collection from loading.
The current version identified in the provided documentation is 1.6.4.
How ComfyUI-PlagueKind-Nodes Works
A ComfyUI workflow passes images, masks, models, and other data from one node to the next. This collection adds nodes that keep related parts of that workflow together:
- Images and masks share a transformation. Think of an image and its mask as two transparent sheets that must stay lined up. Unified Resize applies the same sizing and cropping geometry to both, helping a selected area remain in the right place.
- A visual crop replaces manual coordinate math. You draw a box over a preview, then adjust its corners or aspect ratio. The crop is stored as a relative position, so it can remain in a similar place if the source image changes size.
- Loaders and bundles organize inputs. A media loader collects image and video files; bundle nodes carry multiple connections through one wire. This is useful when a reference-heavy workflow would otherwise become difficult to read.
- H3 reference nodes prepare conditioning. These nodes arrange prompts, keyframes, reference media, and optional RefMods into inputs for MiniMax H3 generation. Reference images can be shown to the text-vision encoder and, depending on your settings, encoded as visual references for generation.
- Model patches change how sampling runs. Sparse attention skips some attention work, caching can reuse intermediate work, and the AdaLN fix handles a particular LoRA compatibility issue. They affect the generation process rather than editing finished frames.
- The video upsampler works on a finished clip. It enlarges the clip in latent space, lightly re-draws detail in overlapping pieces, and joins the results. The overlap helps avoid obvious seams between processed sections.
ComfyUI-PlagueKind-Nodes Features
Unified Resize Image / Mask
Unified Resize Image / Mask (PlagueKind/image) outputs an image, mask, width, and height. Use it when an image and mask must undergo matching changes—for example, before inpainting or compositing.
You can choose a target by width and height, a multiplier, the longer or shorter side, or a total megapixel count. Aspect-ratio presets include square, portrait, landscape, and widescreen shapes; presets do not apply in Dimensions mode. You can also choose an upscale method, enable or disable center cropping, and require dimensions to be divisible by a chosen number. Set divisible by to 1 to disable that constraint.
If no image is connected, the node can serve as a resolution selector for text-to-image or text-to-video workflows.
Visual Crop nodes
Both crop nodes appear under PlagueKind/image:
- Visual Crop (BBox) crops an image or mask without resizing it. It also outputs the cropped dimensions and the crop’s original
xandyposition, which can help when placing the result back onto the source. - Visual Crop + Resize (BBox) crops first, then applies the Unified Resize options in the same node.
After running the node to obtain a preview, you can drag and resize a crop box directly on it. Choose Free, a preset ratio such as 1:1 or 16:9, or a custom numeric ratio. A numeric crop-box override is also available if you need precise adjustments.
LoRA Loader Stack
LoRA Loader Stack ( LTX / MiniMax H3 Compatible ) (PlagueKind/loaders) provides up to 10 LoRA slots. You can enable slots individually, adjust their strengths, reorder them by dragging, and find files through a searchable folder browser. The stack is saved with the ComfyUI workflow.
Its modes control how a slot’s strength is applied:
normalapplies one strength to the LoRA.ltxoffers separate video and audio branch control for LTX workflows.minimaxoffers control for joint, video, audio, and text-related MiniMax H3 weights.
The node reports key counts for each of those H3 groups, helping you inspect a LoRA before deciding how to apply it. It can also function as a standard loader for compatible models outside these specialized workflows.
RefMod Loader Stack
RefMod Loader Stack ( MiniMax H3 RefMods ) (PlagueKind/loaders) organizes up to eight H3 RefMods. Each slot has a mod choice, strength, copy count, component filter (All, Visual, or Audio), and separate visual and audio strengths. You can set an optional total-token budget; 0 disables that limit.
This node uses the RefMod loading behavior supplied by ComfyUI-MiniMaxH3Mod rather than replacing it. It outputs an H3_REF_MODS bundle and a prompt_hint. If that other node collection is unavailable, this node returns an empty bundle and prints a warning instead of stopping the workflow.
Media Loader and bundle nodes
PK Media Loader ( 9 Images + 3 Videos ) (PlagueKind/loaders) puts nine image slots and three video slots in one tabbed node with previews. Each video has an audio toggle, on by default. Empty slots are skipped, and an unreadable file produces a warning rather than stopping the run. The node outputs a bundle containing the selected media and a separate image_1 output for workflows that need a direct image connection.
Bundle In (PlagueKind) and Bundle Out (PlagueKind) (PlagueKind/utils) carry up to 20 mixed-type connections through a single bundle wire. Their slots match by position: input_1 is retrieved as output_1, and so on. This is particularly helpful when routing many references across a large canvas.
MiniMax H3 prompt and reference nodes
PK MiniMax H3 Combined Reference to Video (PlagueKind/minimax) prepares conditioning and an empty latent for video generation. In one node, you can use first and last keyframes alongside reference images, reference videos with paired soundtracks, standalone reference audio, a media bundle, and a RefMod bundle. It also outputs a reference map explaining what each generated <Picture n>, <Video n>, or <Audio n> tag refers to.
