ComfyUI-FBnodes Introduction
ComfyUI-FBnodes is a collection of extra nodes and interface tools for ComfyUI. The author built them for practical image, video, audio, and model-selection workflows. You can use individual tools where they help; you do not need to build a workflow around the entire collection.
The extension helps with tasks that can otherwise interrupt an artistic workflow: finding a model in a large library, choosing a LoRA by its thumbnail, cropping an image inside a node, reviewing a video before a second processing pass, or joining clips into a finished sequence. It also includes saving tools, audio trimming, workflow switches, and previews that make it easier to see what is happening while a workflow runs.
How ComfyUI-FBnodes Works
A ComfyUI workflow passes information between connected nodes. FBnodes adds nodes that handle particular steps in that flow. For example, Load Image+ supplies an image and mask, Crop Image+ adjusts their framing together, and Save Image+ writes the result to a file.
Some FBnodes tools work at the interface level instead. Remap Missing Models checks the model selections in your current graph when you click its button. Animated Latent Preview displays an evolving preview while a compatible video workflow samples.
It helps to distinguish three common kinds of video operation:
- Clip Stitcher joins existing clips by cutting or blending them. It does not generate new frames with an AI model.
- VACE Stitcher uses models to generate transitions between clips, then assembles the sequence.
- LTX Review pauses a two-pass workflow so you can inspect an intermediate result and decide whether to continue.
A latent is an intermediate representation used by a generation workflow. Where supported, saving or reusing a latent can avoid converting a clip to a conventional video and back. You do not need to edit the latent yourself to benefit from that option.
ComfyUI-FBnodes Features
Saving images and video
Save Video+ saves video using H.264, an 8-bit choice aimed at broad compatibility, or H.265/HEVC, a 10-bit choice with better compression. You can select yuv420, yuv422, or yuv444 chroma, and adjust CRF to control quality: lower values mean higher quality and generally larger files; 0 is lossless, while 18–23 is described as a high-quality range. The node can combine audio with video, embed workflow metadata, and optionally save a matching latent for later use. A preview-only mode skips the normal save. If you select yuv444, it also makes a temporary preview clip that can be viewed in a browser.
Save Image+ provides date-token filenames, a choice between output and temporary saving, and an optional Compare image input. Its additional settings sit under Controls by default.
- PNG: Save an 8-bit image, the default, or a 16-bit image from sRGB input.
- EXR: Save 16-bit or 32-bit floating-point files from sRGB, HDR (HLG), or linear input. The saved file uses scene-linear values.
- AVIF: Choose Auto, 8-bit YUV420, or 10-bit YUV420, with sRGB, HDR (HLG), or HDR PQ input and CRF quality control. The node saves one still image for each batch item. If the input has transparency, its alpha channel is removed for AVIF saving and preview without compositing it onto a background.
EXR and AVIF previews are separate, temporary 8-bit PNGs made from the input. They are useful for inspection but are not color-managed HDR previews and do not change the saved image. Newer formats require a ComfyUI version with Save Image (Advanced); the original 8-bit PNG option remains available on older versions. FBnodes also lets ComfyUI import a workflow by dropping an EXR that contains workflow or prompt header metadata.
Loading and adjusting media
Load Image+ lets you browse input or output folders, navigate subfolders through thumbnail previews, or drop an image or video onto the node. For video files, a frame-position slider lets you choose a screenshot. You can enlarge the preview, make a quick crop, and flip it horizontally or vertically. The node outputs both IMAGE and MASK; it can extract a mask from an image’s alpha channel, and painted masks from ComfyUI’s Mask Editor appear in its preview. Use Crop Image+ when you need more precise framing controls. Load Image+ is based on Prompt Extractor from Prompt Manager.
Load Video+ has a similar browser, drag-and-drop behavior, and an enlarged preview with playback controls. It outputs a VIDEO connection that you can send to Get Video Components+.
Load Audio+ offers an audio-focused browser and a trimming timeline. Drag the in and out handles, scrub with the playhead, or loop the selected clip while deciding what to keep. It outputs the trimmed AUDIO and its duration in seconds.
Crop Image+ displays a draggable crop box and a live preview. You can lock it to a preset aspect ratio and switch between portrait and landscape orientation. It crops connected IMAGE and MASK outputs together. When a new image arrives, the crop resets to the full frame and the ratio resets to None.
