ComfyUI-FBnodes

A grab bag of handy ComfyUI nodes I built for my own workflows and figured someone else might enjoy too. Video saving, universal switches, LoRA helpers, animated latent previews, Crop tool and whatever else I end up needing next.

Nodes

Save Video+

Save video with H.264 or H.265 (HEVC) codec and quality control. Includes audio muxing and workflow metadata embedding. If yuv444 is selected, will generate a preview clip, so it can still be seen in browser. (Saved in temp)

• Codec: H.264 (8-bit, max compatibility) or H.265/HEVC (10-bit, better compression)
• Chroma: yuv420 / yuv422 / yuv444
• CRF: Constant Rate Factor quality control (0=lossless, 18-23=high quality)
• Preview mode: Toggle save off for fast preview-only encoding
• Latent saving: Optionally save the latent alongside the video for easy re-generation

Load Image+:

• Image/Video Screenshot Loading: Image loader, based on Prompt Extractor from Prompt Manager
• Input/Output Folder Switching: Toggle between browsing your input or output folder directly from the node
• File Browser: Same thumbnail browser as Prompt Extractor with subfolder navigation
• Video Frame Scrubbing: Load any frame from a video using the frame position slider
• Drag-and-Drop Support: Drop images or videos directly onto the node
• Image Preview: Built-in preview with click-to-enlarge modal for images and videos
• Mask Editor Support: Right-click to open ComfyUI's MaskEditor, painted masks are displayed in the preview
• IMAGE + MASK Output: Outputs both IMAGE and MASK tensors, with alpha channel extraction from images

Load Video+

Video loader with the same file browser and UX as Load Image+, but for videos. Outputs a VIDEO type for use with Get Video Components+.

• Input/Output Folder Switching: Toggle between browsing your input or output folder
• File Browser: Thumbnail browser with video-first filtering and subfolder navigation
• Video Preview: Click-to-enlarge modal with playback controls
• Drag-and-Drop Support: Drop video files directly onto the node
• VIDEO Output: Outputs a VIDEO, ready to pipe into Get Video Components+

Load Audio+

Audio loader with the same file browser and UX as Load Image+/Video+, plus in/out trimming.

• Input/Output Folder Switching: Toggle between browsing your input or output folder
• File Browser: Thumbnail browser with audio-first filtering and subfolder navigation
• Drag-and-Drop Support: Drop audio files directly onto the node
• Trim timeline: Draggable in/out handles with playhead, scrubbing, and clip looping
• Outputs: AUDIO (trimmed) and duration (seconds)

LTX Review

Review gate node for 2-pass LTX-style workflows. It pauses execution, decodes video/audio latents internally, opens a review popup with playback, and continues only when you proceed (or timeout policy allows it).

• Inputs: video_latent, audio_latent, video_vae, audio_vae, fps, timeout, on_timeout, enable
• Popup controls: Proceed, Cancel, Requeue
• Audio + video playback: Always generates a browser-compatible H.264/yuv420 preview in temp from decoded latents
• Timeout behavior: Auto-proceeds or auto-cancels based on on_timeout
• Enable switch: When OFF, passes through latents immediately and skips both VAE decodes/review popup
• Preview passthrough when disabled: Optional preview_video preserves review_path output for preview workflows without running decode
• Outputs: video_latent, audio_latent, review_path
• review_path: Path to the generated temp preview clip

LTX Review Preview

Minimal companion preview node for displaying a video from a path, useful with LTX Review's review_path output.

• Input: review_path (supports absolute paths and optional [input]/[output]/[temp] annotations)
• Behavior: Resolves the path and displays the video preview when executed
• No pass-through outputs: This node is display-only

Load Checkpoint+

Checkpoint loader with a searchable, grouped dropdown designed for large model libraries.

• Text filter: Supports comma-separated terms (example: sdxl, anime)
• AND matching: All typed terms must match in model name or path
• Grouped list UI: Displays a flat list with section headers (example: ----Anime----)
• Clean labels: Hides common extensions in the dropdown display (.safetensors, .ckpt, .pt, .bin, .pth)
• Safe header selection: Selecting a section header auto-selects the first model in that section
• Real path output: Node still passes the real Comfy model path internally
• Outputs: MODEL, CLIP, VAE, ckpt_name (COMBO)

Load Diffusion Model+

Diffusion/UNET loader with the same searchable grouped dropdown UX as Load Checkpoint+.

