ComfyUI-PlagueKind-Nodes

H3_refmod_loader

Loads an H3 reference-mod stack from JSON through the ComfyUI-MiniMaxH3Mod backing node and returns mods with a prompt hint. If the backing node is unavailable or stack data cannot be parsed or loaded, it returns an empty bundle.

PlagueKind/Loaders · 0 entradas · 3 parámetros · 2 salidas

H3AdaLNLoRAFix

Makes dense (full-base) MiniMax-H3 LoRAs — including turbo LoRAs — work on pruned / curve-form H3 checkpoints, instead of being silently skipped with 51 ERROR lora ... adalnproj lines per model load. Drop it on the MODEL wire anywhere after your LoRA loader(s) — rgthree's Power Lora Loader, ComfyUI's own, any stack. It operates on the patches those loaders already attached, so it needs no knowledge of which LoRAs were chosen and nothing upstream has to change. A pruned checkpoint stores its AdaLN (timestep-modulation) projections over an 8-wide curve basis instead of the dense 2688-wide time embedding used by full H3 LoRAs. Without this fix, ComfyUI can't reconcile the shapes, logs one ERROR line per incompatible key, and drops all 51 — including on H3 turbo LoRAs, which is where this most commonly shows up. This node rebases those weights onto whichever basis the target model actually uses, so they apply instead of being dropped, and works in both directions (dense LoRA → pruned base, or curve-form LoRA → full base). Measured on real H3 turbo LoRAs, the restored contribution is only 0.02% of the modulation signal — so the practical benefit is a clean log, not a visible quality change. mode: port — rebase the AdaLN weights onto the basis the model actually uses. Quiet log, timestep modulation restored. strip — drop the incompatible weights. Quiet log, output identical to having no fix at all. off — passthrough, leaves the errors in place. Fails safe: if the rebase itself errors, the unfixed model still runs and generates — it just logs the same errors it would have without…

1 entrada · 1 parámetro

H3MiniMaxCache

Reuse MiniMax-H3's whole-block-stack residual across similar steps. H3 is a 33B model, so every sampling step is expensive and video sampling means a lot of steps. This patches a *cloned* model patcher only -- via ``ModelPatcher.add_object_patch`` on ``diffusion_model._forward`` plus a ``block_loop`` replacement and an ``OUTER_SAMPLE`` wrapper -- so nothing about the Core model class is touched globally. If the cache misbehaves, only the clone you fed the sampler is affected. Ported from silveroxides/ComfyUI-UtilsCollection's UC_MiniMaxH3Cache (MIT), with permission, in exchange for this pack's H3 SLA Attention node. See h3cache/patch.py for the implementation.

1 entrada · 6 parámetros

H3SLAAttention

Block-sparse attention for MiniMax-H3. sparsityratio — fraction of key blocks skipped (default 0.90). 0.85 is lightx2v's shipped value; 0.90 measured 15% faster. Below 0.60 the kernel is slower than dense, so it's a real floor, not a safe fallback. blocksize (64 / 128) — how many sequence tokens share one key selection. Matters far more for audio than video: 128 forces 1.6 s of audio down one attention pattern and speech comes out robotic; 64 is clean for 2% more time. minseqlen — sequences shorter than this stay dense, protecting the short text-refiner attention and short/low-res clips where sparsity would cost more than it saves. denselaststeps — run the final N sampling steps at full attention to recover fine detail, since the last step's error is the one you actually see. protectaudio — always attends the [text | cond | audio] prefix regardless of top-k, since audio is only 1% of the packed sequence and plain top-k can drop it entirely. Costs 7%. enabled — bypass toggle for a like-for-like dense speed baseline without rewiring the graph. Follows ComfyUI's actual WrapperExecutor chain and derives its position in the sampler from samplesigmas/sigmas rather than counting model invocations, so it stays correctly synced and composes cleanly with step-caching/forecasting accelerators (e.g. Spectrum) that skip some denoiser evaluations. Automatic dense fallback if Triton is missing, the GPU is unsupported, or the ComfyUI attention API changes — a broken patch never blocks a run, it just runs dense. Measured end-to-end on a 5090 at 768p/15s with the SLA turbo LoRA: 44 s/it den…

1 entrada · 14 parámetros

H3UpscaleSizeLD

Turn the generation size + a target into ONE upscale width/height,

PlagueKind/Upscaling · 0 entradas · 7 parámetros · 4 salidas

LTX_lora_loader

Applies enabled LoRA entries from a JSON stack to a model, using per-entry video, audio, and temporal multipliers and a selected loading mode. An optional CLIP input is passed through with the loaded stack.

