ComfyUI-UtilsCollection Introduction
ComfyUI-UtilsCollection is a collection of nodes for building image, video, audio, and text workflows in ComfyUI. Instead of serving one generation model, it provides tools you can combine with compatible models and existing ComfyUI nodes.
You can use it to prepare prompts and visual references, transcribe audio, create masks and poses, arrange foreground layers, manage video clips, adjust sampling, and handle everyday tasks such as resizing images or joining text. Its MiniMax H3 tools cover more specialized work, including reference media, clip continuation, and video upscaling.
The collection includes both straightforward utilities and experimental controls. If you are new to it, start with a workflow example or one task you want to solve, rather than trying to use every node at once.
How ComfyUI-UtilsCollection Works
A ComfyUI workflow passes information between connected nodes. Think of each node as a tool at an artist’s workstation: one prepares a reference image, another interprets it for a model, and another saves or combines the result. ComfyUI-UtilsCollection adds tools for those steps.
For example:
- To remove an object, you can load an image and mask, use a LaMa model with
UC_LaMaInpaintto fill the masked area, then save the result. - To build a scene, you can stage foreground cutouts, position them over a background, and composite the final image.
- To guide a MiniMax H3 video, you can prepare ordered reference images or video frames, encode their visual information, and connect the resulting conditioning to your generation workflow.
Some nodes return an edited image, mask, or audio clip. Others return conditioning—the instructions and references a generation model uses—or a patched model with different sampling behavior. Many model-backed nodes use ComfyUI’s model management so they can work within its normal loading and device-selection system.
Several advanced tools are experimental. A reference can encourage a subject or composition, but it does not guarantee exact identity, framing, or timing. Compare results with your usual workflow before relying on a new setting.
ComfyUI-UtilsCollection Features
Prompts, visual conditioning, and MiniMax H3 references
The text-encoding and preset nodes help you assemble model instructions without rebuilding the same prompt structure every time. They include system-prompt, weighted-text, visual-conditioning, and prompt-preset tools. The encoders follow current ComfyUI Core behavior while retaining compatible older node IDs where practical.
The MiniMax H3 nodes offer several ways to guide video generation:
- Image-to-video encoders combine prompts with pictures, video, and—in supported workflows—audio. Picture and Video inputs have separate resolution controls, so you can give still references and moving references different levels of visual detail.
- Temporal fusion adds offset video samples within the usual video-token budget. A temporal density of
1keeps ordinary sampling. You can configure the density and fusion method withUC_MiniMaxH3MediaConfig. - VLM guides add independently encoded image-and-timestamp guidance before the prompt.
UC_MiniMaxH3VLMFrameGuideselects frames by index:0is the first frame,1the second, and-1the last. These guides do not jointly re-encode the original prompt. - Reference extraction, saving, loading, and applying let you reuse H3 image, video, or audio references.
UC_MiniMaxH3RefExtractcan average an image batch into one reference; this saves reference blocks but may lose details unique to individual images. Choose itsvideomode only for an ordered 24 fps clip of at least five frames.UC_MiniMaxH3AudioRefExtractcreates a separate audio reference. Saved references can be stored individually or in a bundle underComfyUI/models/minimax_h3_refs.
When applying saved references, keep their Picture or Video numbers distinct from other references of the same type. The retention control determines how much detail from newly supplied native references is kept; it is not an attention-strength slider. max_ref_tokens reports excess native reference tokens as an error rather than silently discarding references. Stored Qwen embeddings also reflect the text encoder and patches used when the reference was extracted.
Clip continuation carries the end of one generated clip into the next. Save decoded tail frames—and optionally matching audio—from clip 1 to slot 1. Load slot 1 when making clip 2, then save clip 2 to slot 2, and continue in that pattern. Use the same filename_prefix throughout. Load slot 0 for the first clip, which has no preceding context; connect Load’s continuation_media output to the continuation encoder, not Save’s path output. A video VAE is required, plus an audio VAE if saved audio is present.