Reference settings let you balance fidelity and workload:
ref_image_size: matchkeeps each reference’s aspect ratio and scales it down, if needed, toward the generation’s pixel area.ref_image_size: maxuses the reference pipeline’s larger presentation for stronger identity fidelity, but can make sampling several times slower.ref_image_size: nonepresents references to the text-vision encoder without adding a VAE-encoded visual reference block.vlm_resolutionandvlm_video_resolutionseparately control how stills and sampled video frames are presented to the text-vision encoder. A value of0follows the encoded reference resolution.reference_fpscontrols the playback rate used to reconstruct RefMod video timestamps;max_total_tokenscan reject RefMod bundles above your chosen limit, with0meaning no limit.
PK MiniMax H3 Image (PlagueKind/minimax) is the still-image counterpart. It accepts reference images and RefMods, and outputs conditioning, an empty latent, and a reference map. Setting width or height to 0 uses the first reference image’s scaled size where applicable; without images, the stated fallback is 2912 × 2912. Like the native H3 image node, it requires ModelSamplingMiniMaxH3 set to 1.0/1.0.
Both nodes rely on a ComfyUI build with native MiniMax H3 support.
MiniMax H3 model patches
H3 SLA Attention (PlagueKind/model_patches/minimax) uses block-sparse attention: it checks which groups of tokens appear most relevant, then spends full attention work on a selected fraction. No separate sparse-attention model is loaded; sparsity is chosen at runtime. An SLA LoRA can help the model tolerate that change.
Its main controls include sparsity_ratio (how much is skipped), block_size (how many tokens share a selection), and min_seq_len (shorter sequences remain dense). You can keep early or final sampling steps dense, protect audio or reference regions, and choose an attention engine. The default block size is 32; larger blocks may be faster but can harm speech quality. The node includes an enabled bypass for comparing a dense run without rewiring.
The author reports approximately 44 seconds per iteration dense, versus 31 seconds at 0.85 sparsity and 25 seconds at 0.90, measured on an RTX 5090 at 768p for a 15-second clip with an SLA turbo LoRA. These are examples from one setup, not guaranteed speeds. If the chosen kernel, GPU, or attention API is unsuitable, the node falls back to dense attention with a warning.
H3 MiniMax Cache (PlagueKind/model_patches/minimax) experimentally skips re-running H3’s transformer block stack when its intermediate features have changed little. reuse_threshold controls how much change is acceptable: raising it may skip more work but can reduce fidelity. start_percent and end_percent keep caching away from early structure-building and late detail-polishing steps; max_steps caps consecutive skips. The cached data can stay with the model or be moved to CPU memory to ease VRAM pressure. It can be combined with H3 SLA Attention in either order.
H3 AdaLN LoRA Fix (PlagueKind/model_patches/minimax) addresses a specific mismatch between dense, full-base H3 LoRAs and pruned or curve-form H3 checkpoints. Place it on the MODEL connection after your LoRA loader. Its default port mode adapts the affected weights; strip drops incompatible weights; off passes the model through unchanged. The author notes that the restored contribution measured on H3 turbo LoRAs is very small, so its practical benefit is usually removing repeated compatibility errors rather than a visible image-quality change.
H3 Video Upsampler and video utilities
H3 Video Upsampler (PlagueKind/minimax/upscaler) takes finished frames, optional audio, and optional text conditioning, then produces enlarged frames, audio, and a report. It targets an output of about 2 megapixels. The file-to-file variant processes a video file into an MP4 using streaming, avoiding the need to keep the entire film’s frames in memory.
The suite offers two re-draw methods:
regen2kuses an Open-H3-Regenerate-2K LoRA for a faster detail re-draw.ic, or in-context, keeps each section’s low-resolution draft available as a video reference while re-drawing it and supplies its soundtrack. This slower approach is intended to better preserve details such as faces, lettering, and mouth movement. It has additional model requirements.
Long clips are processed in overlapping sections, with frame tiling when needed for GPU memory. tile_tokens can learn a workable capacity for your card and remember it. An optional small-face pass detects, tracks, re-draws, and blends faces back into the clip; without OpenCV, that pass leaves the output unchanged. The original soundtrack is retained and retimed for clips that are not 24 fps.
Supporting nodes in the same category are:
- H3 Video Frames (lean): selects a frame window, frame rate, and size while decoding video.
- H3 Load Video (path): reads a video from a file path without an upload or copy.
- H3 Save Video (long-form): streams frames and optional audio into an MP4. It supports RTX hardware encoding or CPU encoding, quality presets, regular and hard-cut keyframes, and optional NVIDIA Video Effects frame generation when its additional requirements are present.
