Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-DaSiWa-Nodes

ComfyUI Extension: ComfyUI-DaSiWa-Nodes

Repo Name

ComfyUI-DaSiWa-Nodes

Author
darksidewalker (Account age: 1999 days)
Nodes
View all nodes(3)
Latest Updated
2026-11-08
Github Stars
0.22K

How to Install ComfyUI-DaSiWa-Nodes

Install this extension via the ComfyUI Manager by searching for ComfyUI-DaSiWa-Nodes
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-DaSiWa-Nodes in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-DaSiWa-Nodes Description

ComfyUI-DaSiWa-Nodes provides custom ComfyUI tools to direct AI video and audio, refine images and frames, manage prompts and LoRAs, prepare inpainting, and export finished media.

ComfyUI-DaSiWa-Nodes Introduction

ComfyUI-DaSiWa-Nodes is a collection of custom nodes for creating, refining, organizing, and exporting AI images and videos in ComfyUI. It brings several kinds of work into one collection: directing MiniMax H3 video and audio, improving image and video quality, managing prompts and LoRAs, preparing images for inpainting, and saving finished work.

You do not need to use every node. For example, you might use MiniMaxH3 Director to plan a video, MiniMaxH3 Enhanced Upscale to enlarge it, and Enhanced Video Combine to export it. For an image workflow, you might use Torch Resize, Inpaint Crop Prep & Composite, and Metadata Image Saver instead.

The collection also includes workflow conveniences such as a seed controller, editable canvas labels, a node status switch, and a system monitor. Its installed version appears in ComfyUI → Settings → About. Additional options are under ComfyUI → Settings → Other → DaSiWa.

Read the collection’s news and complete changelog.

How ComfyUI-DaSiWa-Nodes Works

A ComfyUI workflow passes information between connected nodes. Think of each node as a workstation with a particular job: one prepares a prompt, another generates frames, another improves the result, and a final node exports a file. DaSiWa nodes fit into those chains or provide controls alongside them.

A few distinctions will help you choose the right tool:

  • Frames versus video files: Many video-related nodes work with a batch of images, where each image is one frame. An export node such as Enhanced Video Combine turns those frames into a playable video and can add audio.
  • Latent versus decoded frames: A latent is the model’s working representation of an image or video before it becomes ordinary visible frames. MiniMax H3’s latent refinement nodes work at this stage, so they belong before decoding and export. RTX enhancement works on decoded image frames.
  • References versus prompts: A prompt describes what you want. Reference images, video, and audio give the model material to draw from. MiniMaxH3 Director can organize both on a timeline and pass them to native H3 nodes.
  • Repeatability versus variation: A fixed seed helps you reproduce a result with the same workflow and settings. Random seeds, wildcards, and prompt presets help you explore alternatives.

For instance, an H3 video workflow can go from Director and Guide → sampler → latent upscale or regeneration → video/audio decode → Enhanced Video Combine. If you only need to resize an existing image, Torch Resize can be used without any H3 nodes. The connections and limitations for each tool are covered in its linked guide.

ComfyUI-DaSiWa-Nodes Features

Direct and extend MiniMax H3 videos

MiniMaxH3 Director is the collection’s central workspace for H3 video/audio creation. You can begin with text, images, or an existing H3 video. It supports text-to-video/audio, image-guided modes, first/last-frame guidance, multi-reference generation, and an image-inpaint pass that produces a still frame. Its reference lanes keep image, video, and audio sources separate, while allowing you to reorder media, trim clips, preview sound, and write prompts for individual references.

The Director also offers:

  • Prompt editing and Prompt Forge: Write your own prompt, or review a draft produced with an optional local LLM, Ollama, or an operator-configured compatible server. Prompt Forge can use character, place, style, pose, and shot information. You choose whether to apply its draft.
  • RefMods: Select reusable reference bundles for people or other concepts, adjust their strength, or create a bundle from selected media. Saved bundles can be edited as copies without overwriting the original.
  • Resolution and frame-rate controls: Choose aspect and size presets, resize inputs, and pass a matching frame rate downstream. The default frame rate is 24 fps.
  • Continuity: Optionally extend a completed H3 take from a saved latent checkpoint. The duration setting controls how many seconds to add; a separate prompt describes what happens next. Capturing a new take and advancing the source are explicit choices.
  • Reference packs: Save and reload reference files, prompts, and RefMod selections for later work.

