FM_nodes Introduction
FM_nodes is a collection of custom nodes for ComfyUI. It brings several image and video research projects into visual workflows, so you can use them for tasks such as restoring faces, brightening dark images, cleaning up weather effects, improving video detail, creating intermediate frames, and stitching or stabilizing footage. The author provides example workflows for all seven project families. (github.com)
Choose a tool based on the problem you see in your work:
| Your goal | FM_nodes option |
|---|---|
| Improve a small or soft face | WFEN |
| Upscale real-world video | RealViformer |
| Help an added element fit a painterly image | ProPIH |
| Brighten a low-light image or sequence | CoLIE |
| Add frames between existing video frames | VFIMamba |
| Remove rain, haze, snow, defocus, or blur | ConvIR |
| Join video views or steady stitched footage | StabStitch |
These are separate tools, not stages you must run in order. Start with the example workflow closest to your goal, then compare its result with your original image or footage.
How FM_nodes Works
A node is a step in a ComfyUI workflow. Think of the workflow as an editing desk: one step supplies an image or frames, another applies a treatment, and a final step lets you inspect or save the result. FM_nodes supplies specialized treatment steps based on the projects listed below.
Most of these tools use a model file: a set of learned information that the node reads when it runs. A workflow file, usually ending in .json, is different. It is an example arrangement of connected nodes, not the model itself. CoLIE is the exception noted by the author: its example does not require a separate model download. (github.com)
For video, the distinction between upscaling and interpolation is useful. RealViformer aims to improve the detail of existing frames. VFIMamba creates frames between existing ones, which can make motion appear smoother. StabStitch addresses a different problem: joining views and reducing visible shaking in stitched footage.
FM_nodes Features
WFEN: Face Restoration
WFEN focuses on improving facial detail. Try it when a face occupies only a small part of an image or looks soft after enlargement. The supplied face-crop workflow shows how the author uses it.
WFEN is based on Efficient Face Super-Resolution via Wavelet-based Feature Enhancement Network. The model file expected by the example is models/wfen/WFEN.pth. (github.com)
RealViformer: Video Upscaling
RealViformer is intended to increase the apparent detail of real-world video. Use it when you want a larger, clearer version of existing footage rather than additional frames. Results will depend on the source material; inspect moving details as well as individual still frames.
- Example workflow
- Workflow image
- Video comparison, shown separately from the workflow image
The example expects models/realviformer/weights.pth. Read more about the underlying project in Investigating Attention for Real-World Video Super-Resolution. (github.com)
ProPIH: Painterly Image Harmonization
ProPIH is for making a composited element look more at home in a painterly image. For example, when you add a subject to an illustration, harmonization can help the added subject better match the image’s visual style. It is about visual integration, not simply increasing resolution.
- Example workflow:
workflow_propih.jsonin the repository’sworkflowfolder - Workflow image
- Underlying ProPIH project
The author lists two required files: models/propih/vgg_normalised.pth and models/propih/latest_net_G.pth. Note: the original README labels a link “workflow_propih.json,” but that link points to the RealViformer workflow. Check the workflow folder for the ProPIH example rather than using that mislabeled link. (github.com)
CoLIE: Low-Light Enhancement
CoLIE brightens dark images and can also work with batched images or video. It is a useful starting point when details are difficult to see because a scene was captured or created with very little light.
The author notes that a lower loss_mean appears to produce brighter results. Treat that as a practical observation, not a guarantee: try a small change, compare the output, and stop when the scene looks suitably lit. No separate model download is listed for CoLIE.
VFIMamba: Video Frame Interpolation
VFIMamba generates frames between existing video frames. For example, you might use interpolation when you want motion to appear smoother while keeping the original frames as the starting points. The author’s video example compares an original clip with a result labeled 5×, 20 FPS; that label describes the example, not a promised result for every video.
Two model files are listed: models/vfimamba/VFIMamba_S.pkl and models/vfimamba/VFIMamba.pkl. VFIMamba also needs additional software dependencies. The author specifically warns that mamba-ssm has no prebuilt Windows binary in the setup described in the original README; see the Mamba project and the author’s Windows build resources when diagnosing a dependency error. Those build resources specify Python 3.11 and PyTorch 2.3.0+cu121, so they should not be assumed compatible with every ComfyUI environment. (github.com)
ConvIR: Image Restoration
ConvIR offers models for several kinds of visible damage or obstruction. Choose the task that matches your image rather than trying every model on every picture:
- Deraining: rain obscures the scene.
- Defocus: parts of the image appear out of focus.
- Dehaze: haze reduces clarity and contrast.
- Desnow: snow obscures image details.
- Motion blur: movement has smeared details.
The available checkpoints include task and dataset names such as haze4k-small.pkl or snow100k-base.pkl. These names identify different pretrained options; the supplied README does not establish which one will look best on your images. Compare a few relevant outputs at the size you intend to use.
StabStitch: Stitching and Stabilization
StabStitch provides examples for stitching video views and stabilizing the result. The author recommends using StabStitch Crop Resize and notes that inputs are resized to 360 × 480. Keep that resizing in mind when judging fine detail or preparing footage for a larger final composition.