PlagueKind/Loaders · 2 entradas · 2 parámetros · 2 salidas

LTX25ICLoRALoader

Load an MSR multi-reference IC-LoRA for LTX2.5. Applies the regular LoRA weights to the model and extracts the learned Fourier-MLP reference-slot embedding tensors, which 'LTX25 Reference Params' uses to embed each reference still into its own slot. The output type is compatible with the ComfyUI-LTX2.5-MSR package's IC-LoRA Loader.

1 entrada · 2 parámetros

LTX25LatentUpscaleParams

Bundle the LTX2.5 latent upscale settings for the Ultimate Upscale node.

LTX25ReferenceParams

Encode reference stills into LTX2.5 guide latents for 'LTX25 Ultimate Upscale'. Works with or without ComfyUI-LTX2.5-MSR: with the MSR IC-LoRA Loader output connected, learned slot embeddings are applied and consecutive negative temporal offsets are assigned (MSR training layout); without it, plain guides at offset 0. The guides are encoded ONCE here; the main upscale node resizes them to each chunk's grid and appends them as near-clean conditioning tokens.

2 entradas · 2 parámetros

LTX25SpatialSplitParams

Bundle the spatial tile settings for the LTX25 Ultimate Upscale node.

0 entradas · 14 parámetros

LTX25TemporalSplitParams

Bundle the temporal split settings for the LTX25 Ultimate Upscale node.

0 entradas · 4 parámetros

LTX25UltimateUpscale

One node for the full LTX2.5 latent re-enhancement pipeline.

MMH3LatentUpscaleParams

Bundle model-free latent upscale settings (interpolation) consumed by the Ultimate Upscale node. The video latent is resized spatially, audio passes through. Mirrors ComfyUI's 'Upscale Latent' node but keeps the H3 nested (video+audio) structure intact.

0 entradas · 3 parámetros

MMH3LatentUpscaleWithModelParams

Bundle the H3 3D model-based latent upscale settings consumed by the Ultimate Upscale node.

0 entradas · 5 parámetros

MMH3SpatialSplitParams

Bundle the spatial split settings consumed by the Ultimate Upscale node.

0 entradas · 16 parámetros

MMH3TemporalSplitParams

Bundle the temporal split settings consumed by the Ultimate Upscale node.

0 entradas · 3 parámetros

MMH3UltimateUpscale

One node for the full latent re-enhancement pipeline.

PK_MiniMaxH3Combined

No hay una descripción breve disponible para este nodo.

PlagueKindBundleIn

No hay una descripción breve disponible para este nodo.

PlagueKind/Utils · 0 entradas · 0 parámetros · 1 salida

PlagueKindBundleOut

No hay una descripción breve disponible para este nodo.

PlagueKind/Utils · 1 entrada · 0 parámetros

UnifiedResizeImageMask

A single ComfyUI node that ensures consistent resizing behavior between images and masks using a unified geometric pipeline. Multiple scaling modes: Dimensions (W × H) Multiplier Longer Side Shorter Side Total Pixels (MP) Aspect-ratio preservation option Center crop alignment Divisible-by constraint (useful for latent models like LTX-2.3 / SDXL workflows, where other nodes only do one side.) Set divisible by 1 to disable. Unified image + mask transformation pipeline Stable tensor-based mask resizing (no PIL dependency issues) Default ComfyUI workflows often suffer from: mask stretching inconsistencies image/mask misalignment after resize inconsistent crop behavior between pipelines This node ensures both image and mask follow identical geometric transformations for predictable inpainting and compositing results. Unified Resize Image / Mask (Clean) Category: image

PlagueKind/Image · 4 salidas

VisualBBoxCrop

Provides a visual bounding-box crop tool with an interactive crop box, aspect-ratio options, normalized coordinates, and crop-origin outputs.

PlagueKind/Image · 6 salidas

VisualCropAndResize

Combines a visual bounding-box crop with the unified resize pipeline in one node.

PlagueKind/Image · 6 salidas