Continuation Load offers replace, prepend, and temporal fusion modes. Choose temporal fusion when you want to preserve the reference-video timeline while blending saved interior frames into matching times. No mode is selected automatically. When joining clips, UC_MiniMaxH3ClipContinuationAccumulate searches for an overlapping section and trims video and audio at the chosen splice. It assumes 24 fps; larger overlap-search windows take longer, and ambiguous matches can leave both clips unchanged with a console warning.
H3 Reference Video Components prepares reference frames, audio, dimensions, and frame counts on a 24 fps timeline. Use start_at_timestamp to skip the beginning of both video and audio. Set duration to 0 to use the remainder of the clip. For equal parts of a source video, set Divide into segments to N and select a zero-based Segment index from 0 to N-1; set the count to 0 for manual start and duration controls. The preview distinguishes source frames from end padding. Its separate frames output is aspect-ratio matched and center-cropped; the video output retains the source framing.
Image editing, detection, pose, and compositing
Image and mask utilities cover common preparation work: cropping, padding, resizing, blending through masks, extracting colors, handling alpha transparency, matching image properties, and arranging image batches. Detection nodes can turn a subject or region into masks and bounding boxes for later editing.
Model-backed options include:
- YOLO detection and segmentation: find supported subjects in a still image or batch of video frames, returning bounding boxes and masks. Batch size controls how many frames are processed at once. Optional temporal filtering can remove detections that lack support in nearby frames.
- Depth Anything V3: estimate depth, confidence, sky masks, and camera-pose information. Use
independentmode for separate images ormulti_viewfor related views. Depth-output clipping controls how strongly extreme near and far values affect the result. - OpenPose, DWPose, and Animal Pose: produce pose renderings and keypoint data. Their confidence thresholds determine which uncertain points remain. For an ordered video batch, optional temporal filtering removes unsupported keypoints; leave it off for unrelated still images. Smaller batch sizes can help when GPU memory is limited.
- DensePose: produces a rendered body-surface representation, with detection and rendering controls.
For hands-on layouts, Create Bounding Boxes (Nodes 1.0) provides a canvas where you can draw, move, resize, reorder, and describe object or text boxes. It outputs pixel-space boxes and layout information for supported prompt workflows. If connected to a background image, leaving width and height at 0 uses that image’s dimensions; without one, the fallback is 1024 × 1024. Boxes default to outlines, so an unwanted prompt color is not added unless you select one.
The staged compositor helps you arrange cutouts over a background. run_staging prepares cutouts for the placement editor; run_staged composites retained cutouts without reevaluating the foreground branches; and full_run does both in one queue. foreground_0 is the backmost layer. You can also use face-specific staging, or request a separate full-background composite for each foreground instead of stacking them.
After staging, a foreground’s right-click menu provides Brush and Text tools. Brush can add marks or erase original foreground pixels; text is an editable multiline block that follows the foreground’s placement. Visibility controls let you hide these additions without deleting them. Staged cutouts live in server memory and must be recreated after a ComfyUI restart. Brush and text PNGs are saved in ComfyUI input storage, so keep those files when moving a workflow.
Staged compositor workflow JSON · Workflow overview · Source assets
<img src="workflows/CompositorExampleWorkflow.jpg" alt="Staged MediaPipe face background compositor workflow" width="1200">
Audio transcription
UC_WhisperLoader and UC_WhisperTranscribe turn ComfyUI audio into text. Choose transcription to write what was spoken, or translation to produce English text. Spoken language can be detected automatically or supplied as a language code.
Each recording returns transcript text, timestamped segments, and a spoken-language code. Segment times are measured from the start of the recording, even when a long recording is processed in multiple windows. Timestamps are approximate: word-level alignment, speaker identification, and SRT/VTT subtitle formatting are not included. If you explicitly choose a language, the language output reports that choice; translation output remains English.
Whisper needs tiktoken. If it is unavailable, other collection nodes can still load, but transcription reports an error.
Workflow parameters, sampling, and text utilities
Resolution selectors and length pickers help keep image and video dimensions consistent. Switches, random-number nodes, bounding-box adjustments, and list tools help you build reusable workflows instead of editing connections by hand for every variation.