MiniMaxH3 Director interface Director Prompt Forge Director RefMod controls

Explore the Director UI, modes, and prompting guide · Learn how H3 Continuity works.

MiniMaxH3 Continuity • Segment Prepare and Segment Decode offer an alternative when you want to render and decode only the newly added part of a continuing video. Keep the parent checkpoints for each segment; full-timeline output remains the default. MiniMaxH3 • Join Clips can assemble ordered Continuity exports or independent uploaded clips. It re-encodes at 24 fps and can optionally generate an H3 transition between cuts when the required Director, Guide, and sampler are connected.

Director clip assembly dialog

Compare Continuity output options and clip-joining limits.

Refine H3 results

MiniMaxH3 Enhanced Upscale enlarges a completed H3 video/audio latent and optionally refines it with diffusion while preserving the original audio. Its default scale factor is 2×. You can choose a learned upscaler or interpolation, and use the denoise setting to control refinement: 0 skips refinement. Spatial tiling and temporal chunking are on by default to divide demanding work into more manageable pieces. Memory Planning estimates whether a plan fits, but it cannot guarantee that your GPU will avoid an out-of-memory error.

MiniMaxH3 • Hi-Res Regenerate packages the Open-H3-Regenerate-2K recipe into a node. Choose 1080P, 2K, or 4K to enlarge and refine a draft, or Original to refine it without changing its canvas size. Its default uses the 2K target, the recipe’s refiner LoRA, and three refinement steps. Lower denoise values use fewer steps; 2D sharpness offers six tiers, from the most faithful to the sharpest, when a compatible 2D upscaler is in use. The node outputs a refined latent, not an encoded video.

Compare the H3 upscale and regeneration controls.

MiniMax H3 De-rope addresses fast-motion smearing after generation. Its Prepare and Recover nodes identify troublesome motion spans, run a second pass on a slowed version, then return frames to the original timing. Optional background preservation can retain detail from the original decoded frames in relatively static areas. Use the original audio when assembling the result.

See De-rope wiring and second-pass settings.

MiniMaxH3 Cache is an optional speed optimization attached to the connected H3 model. It reuses some model calculations when they are sufficiently similar. A more permissive cache threshold can skip more work, but may reduce fidelity; compare a cached render with your original before relying on it for a final shot.

Read the Cache quality and compatibility notes.

Improve decoded images and video frames

RTX Upscaler & Refiner uses NVIDIA RTX Video SDK effects on image frames. Turn on denoise or deblur independently, and choose VSR for typical compressed footage or High Bitrate upscaling for cleaner sources. You can also set resize behavior, per-pass strengths, and optional TrueHDR conversion. Frame count and FPS do not change. TrueHDR needs appropriate HDR encoding and tagging later; the node’s IMAGE output alone does not carry that metadata. Long batches still need memory even though the node processes them in smaller internal windows.

RTX Upscaler & Refiner

Read the RTX Upscaler & Refiner guide.

RTX Frame Generation makes intermediate frames between originals. Choose a multiplier from 2–16, or generate one frame at a selected point between a pair. If you use a multiplier of M, export at source FPS × M to preserve playback speed. The node operates on image batches; route audio separately. It requires SDK support for VideoFrameGeneration.

RTX Frame Generation

Check frame timing, quality settings, and limits.

Seamless Loop works on an image-frame batch to soften the transition from the end of a clip back to its beginning. It analyzes overlap candidates and interpolates the transition using a native RIFE/FILM checkpoint or NVIDIA RTX Frame Generation. Leave exact_endpoint off for ordinary continuous playback; turn it on only when identical first and last pixels are required. The output duration can change. Connect audio and frame-rate information when you need matching looped sound. Very different motion or scenes can still leave a visible seam.

Seamless Loop

Learn about loop settings and model compatibility.

Tiled Diffusion (Image) patches a compatible SD, SDXL, or Flux-family image model so it samples overlapping pieces of a large canvas and blends them. Place it between the model loader and sampler. The default tile is 1024 px with 128 px overlap. ControlNet is not supported with this node.

