Timeline-based authoring for MiniMax H3 text/image/video generation, reference-to-video, and single-image inpainting workflows, integrated with ComfyUI's native H3 implementation. Separate Image, Video, and Audio lanes, slot-based layout, drag-and-drop / paste / upload, per-clip trims, save/load packs, and structured prompt builders per mode.

MiniMax H3 Director

FL2VA mode: text-to-video (T2VA), first-frame (I2VA), or first+last frame interpolation; up to 2 image slots; automatic alignment-line insertion in the prompt.
IMAGE INPAINT mode: exactly one image reference, no video/audio; a 5-frame image-to-video pass through the native MiniMaxH3ImageToVideo node with the single image as keyframe — the output is a one-frame batch you extract your single result from with a Get Image from Batch. The mode emits an inpaint_requested output so you can switch sampling paths by mode, and the Guide node routes it to the native image-to-video call with the image as first frame and no last frame.
REF2VA mode: up to 9 images, 3 videos, 3 audio clips, 12 files total; each video has a compact V / A / V+A switch (Video only / Audio only / Video+embedded-audio) using the same trim range for both streams; standalone audio also supports left/right trim handles.
REFERENCE VIDEO THUMBNAILS: uploaded videos show their actual first frame as a background preview behind each clip tile, making it easy to identify references at a glance.
REFERENCE HANDLING: reorder clips by dragging between slots, attach external soundtracks to videos, crop references visually via draggable markers, preserve incompatible media when toggling FL2VA ↔ REF2VA instead of losing assets.
LANE SELECTION & PASTE: click an Image/Video or Audio lane to select it; Ctrl+V pastes clipboard images/videos/audio into the chosen lane; drag-and-drop from your file manager works too.
PROMPT BUILDERS:
- FL2VA/I2VA/L2VA/T2VA: guided fields for integrated_multimodal_description, overall_soundscape, and non_diegetic_music with automatic alignment headers.
- REF2VA: simplified six-section free-text builder (subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music) with helper buttons: Insert [Shot N] places shot markers at cursor, Prefill Labels & Summary auto-generates Picture/Video/Audio labels from your inserted media, and Preview Prompt shows the exact assembled prompt in a popup with copy-to-clipboard.
VALIDATED LIMITS: 2–15 second reference windows; max 15s combined visual and audio duration each; strict path-safety under ComfyUI's input directory.
NATIVE ROUTING & LAZY LOADING: hands validated data to ComfyUI's built-in MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo nodes (Image Inpaint uses the image-to-video node); only the selected FL2VA or REF2VA model is requested. REF2VA binds the native inputs by name, keeping the Guide compatible with future native input reordering (v0.4.36).
RESOLUTION PANEL: Aspect / Resolution / Input scaling selectors (all default to Auto) drive the output canvas on MiniMax's 16-px grid — Auto aspect follows your first visual reference, Auto resolution sets a 768 px short side, plus fixed aspect, MP and pixel presets with CUSTOM values. The dropdowns are grouped in columns (aspect by orientation, resolution by ###p / MP, ascending) and label the auto options Native (ShortEdge 768px) / Native (ShortEdge 2048px). Input scaling (Off / Auto / Target / Fit / Fill / Fit+pad / Divisible crop) preprocesses visual references via the included Torch Resize before they reach H3.
PROMPT MODE TOGGLE: a Simple / Structured switch in the mode bar changes how builder fields are assembled into the final prompt — Structured keeps the labelled sections, Simple renders one flat block. The choice is persisted in the workflow and honored by Preview Prompt.
FRAME RATE: a frame_rate FLOAT input (0.1–240, default 24) sets the output FPS and is also emitted as a frame_rate output so downstream nodes can read the effective value.
CROP PREVIEW: a ▶ Play crop button previews only the current crop range, and the preview crop range itself is draggable for quick scrubbing.
PASTE-REPLACE: Ctrl+V onto a selected media tile replaces that tile in place, preserving its slot position instead of appending.
EXTERNAL OVERWRITE INPUTS: optional external_prompt_overwrite (STRING) replaces the assembled builder output; connect both external_width_overwrite and external_height_overwrite (INT) to replace the Director canvas and bypass its sizing and input preprocessing entirely.

Full documentation, UI guide, and prompting reference →

⚡ MiniMax H3 Cache

An approximate, model-scoped whole-block-stack residual cache for ComfyUI's native MiniMax H3 model.