The collection also includes scheduler and model-patching tools for artists who want to experiment with generation behavior:
- MiniMax H3 latent upscaling enlarges video—or joint audio-video—latents without a video-VAE decode-and-re-encode step; attached audio passes through unchanged. The 2D v1 model supports continuous
1.1×–4.0×scaling andsharpnesslevels1–6, from most faithful to crispest. The 3D v1 model targets2×,2.25×,2.5×, or4×and ignores sharpness. Spatial tiling can help with a high-resolution second refinement pass that would otherwise exceed GPU memory. - Unified Attention Patcher selects FlashAttention, SageAttention, or experimental sparse MiniMax H3 SLA where supported.
disabledleaves the model unchanged. SLA can reduce work on eligible H3 attention calls, but it may fall back to dense attention; compare both speed and image quality. - Krea2 Normalized Attention Guidance uses a negative prompt at
CFG 1.0.nag_scalestrengthens avoidance of unwanted concepts, whilenag_tauandnag_alphamoderate its effect. Start and end percentages limit the part of sampling where it runs. - Krea2 Edit Patcher supports reference-guided editing and two-image composition. Reference boost changes how strongly appearance is retained; reference position and fit settings affect placement.
- MOSS-Image Edit Patcher supports reference-guided editing, style or character transfer, and masked inpainting. Its presets provide starting profiles, while
strengthcontrols reference influence andstructure_retentioncontrols how closely the original layout is kept. - Scheduler nodes shape the progression of denoising steps.
UC_SigmaRescalechanges a schedule’s exact start and end values without changing its shape or step count. Some samplers needUC_DiscardPenultimateSigmaas a separate step.
Additional loading, video, LoRA, prompt-building, Markdown-preview, JSON, and text-combination nodes are available when a workflow calls for them.
ComfyUI-UtilsCollection Models
ComfyUI-UtilsCollection uses different checkpoints for different tasks; there is no single model that powers the entire collection. Select a checkpoint that matches the node and task. Some nodes download a missing selected checkpoint when executed, while others require you to provide a compatible file.
| Task | Available models or checkpoint choices | When to choose them |
|---|---|---|
| Whisper transcription | tiny, base (default), small, medium, large-v2, large-v3 | Choose the checkpoint you want to use for speech transcription or English translation. The loader downloads only the selected missing model; an invalid existing checkpoint raises an error. |
| YOLO detection and segmentation | Segmentation: face_yolov8m-seg_60, person_yolov8m-seg, skin_yolov8m-seg_400, hair_yolov8n-seg_60. Bounding boxes: face_yolov8m, hand_yolov8s, yolo11x-train28-best | Match the checkpoint to the subject you want to locate and whether you need a region mask or a box. These nodes use converted .safetensors checkpoints. |
| Depth Anything V3 | da3_small, da3_base, da3_large, da3mono_large | Use with the depth nodes to create depth-related outputs. Checkpoint availability is separate from your choice of independent or multi_view processing. |
| Human and animal pose | OpenPose body, optional hand and face checkpoints; DWPose YOLOX detector and pose checkpoint; Animal Pose AP10K checkpoint; DensePose R50-FPN checkpoint | Choose the pose tool for the kind of keypoints or body representation your workflow needs. The pose checkpoints are converted .safetensors files. |
| Big LaMa inpainting | Big LaMa; Anime/Manga Big LaMa | Use the general checkpoint for ordinary imagery or the anime/manga option for that style. These files are not downloaded automatically. |
| MiniMax H3 CLIP projection | 4B v3.1, 4B v3.1 residual MLP, 8B v3.1, 8B v3.1 residual MLP | Match the projection to your Qwen3-VL 4B or 8B text encoder. UC_MiniMaxH3ClipProjectionPatcher uses one projection at a time. |
| MiniMax H3 PDD Acc | FL2VA 8-step; Ref2VA 8-step | Select the file matching your MiniMax H3 diffusion model. |
| MiniMax H3 latent upscaling | 2D v1 FP16, recommended or BF16; 3D v1 FP16 or BF16 | Choose 2D for flexible scaling and sharpness control, or 3D for its listed fixed scales. The loader accepts .safetensors, not legacy .pth files. |
The pose, YOLO, and Depth Anything V3 descriptions concern converted checkpoint loading and available outputs. They do not establish that these ports exactly match the speed or predictions of their original implementations.
