ComfyUI-CoachBate Introduction
ComfyUI-CoachBate is a collection of quality-of-life tools for making and managing AI art workflows in ComfyUI. The author built these tools for daily production work. They are especially useful when you have a large workflow library, generate many variations, or work with video projects that involve multiple shots and processing stages.
The extension has four main areas:
- Workflow organization: Workflows+ helps you find, open, and manage saved workflows, including libraries with thousands of files.
- Repeatable generation: Batch Prompter runs a workflow with a different text value each time, while Shot Loader steps through a planned list of video shots.
- Text, model, and metadata helpers: Preview generated text, repair model paths when opening workflows from another setup, and recover or protect information stored in media files.
- Video refinement and output: Refine subjects in MiniMax H3 clips, load videos with audio, and combine upscaling, frame interpolation, and video encoding.
You do not need to use every tool. For example, you can use Workflows+ solely to organize image workflows without building an H3 video workflow.
How ComfyUI-CoachBate Works
A ComfyUI workflow is a set of connected nodes: one node supplies something, such as an image or prompt, and another uses it. ComfyUI-CoachBate adds nodes and interface tools at different points in that process.
Think of Batch Prompter as a list of instructions for repeated runs. If you enter three non-blank lines, each line can supply a different value to the connected workflow. That value might be an image prompt, but it could also be a filename, a model name, or another piece of text.
Shot Loader works more like a production checklist. It reads a shotlist.json file containing shot prompts and details, skips shots marked DONE, and supplies one shot’s information per run.
Workflow Model Path Auto-Fix helps when you open a workflow made on another ComfyUI setup. If a model widget points to a file under a folder name you do not have, the extension checks the choices available in your setup and updates the widget when it finds a valid match.
For video, H3 Subject Refine follows a selected subject through a clip. It enlarges crops of that subject for a low-denoise MiniMax H3 refinement pass, then places the refined result back into the original frames. This can give a small subject more room to be rendered in detail without regenerating the entire shot.
ComfyUI-CoachBate Features
Workflows+: Find and manage saved workflows
Workflows+ is a separate sidebar tab alongside ComfyUI’s built-in Workflows tab; it does not modify the built-in tab. Its file list and folder tree are designed to remain responsive with very large libraries. The author reports testing it with nearly 8,000 workflows.
- Find by name: A search for
black catmatches that contiguous phrase. Useblack AND catwhen both words can appear separately, orblack OR catwhen either word will do. The operators must be uppercase. - Find by contents: Switch to Contains and press Enter to search inside workflow JSON, including prompts, node titles, and values. Large searches can take time; a progress bar appears, and you can stop the scan with Cancel or Esc.
- Search one folder: Click a folder to limit Name or Contains searches to it. Search subfolders determines whether nested folders are included. Use Show All, or click the selected folder again, to clear the scope.
- Sort and pin: Sort by name, modification date, or creation date. Right-click a workflow to pin it above other files in its folder. A compact Recent menu shows the last 10 workflows you opened.
- Open dropped files: Drop a workflow JSON file, or a PNG or MP4 containing an embedded workflow, onto the panel. JSON and PNG workflow files appear in the file list; clicking a listed PNG opens its embedded workflow.
- Manage files and folders: Rename, move, duplicate, or delete workflows from the panel. Use Ctrl/Cmd-click or Shift-click to select multiple files for bulk actions. You can also rename, move, or delete folders. Deleted items go to
_trashfirst; deleting an item already there removes it permanently.
Workflows+ acts on the workflows folder visible to the ComfyUI server, which matters if you use a remote or cloud-hosted ComfyUI instance. Closing the tab clears its current search, folder scope, and expanded folders; your sort order and Search subfolders preference remain saved. To hide the tab, turn off Enable Workflows+ panel under Settings → CoachBate → Workflows+, then reload ComfyUI.
Workflow Model Path Auto-Fix: Repair missing model selections
A shared workflow may refer to a checkpoint or LoRA stored under a different folder name on your machine. Auto-Fix checks model-type widgets as the workflow loads. If the referenced file is not an exact available choice, it attempts to select a valid local match.
You can also set deliberate Source → Target replacements under Settings → CoachBate → Model Path Replacements. For example, you might redirect a workflow from one version of a model file to another version you prefer. Rules are checked in order, and a replacement is used only if the target is available for that widget. Saved rules take effect without restarting ComfyUI.