MODEL PATCH: clones only the connected MiniMax H3 MODEL; no global model-class monkey patch.
CONTROLLED REUSE: sampled audio/video-token relative-L1 threshold, 15–90% sampling window, and a bounded number of consecutive cache hits.
STORAGE: auto / CUDA / CPU cached-residual storage with CPU fallback if automatic storage runs out of VRAM.
COMPATIBILITY: preserves ComfyUI block replacements and transformer options; can be chained with Patch Comfy Kitchen Attention.
PDD HEAD BANK: works with ComfyUI's PDD LoRA head bank (0.34+). The node passes the PDD sigma-schedule arguments automatically, so cache and PDD coexist with no extra setup.
PER-TOKEN MASKS: honors per-token video and audio denoise masks, running masked rows at their own strength exactly like Core, so cached and region-masked generations match Core quality.
QUALITY: approximate optimization—higher cache thresholds trade fidelity for more skipped block-stack evaluations.

Full documentation, usage, compatibility, and provenance →

🔥 Patch Comfy Kitchen Attention

A one-input model patch that swaps the connected model's attention backend to Comfy Kitchen INT8 attention at runtime.

Model-scoped: clones only the connected MODEL and sets its optimized-attention override; it never monkey-patches ComfyUI globally.
Safe fallback: if Comfy Kitchen INT8 attention is unavailable in your ComfyUI build, it falls back to the ComfyUI default attention and logs the decision.
Chainable: works before or after MiniMax H3 Cache — both are model-clone patches and compose in either order.

MiniMax H3 Model Loader
          │
          ▼
MiniMax H3 Cache ──► Patch Comfy Kitchen Attention ──► Guider / Sampler

💎 RTX Upscaler & Refiner

State-of-the-art image and video enhancement using NVIDIA RTX Video SDK. It executes up to three sequential passes (Denoise, Deblur, and Upscale) in a single node, processing frame-by-frame to keep VRAM usage predictable and low.

Refine: Independent Denoise and Deblur passes (both off by default).
Upscale: AI-powered VSR and High Bitrate upscaling.
Smart Sizing: Multiple resize modes including Constant Megapixel targets.
Efficiency: Frame-by-frame processing for minimal VRAM usage.
Memory Control: The output batch is allocated lazily (like the reference NVIDIA node — the kernel decides, no up-front memory pressure, no temp file). A disk-backed (mmap) fallback (use_mmap, off by default) is opt-in for very long video batches: when enabled it is the last tier of the VRAM -> RAM -> disk chain, taken only when available RAM is still short after automatic model unloading (auto_unload_models, on by default). Warning: enabling use_mmap writes a multi-giB .mmap temp file to your temp drive for the whole run.

Full documentation →

📐 Resolution Scale Calculator

The DaSiWa Scale Calculator provides mathematically precise resolution management for high-performance video models. It uses a Constant-Area Square-Root method to ensure that your GPU VRAM usage remains stable regardless of the aspect ratio.

Unified Resolution Presets: Pick standard p targets from 144p to 2160p/4K or optimized megapixel tiers from one dropdown.
Clear Aspect Modes: IMAGE ASPECT uses the connected image shape; USE ASPECT BELOW uses the always-visible aspect controls.
Video-Safe Snapping: Standard, Div32, Div64, and custom divisor modes keep dimensions aligned for different model families.

Full documentation →

⚡ Torch Resize

A drop-in replacement for ComfyUI's built-in resize nodes that keeps images sharp and video workflows fast without extra dependencies.

Sharper results: Lanczos resampling with optional sRGB-to-linear gamma correction produces cleaner upscaling and downscaling than native bilinear/bicubic.
Video-friendly batching: Automatically splits long frame sequences into memory-safe chunks so you never run out of VRAM, while keeping output order intact.
Zero extra installs: Runs entirely on the PyTorch build ComfyUI already uses — no Pillow, torchlanc, Triton, or vendor SDK required.
Precise sizing control: Divisible-by alignment, five aspect modes (fit, fill/crop, pad, stretch, long-side crop), and configurable crop/pad placement eliminate guesswork for downstream model constraints.
Alpha preserved: Transparency channels are resized independently without gamma conversion artifacts.

Full documentation →

🎛️ Node Status Switch

The DaSiWa Node Status Switch lets you mute or bypass any node in your workflow using a single toggle. Targets are registered by wiring their outputs into the switch's input slots, which grow dynamically as you connect more nodes (up to 99).

Full documentation →

Quick start:

Add a DaSiWa Node Status Switch to your workflow
Drag any output from the node(s) you want to control into the switch's target_01 input — new slots appear as you connect more
Set action to mute or bypass and configure trigger_on to taste
Toggle enabled directly on the switch

🎬 Advanced LoRA Loader

The Advanced LoRA Loader is a 10-slot stacker for ordinary image/video LoRAs and LTX-2.3. In Basic mode, it loads the complete LoRA map, so it is compatible with standard image and video models. Its VIS control means visual strength: it affects the whole LoRA map in Basic mode, including image models. LTX-2.3 additionally supports independent audio separation.