ComfyUI-PlagueKind-Nodes Models
The collection does not provide a list of alternative base MiniMax H3 checkpoints. Its video upsampler does, however, use separate model files for enlargement and re-draw. Choose them according to the method you want to use:
| File or model | Role | When you need it |
|---|---|---|
| Minimax_h3_latent_Upscaler, from LBH-123-AI | Enlarges the clip in latent space before re-draw. The documented 3D and 2D versions are alternatives; the 2D version uses the sharpness control. | H3 Video Upsampler workflows |
open_h3_regenerate_2k_r256pp_bf16.safetensors, from NVlabs | Re-draw LoRA for the faster method. | regen2k |
minimax_h3_ref2va_int8_convrot | Official H3 model required by the in-context method. | ic |
An IC LoRA, such as minimax_h3_open_regenerate_2k_ic_lora_avg_rank_21_bf16.safetensors | Supports re-drawing with the original section kept in view. The author recommends Kijai’s rank-21 resize. | ic |
minimax_h3_hyperflow_8step_v1.0.safetensors, with minimax_h3_lms_v1.0_r64.safetensors also listed | LoRAs used in the documented in-context re-draw setup. | ic |
The latent upscaler belongs in ComfyUI’s models/latent_upscale_models directory; the listed re-draw LoRAs belong in models/loras. The upsampler’s in-context preflight checks its requirements and reports what is missing before processing begins.
What’s New with ComfyUI-PlagueKind-Nodes
The provided documentation identifies version 1.6.4 as current, but does not include a version-by-version changelog. It does identify one replacement worth noting: in version 1.6.0, the H3 Video Upsampler suite replaced the earlier H3 Ultimate Upscale (LD) and H3 Upscale Size nodes. If you are following an older workflow, look for the newer upsampler nodes rather than assuming the old node names still describe the current process.
The documented upsampler suite also credits work derived from H3 Video Upsampler 0.7.0 for the in-context re-draw method and 0.7.1 for 2D latent-upscaler support. Those are source-project version references, not a changelog for this extension.
Troubleshooting ComfyUI-PlagueKind-Nodes
Start by checking the ComfyUI console when a node warns or produces an empty output. Optional features commonly report what is missing there.
My image and mask no longer line up. What should I check?
- Connect both to Unified Resize Image / Mask so they receive the same transformation.
- Check whether center cropping is enabled and whether an aspect-ratio or divisible-by setting changes the final dimensions.
- Use the node’s width and height outputs to verify the resulting size before passing the pair to later nodes.
A RefMod node returns an empty bundle. Why?
- Check the console for a warning.
- Confirm that ComfyUI-MiniMaxH3Mod is available in your ComfyUI setup; this stack calls that collection’s loader for real RefMods.
- If you are not using RefMods, leave the empty bundle in place or omit that connection.
H3 SLA Attention is not speeding up my run. What should I try?
- Look for a console warning indicating a fallback to dense attention.
- Check that the selected engine is available:
comfy_kitchenrequires a suitable comfy-kitchen version, whiletritonrequires Triton. - Remember that sequences below
min_seq_lenstay dense and that low sparsity can be slower than dense attention. - Compare against a run with
enabledswitched off, keeping the rest of the workflow the same.
I see repeated adaln_proj LoRA errors with an H3 checkpoint. What do I do?
- Add H3 AdaLN LoRA Fix after the LoRA loader on the MODEL wire.
- Try the default
portmode to adapt the affected weights. - If you only want to suppress incompatible weights, choose
strip.offleaves the original behavior and errors in place.
The video upsampler stops before the in-context re-draw begins. Why?
- Read its preflight message for the specific missing requirement.
- Check the official
minimax_h3_ref2va_int8_convrotmodel and the documented IC and HyperFlow LoRAs. - If you want to use the faster method instead, select
regen2kand use its corresponding re-draw LoRA.
Do I need every optional package to use these nodes?
No. The collection is designed to load without external dependencies beyond a standard ComfyUI installation. Individual options have their own requirements: for example, OpenCV enables the upsampler’s small-face pass, and NVIDIA Video Effects frame generation requires an RTX card and the nvvfx package.
Learn More about ComfyUI-PlagueKind-Nodes
- ComfyUI-PlagueKind-Nodes on GitHub — the extension’s project page and a place to check its current documentation.
- Minimax H3 latent upscaler on Hugging Face — information about the latent-upscaler files used by the video suite.
- ComfyUI’s native MiniMax H3 nodes — useful examples when you want to understand the base H3 conditioning and image-generation workflow before combining several reference types.
The author credits the video upsampler’s development to work derived from ComfyUI-Hand-Tie-Clips, the cache to ComfyUI-UtilsCollection, and RefMod-bundle helper work to ComfyUI-Fantastic-MiniMaxH3-PromptBuilder. Comparing those projects’ workflow approaches may be helpful if you want to explore a particular H3 feature in greater depth.