Read the Tiled Diffusion guide.

Inpaint Crop Prep & Composite is a two-node route for fixing a small area of a larger image. Crop Prep finds the masked region, adds your chosen padding, and prepares a crop for an inpainting model. Composite places the edited patch back into the original image; optional color matching helps it blend. This can be useful when you want the inpainter to spend its resolution on a face, hand, or object rather than the entire canvas.

Original image + mask → Inpaint Crop Prep → your inpainter
Original image + edited patch + crop coordinates → Inpaint Composite → finished image

Control size, prompts, and model settings

DaSiWa Scale Calculator helps choose dimensions for a desired aspect ratio and resolution. Its constant-area calculation aims to keep the approximate pixel area stable as the aspect changes, which can make size choices more predictable for video models. Select a standard resolution or megapixel target, then align dimensions to the divisor your downstream model needs.

DaSiWa Scale Calculator

See resolution and aspect controls.

Torch Resize resizes images or frame batches with Lanczos resampling. Choose whether an image should fit, fill and crop, pad, stretch, or use a long-side crop; set crop or pad placement and divisible-by alignment. Optional gamma-aware resizing can change how edges and tonal transitions look. Transparency is handled separately, and long batches are processed in chunks while retaining their order.

Torch Resize

Read the Torch Resize guide.

Advanced LoRA Loader stacks up to 10 LoRAs. Use Basic mode for ordinary image and video models, including MiniMax H3. Its VIS control scales the whole LoRA in Basic mode, not just video. Use LTX-2.3 mode when you need separate visual and audio multipliers. Each slot can be enabled, adjusted, searched, or reset; the information panel can show available trigger words and previews. Increasing strength makes a LoRA’s influence more pronounced, but the useful amount depends on the LoRA and your model.

Advanced LoRA Loader

Learn the LoRA modes and controls.

Seed Control provides Random and Fixed modes for any workflow. Random chooses a fresh seed on each queue; Fixed keeps the current one. New chooses another seed, Use Last restores the previous seed, and a history holds the last ten. You can connect an external seed instead of using the panel.

Read the Seed Control guide.

Random String Picker adds simple variation to connected text. For example, {red|blue} car in {sun|rain} picks one choice from each complete group while leaving the rest of the text alone.

Random String Picker

See more prompt-variation examples.

Wildcard & Preset Prompt Builder assembles positive and negative prompts from its bundled Booru or Natural Language libraries. Choose subjects yourself or use Random Select for inspiration. Weights affect emphasis; a token budget removes lower-weight complete subjects when needed. A seed and stored reroll value can reproduce choices, while New Picks changes them. Connected prompt text can be prepended, and your own variables can be defined in variables_json.

Wildcard & Preset Prompt Builder

Explore presets, variables, and rerolls.

LLM Prompt Writer & Analyze can help draft model-oriented prompts or describe connected text, still images, and sampled video frames. Prompt Writer includes guides for Anima, Illustrious, Krea2, Wan 2.2, LTX 2.3, and YuE2. You can adjust style, detail, and creativity, and optionally supply reference pictures. Analysis controls limit how many video frames are examined and how much memory the task uses. These tools use configured local or external language/vision models; they do not download a model during a run.

Read the LLM and vision-node guide.

The beta57 scheduler adds a scheduler choice to KSampler and BasicScheduler lists without adding a node to the graph. It pairs with ComfyUI’s res_multistep sampler.

Read about the scheduler.

Export and manage finished work

Enhanced Video Combine turns an image-frame batch into a video, with optional audio. Its automatic codec choice tries AV1/MP4, then VP9/WebM, then H.264/MP4; you can also select a codec and container yourself. Adjust quality, choose audio settings, make a forward/reverse ping-pong loop, or save the first and last frames beside the video. An in-node player lets you review the result. Animated WebP and AVIF are available for looping outputs without audio; supported WebP and VP9 paths can retain transparency.

Enhanced Video Combine

Review codecs, previews, audio, and naming options.

Metadata Image Saver saves gallery-friendly image metadata, including prompts and detected LoRA information. It supports PNG and WebP workflow reconstruction and filename placeholders such as %seed%, %date%, and %model%. Turn off workflow JSON embedding when you want to share an image without sharing its full node graph.