• Text filter: Supports comma-separated terms (example: flux, fp8, wan, low)
• AND matching: All typed terms must match in model name or path
• Grouped list UI: Displays a flat list with section headers (example: ----Wan22----)
• Clean labels: Hides common extensions in the dropdown display (.safetensors, .ckpt, .pt, .bin, .pth)
• Safe header selection: Selecting a section header auto-selects the first model in that section
• Real path output: Node still passes the real Comfy model path internally
• Outputs: MODEL, unet_name (COMBO)

Load LoRA+

LoRA loader/applier with the same searchable grouped dropdown UX as the other Load Model+ nodes.

• Text filter: Supports comma-separated terms (example: ltx, turbo)
• AND matching: All typed terms must match in LoRA name or path
• Grouped list UI: Displays a flat list with section headers
• Clean labels: Hides common extensions in the dropdown display (.safetensors, .ckpt, .pt, .bin, .pth)
• Apply strength: strength_model controls how strongly the selected LoRA is applied to connected MODEL
• Selector mode: Works without a connected model and still outputs selected lora_name for downstream nodes
• Real path output: Node still resolves and uses the real Comfy LoRA path internally when applying
• Outputs: model, lora_name (COMBO)

Visual LoRA Picker+

Pick LoRAs by thumbnail image instead of scrolling through filenames.

• Image browser: Opens a thumbnail file browser pointing at your LoRA folder; shows matching images (webp, jpg, png) that share the same path/name as the LoRA
• LoRA resolution: Selecting an image resolves the corresponding .safetensors LoRA automatically
• Stack friendly: Accepts an incoming LORA_STACK and appends the picked LoRA, or returns a single-entry stack
• Numeric strength: Quick numeric input sets both model and clip strength
• Outputs: LORA_STACK containing the picked (path, model_strength, clip_strength) tuple

Crop Image+

Interactive crop node with draggable crop box, optional aspect-ratio lock, and live preview.

• Interactive crop UI: Drag the crop box and handles directly in-node
• Aspect ratio lock: Includes ratio presets with landscape toggle
• Auto reset on new image: New input image resets crop to full frame and ratio to None
• Dual output: Crops both IMAGE and MASK outputs together

Clip Stitcher

Merge multiple video clips into one with optional cross-dissolve transitions. Pure blending, no generation.

• File browser: Multi-select clips from input/output or anywhere on disk; remembers last browsed folder
• Reorderable clip list: Drag-to-reorder, enable/disable individual clips, hover to preview
• Clip duration: Force fixed segment lengths — useful when source clips vary in length. Clip 0 plays 0 → clip_duration; each later clip overlaps the previous by blend_duration (e.g. 5s duration + 1s blend = segments start at 0, 4, 9, 14…). Short clips are padded with black
• Blend duration: Seconds of cross-dissolve between clips (0 = hard cut). Audio is crossfaded to match
• Output: 10-bit VIDEO via ComfyUI's CreateVideo

Crop Audio+

Trim an incoming AUDIO connection between in/out points. Minimal node — just the two sliders. (Usefull if you need to trim audio in a workflow, from incoming Video Components)

• Inputs: audio, in_point, out_point (seconds; 0 = full length)
• Outputs: audio (trimmed), duration (trimmed length in seconds)

VACE Stitcher

Generate smooth AI-powered transitions between video clips using VACE conditioning and 2-stage sampling with a single node featuring a built-in clip browser, drag-to-reorder list, and cached .latent transitions for resumability.

• File browser modal: Browse input/output folders, multi-select clips, subfolder navigation
• Reorderable clip list: Drag-to-reorder, enable/disable individual clips, hover thumbnails
• 2-stage sampling: High-noise + low-noise models for quality transitions
• Pixel-space stitching: Crossfade with easing curves, optional color matching
• Lossless latent support: Load .latent clips directly — skips lossy video decode, with memory-efficient on-demand decoding
• Cached transitions: Transitions cached as lossless .latent files — skip already-generated pairs on re-run
• Options node: Connect a separate "VACE Stitcher Options" node to tune all parameters, or use sensible defaults

Inspired by Bob's Wan VACE Clip Joiner workflow on CivitAI.