Dark edges can appear around a warped result. The author suggests setting interpolate_mode to NORMAL or setting do_linear_blend to True to eliminate them. Try one option and review the borders before changing anything else.
| Task | Example |
|---|---|
| Stitch footage | StabStitch workflow · workflow image |
| Stabilize footage | StabStitch Stabilize workflow · workflow image |
Example input videos are noted in examples/stabstitch. The workflow images above are separate examples, not images embedded in the workflow files. The three listed model files are temporal_warp.pth, spatial_warp.pth, and smooth_warp.pth, all under models/stabstitch/. See the underlying StabStitch project. (github.com)
FM_nodes Models
The names below are checkpoint files used by different tools, not interchangeable versions of one FM_nodes model. A filename often indicates its task or source dataset; it does not, by itself, guarantee a particular visual result or speed. (github.com)
| Tool | Models listed by the author | When to use them |
|---|---|---|
| WFEN | models/wfen/WFEN.pth | Face restoration |
| RealViformer | models/realviformer/weights.pth | Video upscaling |
| ProPIH | models/propih/vgg_normalised.pth, models/propih/latest_net_G.pth | Painterly image harmonization; both files are listed |
| CoLIE | No separate model download listed | Low-light enhancement |
| VFIMamba | models/vfimamba/VFIMamba_S.pkl, models/vfimamba/VFIMamba.pkl | Video frame interpolation; the supplied context does not document a visual or performance comparison between these two files |
| StabStitch | models/stabstitch/temporal_warp.pth, spatial_warp.pth, smooth_warp.pth | Stitching and stabilization; all three are listed |
ConvIR checkpoints are grouped by restoration task. Use the model list to identify the appropriate family, then follow the project’s pretrained-model section for the files:
| Task | Checkpoints listed in the original README |
|---|---|
| Deraining | deraining.pkl |
| Defocus | defocus/dpdd-base.pkl, dpdd-large.pkl, dpdd-small.pkl |
| Dehaze | dehaze/densehaze-base.pkl, densehaze-small.pkl, gta5-base.pkl, gta5-small.pkl, haze4k-base.pkl, haze4k-large.pkl, haze4k-small.pkl, ihaze-base.pkl, ihaze-small.pkl, its-base.pkl, its-small.pkl, nhhaze-base.pkl, nhhaze-small.pkl, nhr-base.pkl, nhr-small.pkl, ohaze-base.pkl, ohaze-small.pkl, ots-base.pkl, ots-small.pkl |
| Desnow | desnow/csd-base.pkl, csd-small.pkl, snow100k-base.pkl, snow100k-small.pkl, srrs-base.pkl, srrs-small.pkl |
| Motion blur | modeblur/convir_gopro.pkl, convir_rsblur.pkl |
The ConvIR paths in the table are relative to models/convir/. For example, the full path for haze4k-small.pkl is models/convir/dehaze/haze4k-small.pkl.
Troubleshooting FM_nodes
A workflow says a model is missing.
- Identify which tool the workflow uses.
- Compare the required filename and folder with the FM_nodes Models section. Check the complete name, including
.pthor.pkl. - For tools that list more than one required file, check each file. Do not substitute a checkpoint from a different tool.
- Reopen or rerun the workflow after correcting the files.
The ProPIH workflow link opens RealViformer.
The original README contains a mismatched link. Look for workflow_propih.json in the repository’s workflow folder, and use the ProPIH workflow image as a visual reference. (github.com)
VFIMamba reports a missing dependency, particularly on Windows.
Check the error for mamba-ssm, causal-conv1d, or Triton. These are software dependencies, not missing video frames or model checkpoints. The original README links to the Mamba project, causal-conv1d, Triton, and Windows build resources. Because the author’s provided Windows build targets a specific Python, PyTorch, and CUDA combination, ask someone familiar with your ComfyUI environment to check compatibility before using a build intended for another setup. (github.com)
StabStitch leaves dark borders.
- Try
interpolate_mode = NORMAL. - If borders remain, test
do_linear_blend = True. - Compare the edges of the output and consider the author’s recommended StabStitch Crop Resize node.
Frequently asked questions
- Do I need every model in this README? No. Choose the tool for your task, then check the files listed for that tool.
- Does CoLIE need a separate downloaded checkpoint? The author says no.
- Will lowering CoLIE’s
loss_meanalways improve an image? Not necessarily. The author observes that lower values seem to produce brighter images; judge whether that brightness suits your scene. - Does VFIMamba upscale my video? Its listed purpose is frame interpolation: adding frames between existing frames. For video upscaling, see RealViformer.
- Why does StabStitch change my input dimensions? The author states that its inputs are resized to 360 × 480. Account for that when planning the rest of your workflow. (github.com)
Learn More about FM_nodes
Start with the FM_nodes repository and its example workflows. The linked workflow images throughout this README show what the examples look like; the separate video comparisons help you assess motion and detail that a still image cannot show.
For a deeper explanation of each technique, visit the original projects: WFEN, RealViformer, ProPIH, CoLIE, VFIMamba, ConvIR, and StabStitch. The FM_nodes issue tracker is the repository’s place to check for reported problems or describe a reproducible issue.