Metadata Image Saver

Read the image metadata guide.

Watermark Overlay adds a watermark to images or frame batches. You can choose placement, cycle corners with a seed, or fade the watermark in and out at the beginning and end of a clip.

Watermark Overlay

Explore watermark controls.

Keep the workspace organized

Node Status Switch lets you mute or bypass connected target nodes with one control. Connect an output from each node you want it to manage, choose the action and trigger, then toggle the switch. It can register up to 99 targets.

Node Status Switch

Read the Node Status Switch guide.

Lable (DaSiWa) places editable, workflow-only labels on the ComfyUI canvas. Add text or an embedded image, then change its font, color, opacity, alignment, size, and placement. Labels travel with saved workflows and do not execute as generation nodes.

See label controls and compatibility.

System Monitor displays available CPU, memory, disk, and GPU readings while you work. Choose Lite, Ultra compact, or Full mode, and adjust its placement or appearance. A separate Free Memory button offers explicit ComfyUI cleanup actions; it does not make an oversized live image batch disappear.

System Monitor Lite mode System Monitor Full mode System Monitor Ultra compact mode

Read the System Monitor guide.

ComfyUI-DaSiWa-Nodes Models

This collection is not a single generation model. It provides nodes that work with models or checkpoints selected elsewhere in ComfyUI. The context identifies these model choices:

Model or checkpointWhat it is used forWhen to choose it
MiniMax H3 modelGenerates H3 video/audio and supplies the model connected to Director-related sampling, caching, and regeneration.Use for the H3 workflows described above.
H3 latent upscaler, 2D (2d_conv_v1)Learns to enlarge H3 latents; supports sharpness tiers 1–6.Choose when you want to control the balance between a more faithful result and a sharper one. An installed 2D checkpoint is the Enhanced Upscale default.
H3 latent upscaler, 3D (3d_conv_v1)Another supported learned H3 latent-upscale checkpoint type.Choose when your selected checkpoint is this type; the 2D sharpness control does not apply.
InterpolationEnlarges an H3 latent without a learned upscale checkpoint.Choose as an alternative to learned upscaling; it is a selected method, not a fallback used silently if a checkpoint fails.
Open-H3-Regenerate-2K refiner LoRASupplies the published refinement recipe for Hi-Res Regenerate.Use with the plain connected H3 model for that node’s 1080P, 2K, 4K, or Original-target refinement. Its published Diffusers-key file must first be converted to native H3 keys.
RIFE or FILM frame-interpolation checkpointCreates transition frames for Seamless Loop’s native interpolation option.Choose a compatible checkpoint from models/frame_interpolation/ when using the native RIFE/FILM path. The alternative RTX path uses NVIDIA VFX instead.
Language or vision-language modelHelps Prompt Forge, Prompt Writer, or Analyze draft and interpret text or images.Choose a configured local model, Ollama, or compatible server when you want optional prompt assistance. The appropriate choice depends on whether your task needs text only or image understanding.
LoRA files and RefMod bundlesLoRAs influence compatible generation models; RefMods provide reusable H3 reference material.Use LoRAs for a desired learned influence and RefMods when you want to reuse saved reference concepts or media.

The readme does not specify individual downloadable checkpoint filenames or benchmark one model against another. Use the relevant feature guide to check format and compatibility before selecting a file.

What’s New with ComfyUI-DaSiWa-Nodes

The News & Changelog is the source for the complete release history. The original readme highlights these version-specific changes; it does not provide a complete version-by-version timeline:

  • v0.5.51: Hi-Res Regenerate processing phases can appear under the owning job when paired with DEAS 0.2.29. This makes longer upscale and refinement runs easier to follow; missing telemetry does not change the result.
  • v0.5.29: Metadata Image Saver adds LoRA hashes to its metadata, helping galleries associate an image with its LoRA resources.
  • v0.5.25: Enhanced Video Combine can preserve RGBA transparency in supported Animated WebP and VP9 WebM/MKV outputs. Other formats still discard alpha.
  • v0.4.45: Enhanced Video Combine accepts a connected seed for %seed% in output names, making generated videos and frame exports easier to match to a run.
  • v0.4.40: Enhanced Video Combine uses PyAV for encoding and adds seekable previews, making it easier to pause and scrub generated videos in the node.
  • v0.4.35: Advanced LoRA Loader presents filenames and SHA-256 hashes on separate rows in its information panel, so long paths do not stretch the controls.
  • v0.4.29: Advanced LoRA Loader adds a per-slot trash button that resets the selected LoRA to None without resetting that slot’s strength controls.
  • v0.4.27: Advanced LoRA Loader forwards metadata needed by compatible PDD/ACC head banks and adds an optional, off-by-default LoRA file cache for reuse across slots.