Model Modes: Select Basic for universal image/video compatibility or LTX-2.3 for separate visual/audio branches. MiniMax H3 uses Basic mode because its transformer blocks are shared between video and audio.
Visual Control: STR × VIS is the effective visual strength. In Basic mode, VIS controls the complete LoRA map; it is not video-only.
Dual-Branch Control: LTX-2.3 can adjust visual (VIS×) and audio (A×) multipliers independently per LoRA.
10 LoRA Slots: Stack up to 10 LoRAs with fine-grained strength control (STR: −5.0 to +5.0).
Toggle All: The ALL header button enables every slot; when all slots are enabled, it disables every slot.
Key Count Indicator: Auto-scans each LoRA to show video/audio key counts before generation.
6 Themes: Switch between Jade, Neon, Studio, Chrome, OLED, and Wood color schemes.
Searchable UI: Quick LoRA search with live filtering in the node itself.
LoRA Info Button (ⓘ): an info button at the right edge of each slot row (drawn as a circle with an "i") opens an info panel with the LoRA's Civitai links — looked up by the file's SHA-256 and cached in lorainfo/ — shown as both mirrors, .com (labeled BLUE: in blue) and .red (labeled RED: in red); the lookup falls back to the .red mirror when .com has no entry, and misses are memoized so a missing LoRA doesn't re-hit the API on every open (Refresh forces a re-lookup). The panel also lists trigger/trained words from the safetensors header and Civitai (click-select, copy) and preview images (Civitai plus a local sidecar image next to the LoRA if present). The file name and sha256 sit on their own rows at the top, so long folder paths can't stretch the controls (v0.4.35).
Trash Button (v0.4.29): a small ASCII-drawn trash button sits directly right of the info button. It resets that slot's LoRA back to None (STR / V× / A× kept — same as picking None in the picker) so you can unstack a LoRA without reopening the picker. The info button shifted slightly left to make room.
PDD/ACC Metadata (v0.4.27): LoRA files are read with their metadata and it is forwarded to Core's load_lora_for_models like the native LoraLoader, so PDD/ACC head banks activate; older ComfyUI builds fall back automatically.
Opt-in LoRA Cache (v0.4.27): a cache button in the control strip keeps each unique LoRA file in a small LRU cache so a LoRA reused across slots is read once — off by default.

Advanced LoRA Loader

Full documentation →

💾 Metadata Image Saver (Civitai Ready)

The DaSiWa Metadata Image Saver ensures your images are fully compatible with Civitai, Hugging Face, and other galleries by embedding A1111-style metadata. It automatically detects LoRAs used in the workflow and supports dynamic filenames.

Civitai Compatibility: Writes the standard parameters block for auto-parsing of prompts and resources.
LoRA Detection: Scans your workflow and appends <lora:name:weight> triggers automatically.
WebP Support: Full "Drag-and-Drop" workflow reconstruction support for both PNG and WebP formats.
Dynamic Filenames: Use placeholders like %seed%, %date%, %model%, %width%, and %height%.
Privacy: Toggle workflow JSON embedding to share images without exposing your full graph.

Full documentation →

🎞️ Enhanced Video Combine

Converts an IMAGE batch into a high-quality video with optional AUDIO muxing and an in-node VHS-style preview.

DaSiWa Enhanced Video Combine

Codecs: Auto (AV1 → VP9 → H.264), or explicit AV1 / VP9 / H.264 / H.265(HEVC). Hardware-first encoder chain (NVENC → QSV → AMF → VAAPI → software); mandatory H.264/MP4 fallback.
Containers: Auto-selects per codec (WebM/MKV/MP4 for AV1/VP9; MP4/MKV for H.264/H.265).
Animated images: Animated AVIF (GPU AV1 or software) and Animated WebP (libwebp_anim). Looping, no audio.
Bit depth & quality: Auto-detects 8-bit vs 10-bit source precision; Auto codec forces 8-bit 4:2:0. CRF/CQ-based quality slider (default 20).
Audio muxing: Opus/AAC/MP3 selectable; Auto uses Opus (WebM) or AAC (MKV/MP4). Bitrates 64–320k. Optional crop-to-audio.
In-node preview: Framed player with native hover-reveal controls and hover-to-unmute audio; an optional Mute checkbox keeps the preview permanently silent, and both Autoplay and Mute are remembered with the node. Streamed H.264 transcoding for AV1/H.265 where needed.
Frame exports: Save first/last frame as PNG alongside the video; all assets published to ComfyUI Assets.
Ping-pong mode: Forward/reverse frame loop.
Workflow metadata: Embed prompt/workflow JSON where supported.
Logging: Compact CLI output with codec/container/encoder decisions and resolved audio settings. Built-in ? help dialog.