How to use VACE Stitcher

An example workflow can be found .... If starting from scratch here are some basic instructions.

Required Models

You need both the high-noise and low-noise Wan 2.2 VACE models. Choose one format:

bf16 or fp8 (from Comfy-Org):

• wan2.2_vace_i2v_high_noise_14B-*.safetensors

or GGUF (from QuantStack):

• Wan2.2-VACE-Fun-A14B-*.gguf (high + low noise variants)

Place models in /models/diffusion_models/ or /models/unet if GGUF

Tips

• Use a Wan 2.2 i2v‑distilled LoRA to lower the required step count.
• Lossless clips: Save your source clips with Save Video+ using the "Save Latent" option. VACE Stitcher will automatically detect and use the .latent file for lossless quality — no video decode needed. A magenta dot in the clip list indicates which clips have a latent available. For now only Wan latents are supported.
• First run generates and caches all transitions. Subsequent runs skip cached pairs.
• Delete Transitions clears the cache so you can regenerate with different settings.
• Without an Options node connected, the seed is random each run — just delete transitions and re-queue for a new result.
• Connect a VACE Stitcher Options node to control context/replace frames, steps, sampler, crossfade, color matching, and more.
• Disable clips in the list (uncheck) to skip them without removing.

Load Latent File

Load a .latent file saved by Save Video+. Companion node for video+latent workflows.

Get Video Components+

Memory-efficient video inspector/extractor for VIDEO inputs.

• Metadata-first by default (does not decode frames/audio/latent unless those outputs are connected)
• Internal chunked frame decoding uses a fixed chunk size of 128 to reduce memory needs.
• Outputs: images, audio, fps, filepath, latent, width, height, frame_count, duration
• Automatically loads matching .latent file if one exists beside the video

Audio Mono to Stereo

Convert mono audio to stereo by duplicating the channel. Useful for video models that output mono audio.

Switch Any

Universal switch with up to 10 named inputs. True lazy evaluation — only the selected input is evaluated. Other inputs are completely ignored by ComfyUI, with zero performance cost from inactive branches.

• Custom names via comma/semicolon-separated list
• Dynamic input count (1-10)
• Names display on input slots

Switch Any (Boolean)

Boolean switch — passes through on_true or on_false based on a condition toggle. Only the active branch is evaluated.

Apply LoRA+

Apply a LORA_STACK (list of LoRA tuples) to a model and optional CLIP. Works with Prompt Manager Advanced's LoRA stack output.

LoRA List+

Build a reorderable, toggleable list of LoRAs for quick testing workflows.

• Browse folders and add .safetensors files from disk
• Enable/disable individual LoRAs and drag to reorder
• Outputs enabled entries as newline-separated full paths
• Also outputs basenames (filename-only, extension removed), one per line

Show as Text

Display any input as text in-node, with values persisted in the workflow.

• Accepts any input type (*) and converts values to string internally
• Useful for inspecting intermediate values without extra conversion nodes
• Persists displayed text across workflow reloads and tab switching

Missing Model Remap Tool

Adds a top-bar button in ComfyUI: Remap Missing Models.

When clicked, it scans the current graph and tries to remap missing model-like widget values.

Matching order:

1. Exact basename match (case-insensitive)
2. Precision-token equivalent fallback (fp8/fp16/bf16)
3. Version variations (v1.1, v1_2, v2)

Example:

• workflow value: wan21.safetensors
• local file exists as: wan/wan21.safetensors
• tool remaps node widget to: wan/wan21.safetensors

Supported categories in v1:

• checkpoints/unet
• loras
• text encoders/clip
• vae (including audio-vae related widget keys)
• controlnet
• upscale models (including latent upscale related widget keys)
• Power Lora Loader (rgthree)

Rules and limits:

• Exact filename match first (case-insensitive)
• If exact fails, tries precision-token equivalents (fp8/fp16/bf16)
• If duplicate basenames exist in a category, result is marked ambiguous and skipped
• Updates only the current in-memory graph (no workflow file overwrite)
• Rejects unsafe values (absolute paths or traversal-like paths)

Animated Latent Preview

Provides animated video previews during KSampler execution for video models (Wan, LTX, MiniMax). Compatible with VideoHelperSuite — automatically defers if VHS is installed.