Troubleshooting ComfyUI-UtilsCollection
A Whisper node reports that tiktoken is missing
- Check that the error occurs when you execute transcription; other collection nodes can remain available.
- Add
tiktokento the Python environment used by your ComfyUI instance. - Restart ComfyUI and run the Whisper nodes again.
A model cannot be selected or loaded
- Confirm that the checkpoint matches the node and, where applicable, the model size or diffusion-model variant.
- For manually supplied models, check their expected ComfyUI model category: LaMa uses
models/lama, CLIP projections usemodels/clip_projections, PDD Acc usesmodels/loras, and H3 latent upscalers usemodels/latent_upscale_models. - Refresh ComfyUI’s model files or restart it after adding a checkpoint.
- If a node downloads missing weights on execution, check its console error. An existing invalid Whisper checkpoint is reported rather than overwritten; an unavailable Whisper language for a checkpoint is also reported explicitly.
A saved H3 reference fails when applied
- Check that the base conditioning came from a compatible H3 encoder in this collection and includes the required VLM layout information.
- Check the reference information for duplicate Picture or Video numbers of the same media type.
- If a saved Qwen reference uses a retired timestamp presentation or fixed number-one presentation, extract it again before applying it.
- If you receive a token-limit error, adjust your connected references or
max_ref_tokens; the node will not silently drop them.
Clip continuation produces an unexpected join
- Verify that Load and Save use the same
filename_prefixand that each new clip loads the preceding slot. - Connect Load’s
continuation_mediato the encoder. Do not connect Save’spathto Load. - Select the intended Load mode explicitly. Use temporal fusion if preserving a reference-video timeline is important.
- Check the ComfyUI console for the selected cuts, an unchanged-join explanation, or a warning about ambiguous overlap. The joiner does not stretch time or repair dropped frames and assumes 24 fps.
Staged compositor edits disappear after a restart or workflow move
- Run staging again after restarting ComfyUI; retained cutouts were held in server memory.
- When moving the workflow, keep the Brush/Text PNGs saved in ComfyUI input storage.
- Check Show Brush/Text if an edit exists in the editor but is missing from the output.
Frequently asked questions
Do I need a separate model for every node?
No. Simple utilities work with their inputs alone. Model-backed tasks such as inpainting, detection, transcription, and H3 upscaling require compatible checkpoints.
Are Picture timestamps exact video-frame locks?
No. They are prompt instructions that can guide an approximate timeline; generated results remain variable.
Can I use temporal filtering on unrelated still images?
Leave it disabled. Pose filtering treats the batch as consecutive video frames, and YOLO’s temporal option likewise uses nearby-frame support.
Will H3 references guarantee a character identity or cloned voice?
No. The reference tools add conditioning, not learned trigger words or voice-cloning guarantees.
Learn More about ComfyUI-UtilsCollection
- Eight-image MiniMax H3 storyboard example: Workflow JSON · API workflow JSON · Overview image · Reference images. This example shows eight chronological images guiding the framing and progression of a short video without a native reference video. The images must be available in ComfyUI’s input storage when you open the workflow.
- Staged composition example: Workflow JSON · Overview image · Source assets. Start here if you want to see foreground placement and face-aware compositing in a connected workflow.
- Additional model information: Model migration documentation describes the collection’s model-conversion approach. It is most relevant when working with model files beyond the listed checkpoints.
- ComfyUI basics: ComfyUI is the host application for these connected-node workflows.
<img src="workflows/UC_AdvancedMiniMaxH3ImageToVideo/QwenOnly_8Image_1024VLM_Overview.png" alt="MiniMax H3 image-to-video workflow using eight ordered reference images" width="1200">