Full documentation →

🎬 Watermark Overlay

A professional-grade watermark tool optimized for image and video batches. It uses a stable CPU compositor with high-quality resampling and precise rotation.

Dynamic Random Positioning: Toggle seeded corner cycling while keeping the selected position as the start position.
Splash Mode: Configure dynamic fade-in and fade-out at the start and end of clips for professional branding.
Optical Padding: Automatically adjusts placement by the watermark's visual center of mass for perfect alignment.
Stable Compositing: Output frames are initialized from the source batch before the watermark region is blended, avoiding flicker and black-frame artifacts.

Full documentation →

🩹 Inpaint Crop Prep & Composite

A two-node crop-inpaint-composite pair for any inpainting model. Inpaint Crop Prep tight-crops to the mask and scales it for a high-res inpainter; Inpaint Composite blends the result back onto the original image.

Crop Prep: Gaussian-blurs the mask, extracts its bounding box (with configurable grow_px padding), crops image + mask, and bicubic-scales both to target_width × target_height. Emits cropped_image, cropped_mask, and the original-space bbox_x/y/w/h so you can composite back. can_shrink (default on) allows downscaling; turn it off to keep the crop at least its native size.
Composite: pastes the inpainted source patch back at (x, y) with the (auto-rescaled) mask, applying optional Match Channels or Histogram color correction against the destination region for a seamless blend.
Pure PyTorch: separable Gaussian blur, bicubic resampling, and channel-statistics color matching with no torchvision or extra dependencies.

Wiring:

IMAGE + MASK ──► Inpaint Crop Prep ──► (cropped_image, cropped_mask)
                                     ──► any inpainter ──► source patch
IMAGE ───────────────────────────────────────────────┐
                                                     ▼
                                  Inpaint Composite (x, y, w, h from Crop Prep)
                                                     │
                                                     ▼
                                                   IMAGE

🖥️ System Monitor

A compact system telemetry bar integrated directly into the ComfyUI top toolbar. The adjacent DaSiWa settings button lets you hide the monitor or choose its display mode.

Multi-GPU Support: Separate metrics per GPU device (NVIDIA, AMD, Intel) labeled as GPU0, GPU1, etc.
Resource Metrics: CPU, RAM, SWAP/Pagefile, DISK, GPU Utilization, GPU VRAM, and GPU Temperature.
Visual Feedback: Color-coded borders and proportional background fills (0–100%) for instant at-a-glance assessment.
Lite / Full Modes: Lite is the default compact toolbar view; Full shows every available metric with detailed values and a live 60-second graph.
Responsive Layout: Lite automatically hides lower-priority metrics when toolbar space is limited; Full is a scrollable panel that adapts to narrow screens. Each Lite chip sizes to its label and value (max-content) so text never clips, at any resolution, font, or DPI.
Cross-Platform: Works on Linux and Windows with automatic fallback detection for GPU tools.
Container-safe: In containers and sandboxes where parts of /proc are missing (e.g. /proc/vmstat), probes degrade to n/a instead of warning every second. Set DASWA_SYSTEM_MONITOR=0 (also false/no/off/disable) to fully stop the backend polling thread.
Independent Placement: Renders as its own toolbar element, not dependent on third-party extensions.

Lite mode

Full mode

Full documentation →

🔀 Random String Picker

Bridge any string/text node through DaSiWa Random String Picker to randomize prompt variants inline.

Text passthrough: Accepts a connected STRING input and returns a STRING output.
Inline variants: Replaces every {A|B|C} segment with one randomly selected option.
Multiple groups: Processes any number of groups independently, such as {red|blue} car in {sun|rain}.
Literal passthrough: Text outside complete {...} groups is left unchanged.

Full documentation →

🎯 Seed Control

Seed Control extracts the MiniMax H3 Director's seed panel into a standalone node — the same seed UX for any workflow, without a Director in the graph.