The preview shows a configurable leading slice of the video, re-decoded every sampling step so you watch that segment progressively sharpen as denoising progresses.

Enable in ComfyUI Settings: FBnodes > Video Sampling > Animated Latent Preview

True-color preview injection (MiniMax / LTX)

ComfyUI's latent previewer doesn't have a TAESD decoder mapped for some models, so they fall back to the lower-quality Latent2RGB approximation. FBnodes injects TAESD support for those models so they show true-color previews. You can toggle each model's injection or set a custom TAESD filename in ComfyUI Settings: FBnodes > Preview Injection

Download the TAESD file for each model you use and place it in ComfyUI/models/vae_approx/:

• MiniMax H3 — taeh3.safetensors
• LTX — taeltx2_3.safetensors

Then refresh ComfyUI. If a file is missing, that model keeps the standard preview and a message with the download link is printed to the console.

Installation

ComfyUI Manager

Search for "ComfyUI-FBnodes" in ComfyUI Manager.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/FranckyB/ComfyUI-FBnodes.git
pip install -r ComfyUI-FBnodes/requirements.txt

Requirements

• av (PyAV) — for video encoding

License

GPL-3.0

Changelog

version 1.4.0

• Load Model+: Added Load Checkpoint, Diffusion model and LoRA with filtering system. Allowing, for example, to display only SDXL models.
• Bug Fixes: Multiple big fixes added over time. Including an issue with VACE Stitcher that would fail if clip size wasn't divisible by 32

version 1.3.1

• LoRA List+: Utility node for testing LoRAs, allows adding LoRAs to a list from anywhere on disk. Used with Outputlists-Combiner

version 1.3.0

• Crop Image+: Added a new interactive crop node.

version 1.2.00

• Repath Models Added a new repath utility. Allowing for 1 click repathing of all models and loras. Uses minimal fuzzy logic to find models with different quantizations or versions.
• Show as Text Added a simple Show Text node, that unlike Preview as Text, is saved with Workflow and is maintained when switching tabs.

version 1.1.12

• Load Audio+: Added a load audio node, similar to other loaders. Has the ability to trim the In and Out point.
• Polish: Tweaked the UIs of Load Image+ and Load Video+, matching tweaks that were done to Load Audio+

version 1.1.11

• Apply LoRA+: Added a simple "apply LoRA Stack" node, with a strength multiplier

version 1.1.10

• Load Image+ Mask Support: Added MASK output with alpha channel extraction from images
• Mask Editor Integration: MaskEditor now works correctly — painted masks display in the node preview with transparency
• Fix: Handle MaskEditor's annotated filepath format (file.png [input]) so masked images load properly

version 1.1.9

• Renamed nodes: cleaned up internal node IDs (SaveVideoH26x → SaveVideoPlus, PromptApplyLora → ApplyLoraPlus, BetterImageLoader → LoadImagePlus). Existing workflows will use "Legacy" versions for backward compatibility.
• Fix Save Video+ display issues: Fixed issues found in Load Video+.

version 1.1.8

• Speed Improvement: Thumbnail generation now uses server-side PyAV instead of browser-based video decoding — substantially faster, especially with many clips.
• Improved Handling of H265 Clips
• Renamed nodes: Save Video H264/H265 → Save Video+, Better Image Loader → Load Image+. Existing workflows are unaffected.
• New node: Added Load Video+ — video loader with file browser, preview, and drag-drop. Outputs VIDEO for Get Video Components+.

version 1.1.5

• VACE Stitcher: Added lossless .latent file support for clips and transitions
  • Clips with a .latent file alongside are loaded in latent space, avoiding lossy video decode
  • Transitions now cached as .latent files instead of h265 video — lossless quality
  • Memory-efficient on-demand decoding: only decode what's needed, when it's needed
  • Latent indicator (magenta dot) in clip list UI shows which clips have .latent files

version 1.1.1

• Renamed node from "VACE Transition Builder" to VACE Stitcher
• Added Workflow example.

version 1.1.0

• Added VACE Stitcher node — generates smooth AI transitions between video clips using VACE conditioning
• Added VACE Stitcher Options node — optional parameter overrides for the stitcher
• Credit to Bob for the original workflow

version 1.0.0

• Initial release
• Transferred miscellaneous from ComfyUI-Prompt-Manager to keep it more focused.