Crop Audio+ trims audio already arriving from another node, such as audio extracted from a video. Set in_point and out_point in seconds; 0 means the full length for that boundary. It returns the trimmed audio and its duration.
Get Video Components+ inspects a connected VIDEO and can output images, audio, fps, filepath, latent, width, height, frame_count, and duration. By default, it checks metadata first and does not decode frames, audio, or a latent unless the corresponding output is connected. Frame decoding uses chunks to reduce memory demands. If a matching .latent file sits beside the video, the node loads it automatically.
Load Latent File opens a .latent file saved by Save Video+. It is the companion loader for workflows that keep both a video and its latent.
Audio Mono to Stereo duplicates a mono audio channel to make stereo audio. This is useful when a video model produces mono output.
Choosing models and LoRAs
Load Checkpoint+, Load Diffusion Model+, and Load LoRA+ provide searchable, grouped selections for large libraries. Enter comma-separated search terms to narrow the list: every term must match the file name or path. For example, sdxl, anime looks for entries matching both terms. Section headers organize the results, and common file extensions are hidden from the displayed labels. The nodes still use the real ComfyUI paths behind the scenes.
- Load Checkpoint+ outputs
MODEL,CLIP,VAE, and the selectedckpt_name. - Load Diffusion Model+ outputs
MODELandunet_name. - Load LoRA+ applies a selected LoRA to a connected model using
strength_model, and outputs the model andlora_name. You can also leave the model disconnected and use it solely to select a LoRA name for a downstream node.
Visual LoRA Picker+ lets you choose a LoRA using an image rather than a filename. Its browser looks for a webp, jpg, or png image with the same path and name as the corresponding .safetensors file. Selecting the image resolves that LoRA. You can set a numeric strength for both model and CLIP, then output a LORA_STACK. If a stack is already connected, the picker adds its selection to it.
LoRA List+ is useful for comparing combinations. Browse for .safetensors files, drag to reorder them, and toggle individual entries on or off. It outputs enabled full paths as newline-separated text and, separately, filenames without extensions as basenames.
Apply LoRA+ applies a LORA_STACK to a model and optionally to CLIP. It works with Prompt Manager Advanced’s LoRA stack output.
Remap Missing Models is a top-bar button for workflows whose saved model paths no longer match your local folders. It checks the current graph for missing selections across checkpoints, diffusion models, LoRAs, text encoders, VAEs, ControlNet, upscale models, and supported related widgets. It tries an exact filename match first, then precision-name variations such as fp8/fp16/bf16, and then version-name variations. For example, it can change a selection of wan21.safetensors to wan/wan21.safetensors when that is the matching local path. Ambiguous matches are skipped. Changes affect the current graph in memory, not the workflow file on disk.
Building and reviewing video sequences
Clip Stitcher joins selected clips without generating new content. Browse and select multiple files, reorder them by dragging, disable clips without removing them, and preview them from the list. Set blend_duration to 0 for a hard cut or increase it for a cross-dissolve with matching audio crossfade. You can also set a fixed clip_duration; shorter sources are padded with black. Its output is a 10-bit VIDEO through ComfyUI’s CreateVideo.
VACE Stitcher generates transitions between clips using VACE conditioning and two sampling stages, one high-noise and one low-noise. Its browser supports multi-selection and subfolders; the clip list supports reordering, disabling, and thumbnail previews. The node can crossfade assembled clips with easing curves and optional color matching. A connected VACE Stitcher Options node exposes settings such as context and replacement frames, sampling steps, sampler, crossfade, and color matching; without it, the stitcher uses defaults.
VACE Stitcher caches generated transitions as .latent files. The first run generates them; later runs can skip pairs already cached. Use Delete Transitions when you want those pairs regenerated after changing settings. It can also use supported source .latent files directly, avoiding a lossy video decode. For now, the source latent support described by the author is for Wan latents. A magenta dot in the clip list marks clips with an available latent.
[VACE Stitcher interface](docs/vace_stitcher.png)
The node was inspired by Bob’s Wan VACE Clip Joiner workflow.
LTX Review is a decision point for a two-pass LTX-style workflow. It decodes connected video and audio latents, creates a temporary browser-compatible H.264/yuv420 preview, and opens a popup. Choose Proceed to continue, Cancel to stop, or Requeue. Its timeout and on_timeout settings determine whether it proceeds or cancels if you do not respond. Turning enable off passes the latents through without decoding or opening the popup. An optional preview_video can preserve a review_path for preview workflows while review is disabled. The node outputs video_latent, audio_latent, and the temporary preview’s review_path.