Full 64-bit seeds: unsigned 0..0xFFFFFFFFFFFFFFFF seed field, matching the H3 seed space.
Random|Fixed switch: one segmented switch (a single control, Pixaroma style) — Random rolls a fresh seed on every queue, Fixed keeps the current seed for repeatable results; New rolls a fresh seed and keeps the selected mode, Use Last restores the previous seed and flips to Fixed. Stepping, typing, and the switch all lock the seed as Fixed.
Stacked layout: seed field with an attached ▲/▼ spinner (Pixaroma seed control style; hold to repeat, wraps at the 64-bit bounds) on its own row, then the Random|Fixed switch + New, then Use Last / Last 10 seeds. The panel column has a fixed width (~240 px) and every row is the same 42 px cell height, so resizing the node never stretches or reflows the fields. The seed number auto-fits its font (shrinks as needed) so a full 16-digit seed always displays without clipping.
Lossless 64-bit seeds: the panel keeps the seed as a decimal string and only mirrors it into the hidden INT widget (which coerces 16-digit values to a lossy JS number), so the spinner and display never desync at the top of the 64-bit range.
Last 10 seeds: collapsible history with per-entry copy actions.
External override socket: linking seed disables the local controls (showing an "External seed connected" note) and passes the connected value through — same semantics as the Director's external seed input.
Downstream NOISE output: the seed INT and a NOISE-compatible object are emitted, so the value can be passed straight to any node that accepts a NOISE input (e.g. the LTX sampler's noise socket).
Headless-safe: in Random mode without a local value the backend rolls a fresh seed on every queue, so API clients get the same behaviour without the DOM panel.
Persists with the workflow: mode, last seed and history live in a hidden state widget and survive save/reload.

Full documentation →

🎲 Wildcard & Preset Prompt Builder

DaSiWa Wildcard & Preset Prompt Builder builds positive and negative STRING prompts directly from the bundled dual wildcard library—no downstream picker node needed.

DaSiWa Wildcard & Preset Prompt Builder

Dual style: Switch globally between Booru and Natural Language source keys.
Compact selector: Collapsible categories expose subject checkboxes, weights, deterministic live selections, and right-aligned selected-subject counters that remain visible while a category is collapsed.
Fast inspiration: Random Select replaces the current selection with 1–10 secure-random available Preset/Wildcard subjects.
Reproducible rerolls: Seed plus the stored reroll value reproduce every {A|B|C} choice; New Picks only advances the reroll value, while New picks on every queue opts into fresh output for each queue—including Preview as Text selected-output execution.
Weighted, bounded prompts: Non-1.0 enabled subjects use ComfyUI emphasis syntax. Each positive/negative prompt independently removes complete lowest-weight subjects until it meets the token budget.
Optional prompt prefixes: Connect positive_input or negative_input to prepend an existing prompt to that generated side.
Custom library: Edit or replace data/wildcards_and_presets_dual.json with a compatible library; no checksum sidecar or pinned data version is required.

Wildcard & Preset Prompt Builder documentation →

🧠 LLM / VLM Analyze

The DaSiWa LLM / VLM nodes let you run local transformers chat or vision-language models from inside a ComfyUI workflow. They accept native STRING inputs and native IMAGE batches from nodes such as Load Image or VHS frame loaders.

Native ComfyUI Inputs: Analyze connected text, still images, or video/image-sequence frame batches.
Prompt Presets: Custom system instructions, LTX-2.3/Wan2.2 video prompt enhancement, and image/video caption presets for mixed tags, tag-only, or natural language.
Memory Modes: Keep models cached for speed, or use full cleanup to unload DaSiWa and ComfyUI managed models before/after analysis so later image/video models recover VRAM/RAM.
Frame Sampling: Limit video analysis with max frames, stride, frame strategy, resize controls, context limits, and optional KV-cache reduction.
Local, GGUF, Ollama, or HF Models: Load full Transformers folders, local GGUF through llama.cpp, call Ollama, or download a Hugging Face repo id into ComfyUI/models/llm.

Full documentation →

🛠️ Installation

Manual install

Activate your venv inside your ComfyUI folder

Clone this repo into your custom_nodes folder:

git clone https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes

Install all dependencies:
```
pip install -r requirements.txt
```
Requirement: NVIDIA RTX GPU with drivers 530+. (Windows users may need the NVIDIA Broadcast SDK; Linux usually works out-of-the-box with the pip package).
Restart ComfyUI.

Use ComfyUI-Manager

Search for DaSiWa-Nodes and install.

Credits

The RTX implementation in this collection is based on the excellent work by Deno2026/comfyui-deno-custom-nodes.
Lora-Loader is based on Brojakhoeman/Loradaddyloaderltx.
Ideas for Watermark Overlay are inspired by Artificial-Sweetener/comfyui-WhiteRabbit
MiniMax H3 Director was inspired by the LTX Director concept from whatdreamscost