Check ComfyUI → Settings → About for your installed version before following instructions for a newer feature.

Troubleshooting ComfyUI-DaSiWa-Nodes

An H3 upscale or RTX task runs out of memory

  1. Reduce the input resolution, video length, or target scale.
  2. For MiniMaxH3 Enhanced Upscale, leave spatial tiling, temporal chunking, and Memory Planning on while testing. Turning Memory Planning off bypasses its estimate; it does not create more memory.
  3. For RTX Upscaler & Refiner, remember that internal chunking limits processing intermediates, but the complete input and output batches still need storage. Consider shorter frame batches. Disk-backed output is an optional last resort, not a substitute for all memory requirements.
  4. Check System Monitor for available resources, then try a smaller test before running the full clip.

A loop still has a visible join

  1. Preview the end and beginning of the source clip. A major scene or motion change can remain noticeable after processing.
  2. Try a clip with more compatible movement and lighting at its endpoints.
  3. Keep exact_endpoint off for normal continuous playback; enable it only if identical endpoint pixels are specifically required.
  4. Review the node’s reported residual discontinuities and the Seamless Loop guide.

An RTX effect or frame-generation option is unavailable

  1. Check whether the required NVIDIA GPU and SDK feature are available in your environment. Available RTX Upscaler effects are detected from the installed SDK.
  2. For RTX Frame Generation, confirm that the SDK provides VideoFrameGeneration.
  3. If you are using Seamless Loop, consider its native RIFE/FILM option with a compatible interpolation checkpoint instead of its RTX option.

RTX enhancement details · RTX Frame Generation requirements.

A Hi-Res Regenerate LoRA is rejected

  1. Confirm that you selected the refiner LoRA intended for the Open-H3-Regenerate-2K recipe.
  2. If the node reports Diffusers keys, convert the file to native H3 keys using DaSiWa Quant Station → Diffusers → H3, as described in the source readme.
  3. Connect the plain H3 model to Hi-Res Regenerate; the specified LoRA already contains the additional components noted in the recipe.
  4. See the Hi-Res Regenerate instructions before rerunning a long render.

A video preview looks different from the downloaded file

Enhanced Video Combine may create a browser-compatible H.264/AAC preview for a video exported with another codec. The downloadable video remains in the selected original format. Check the output file when judging codec or container behavior.

Frequently asked questions

Do I need an LLM to use MiniMaxH3 Director?
No. You can write prompts yourself. Prompt Forge is optional, and manual descriptions can be used without vision analysis.

Does MiniMaxH3 Enhanced Upscale change the audio?
It returns the original audio stream unchanged while processing the video/audio latent.

Does RTX Upscaler & Refiner increase the video’s frame rate?
No. It keeps frame count and FPS unchanged. RTX Frame Generation is the separate tool for adding frames.

Why does a multiplied RTX Frame Generation output play slowly?
Check the export frame rate. For a multiplier of M, encode at source FPS × M to preserve the source playback speed.

Can I keep a transparent video or animation?
Enhanced Video Combine retains alpha only in the supported Animated WebP and VP9 WebM/MKV paths described above. Other output formats drop it.

Can I share an image without sharing my complete workflow?
Yes. Turn off workflow JSON embedding in Metadata Image Saver.

Will the same seed always guarantee the same image or video?
A fixed seed preserves one important input, but changing the model, prompt, node settings, or workflow can change the result. Keep those settings the same when comparing runs.

Learn More about ComfyUI-DaSiWa-Nodes

Start with the guide for the task you are doing:

Each feature section above links to its own detailed guide where one is provided.

ComfyUI-DaSiWa-Nodes Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-DaSiWa-Nodes detailed guide | ComfyUI