LTX Review Preview displays the video at a connected review_path. It is display-only and has no pass-through output.
Directing and inspecting a workflow
Switch Any selects one of up to 10 inputs. You can give the inputs names using a comma- or semicolon-separated list and change how many slots are shown. Only the selected branch is evaluated, so inactive branches do not incur their processing cost.
Switch Any (Boolean) chooses between on_true and on_false using a boolean condition. Again, only the active branch is evaluated.
If Any turns the presence of a truthy connected value into a boolean. A disconnected input, None, an empty value, or 0 returns false; a truthy connected value returns true. It can help a workflow decide what to do when an optional image, latent, or text value is present.
Show as Text displays a connected value inside the node. The displayed text is preserved in the workflow across reloads and tab switches, which makes it useful for checking intermediate values.
Animated Latent Preview shows a configurable leading portion of a video as compatible video models sample. That portion is decoded again at each sampling step, letting you watch it take shape. Enable it in ComfyUI Settings → FBnodes → Video Sampling → Animated Latent Preview. It supports the video-model workflows described by the author for Wan, LTX, and MiniMax, and defers automatically when VideoHelperSuite provides its own preview.
FBnodes also offers true-color preview injection for MiniMax H3, LTX, and Krea2. With a suitable TAESD preview decoder, these models can show a better color preview than the fallback approximation. Settings for each are under FBnodes → Preview Injection. These settings affect the sampling preview, not the final saved image or video.
ComfyUI-FBnodes Models
Most FBnodes tools do not supply or require a particular generation model. They operate on the images, audio, video, models, or LoRAs already in your ComfyUI workflow. The model-related choices explicitly described by the author are:
| Use | Model or file | What it changes |
|---|---|---|
| VACE Stitcher transitions | Both high-noise and low-noise Wan 2.2 VACE models | The two stages generate the transition between clips. Both stages are required. |
| Optional VACE workflow adjustment | Wan 2.2 i2v-distilled LoRA | The author suggests it to lower the required sampling step count. |
| MiniMax H3 preview | taeh3.safetensors | Enables the described true-color latent preview for MiniMax H3. |
| LTX preview | taeltx2_3.safetensors | Enables the described true-color latent preview for LTX. |
| Krea2 preview | TAE_Krea2_AcademiaSD.safetensors | Enables the described true-color latent preview for Krea2. |
For VACE Stitcher, the author lists Comfy-Org’s Wan 2.2 VACE diffusion models in bf16 or fp8 format, or QuantStack’s Wan 2.2 VACE GGUF models with high- and low-noise variants. These are alternative formats for the required pair, not four separate stages.
The TAESD files are preview decoders, not replacements for your main generation models. Place the decoder you use in ComfyUI/models/vae_approx/ and refresh ComfyUI. For Krea2, you can leave its filename setting blank for the default or specify another compatible decoder filename, including a path within a subfolder of models/vae_approx. The author notes that Krea2 preview injection does not change Wan previews.
What's New with ComfyUI-FBnodes
The original changelog records these releases, newest first:
| Version | Change | Why it matters to you |
|---|---|---|
| 1.4.0 | Added searchable Load Checkpoint+, Load Diffusion Model+, and Load LoRA+ nodes; included bug fixes, including a VACE Stitcher fix for clip dimensions not divisible by 32. | Easier model browsing and more reliable stitching with varied clip sizes. |
| 1.3.1 | Added LoRA List+ for selecting and testing LoRAs, including use with Outputlists-Combiner. | Faster comparison of LoRA combinations. |
| 1.3.0 | Added Crop Image+. | Frame an image interactively inside your workflow. |
| 1.2.00 | Added the model repathing utility and Show as Text. | Recover missing model selections more quickly and retain displayed intermediate text in a workflow. |
| 1.1.12 | Added Load Audio+ with in/out trimming; refined the Load Image+ and Load Video+ interfaces. | Easier audio selection and more consistent media browsing. |
| 1.1.11 | Added Apply LoRA+ for applying a LoRA stack with a strength multiplier. | More convenient use of stacked LoRAs. |
| 1.1.10 | Added mask output and alpha extraction to Load Image+; improved Mask Editor previews and handling of annotated file paths. | Painted and image-derived masks work more predictably. |
| 1.1.9 | Cleaned up internal node IDs and retained legacy versions for existing workflows; addressed display issues involving the video nodes. | Older workflows remain compatible after the naming changes. |
| 1.1.8 | Added Load Video+, renamed the image and video nodes to their “+” names, sped up thumbnails using server-side PyAV, and improved H.265 clip handling. | Quicker browsing, especially in folders with many clips. |
| 1.1.5 | Added lossless .latent source clips and cached transitions to VACE Stitcher, with on-demand decoding and a latent indicator. | Better preservation of supported source clips and faster repeat runs. |
| 1.1.1 | Renamed VACE Transition Builder to VACE Stitcher and added a workflow example. | Clearer identification of the video-joining tool. |
| 1.1.0 | Added VACE Stitcher and its optional Options node. | Introduced AI-generated clip transitions. |
| 1.0.0 | Initial release; moved miscellaneous tools from ComfyUI-Prompt-Manager. | Established FBnodes as a separate collection of workflow tools. |
Troubleshooting ComfyUI-FBnodes
A workflow opens, but one or more models are missing.
- Click Remap Missing Models in the top bar.
- Check whether the affected selections now point to your local models.
- If an item remains missing, inspect the available files in that model category and select the intended one yourself. The tool skips ambiguous duplicate filenames.
- Save the workflow through ComfyUI if you want to retain the corrected graph; remapping itself does not overwrite the workflow file.
My LoRA does not appear as an image in Visual LoRA Picker+.
- Check that the LoRA is a
.safetensorsfile in the browsed location. - Check that its thumbnail is a
webp,jpg, orpngwith the same name and path as the LoRA. - Browse to that folder again and select the matching image.
VACE Stitcher regenerates transitions, or I want a different result.
- Check whether you are rerunning the same clip pairs with cached
.latenttransitions available. - To force new transitions after changing settings, use Delete Transitions, then queue the workflow again.
- If you need a repeatable seed rather than the default random seed, connect VACE Stitcher Options and set it there.
A source clip has no magenta latent indicator in VACE Stitcher.
- Check whether a matching
.latentfile is available alongside the clip. - For a new source, use Save Video+ with Save Latent enabled where your workflow supports it.
- Keep in mind that the source latent support described by the author currently covers Wan latents.
The sampling preview has poor color or does not improve after enabling preview injection.
- Check that you selected the appropriate MiniMax H3, LTX, or Krea2 preview option under FBnodes → Preview Injection.
- Check that the matching TAESD file is in
ComfyUI/models/vae_approx/. If it is missing, FBnodes keeps the standard preview and prints a download link to the console. - Refresh ComfyUI after adding a decoder. For Krea2 support added by an extension update, restart ComfyUI.
- For Krea2 image previews, enable Display animated or subgraph previews when sampling and choose ComfyUI’s TAESD or Auto preview method.
An EXR or AVIF preview does not look like my HDR file.
Save Image+ creates a separate 8-bit PNG preview from the input for these formats. It is not a color-managed HDR display of the saved file. Judge the saved EXR or AVIF in a viewer suited to that format.
Frequently asked questions
Does Clip Stitcher create new AI-generated transitions?
No. It joins existing footage with hard cuts or cross-dissolves. Choose VACE Stitcher when you want model-generated transitions.
Does turning off LTX Review remove the latent from my workflow?
No. With enable off, the node passes the video and audio latents through without performing its review decodes or showing the popup.
Will Switch Any run every connected branch before choosing one?
No. It evaluates only the selected input. Switch Any (Boolean) likewise evaluates only its active branch.
Do I need every preview decoder listed above?
No. Use the file corresponding to the model whose preview you want to improve.
Will Remap Missing Models change my saved workflow automatically?
No. It updates the current in-memory graph, not the workflow file on disk.
Learn More about ComfyUI-FBnodes
- VACE Stitcher interface image: See the stitcher’s clip-oriented interface.
- Wan VACE Clip Joiner workflow: Explore the workflow that inspired the author’s VACE Stitcher.
- ComfyUI-Prompt-Manager: Learn about the related project behind Load Image+’s browser inspiration and the LoRA-stack workflow mentioned above.
- Comfy-Org Wan 2.2 VACE models and QuantStack Wan 2.2 VACE GGUF models: Review the model-format choices described for VACE Stitcher.