If you prefer to review missing models yourself, disable the feature in its settings or under Settings → Extension. Replacements are logged server-side so you can inspect what changed.
Batch Prompter: Run variations from a list
CoachBate Batch Prompter takes a multiline text block and supplies one non-blank line per queue run. Connect its string output wherever your workflow needs a changing value.
For example, you could enter three scene descriptions to render three versions of an image. Alternatively, enter three image filenames and use prepend_text and append_text to turn each bare name into the path your image-loading setup expects.
Use ComfyUI’s normal Queue button to start. By default, queue_all_at_once submits the jobs together as separate workflow snapshots. Turn it off to run them one at a time, so you can review results between jobs. starting_number selects the first prompt by its displayed number, ignoring blank lines, and max_prompts limits how many run. Turn on randomize to use a shuffled order without repeats within the run; with max_prompts = 1, it chooses one eligible line at random.
Only an active Batch Prompter with a connected output drives a batch. If more than one eligible Batch Prompter is present, the workflow falls back to a normal single queue run.
Text Preview and Edit, and Numbered Text
CoachBate Text Preview and Edit displays and passes through a connected string. Connect a value to its any input to inspect dynamically generated text, such as a prompt assembled by other nodes. If the upstream node is muted, it uses the text you typed into the node instead. Its resolved value is saved in workflow metadata, making it easier to see the prompt used when you later reopen a saved image or video. A Copy Text button copies the displayed text.
CoachBate Numbered Text provides a multiline text box with line numbers beside it. It outputs the complete text as a single string, which is useful when you want to read or edit longer text without losing your place.
Shot Loader: Work through a video shot list
CoachBate Shot Loader reads a shotlist.json file and outputs one shot’s prompt and production details per run. It can provide the shot ID, duration, filename prefix, optional start and end images and their strengths, and optional negative prompts.
Set mode to increment to move forward, decrement to move backward, or fixed to keep using the same shot. shot_index sets the initial position. Shots whose status is "DONE" are skipped. The node displays the current and next shot after execution, and warns when a referenced start or end image cannot be found.
For a simple shot list, each shot needs an ID, output filename prefix, duration in seconds, and video prompt:
[
{
"shot_id": "001",
"video_filename_prefix": "001-GYM-ENTRANCE",
"duration_seconds": 16,
"video_prompt": "A cinematic shot of an athlete entering a gym.",
"start_image": "C:/path/to/images/gymdoor.png",
"start_image_strength": 0.85
}
]
The file may instead contain a top-level "shots" array. Optional fields include end_image, end_image_strength, start_image_prompt, negative_prompt, and negative_audio_prompt. Set "status": "DONE" on a shot you want to skip. If every shot is marked DONE, the node reports an error rather than looping indefinitely.
To advance through the list automatically, choose increment, select Auto Queue in ComfyUI’s queue controls, and press Queue once. Check the node’s status as it proceeds.
Video and metadata helpers
- Video Combine wraps Video Helper Suite’s video-combining feature and protects saved metadata by removing API keys. It also cleans up certain filename prefixes and suffixes; its documented default format is H.265 MP4 with CRF 16.
- Strip API Key Metadata removes API-key fields from supported media metadata. It can process one file or a folder containing supported PNG, MP4, MOV, MKV, and WebM files.
- Load Videos With Audio loads videos from a folder together with their audio for editing or saving into a new video.
- Video Prompt From Metadata retrieves the prompt embedded in a ComfyUI-generated MP4’s workflow metadata.
- Lyrics JSON Parser reads a custom lyrics-and-timing JSON format created for this workflow.
For other supported integrations, Settings → CoachBate → Compatibility Patches lets you independently control the Video Helper Suite metadata-protection patch, Gemma API-key environment fallback, and VRGDG Whisper dtype correction. These patches are enabled by default; changing them requires a ComfyUI restart.
H3 Subject Refine: Improve a subject within a video
MiniMax H3 can struggle with a subject that occupies only a small part of a frame. H3 Subject Refine detects a subject in each frame, follows it through the clip, enlarges it into a crop, runs an H3 image-to-image refinement pass, and blends the result back into its original position. It adapts the approach used by ComfyUI-H3-FaceRefine so it can be used for subjects beyond faces.
You can use the H3 Subject Refine blueprint for one selected subject, or H3 Subject Refine xN to count and process multiple subjects, one pass per subject. The underlying nodes handle tracking and cropping, video-latent preparation, per-subject denoise control, segmentation masks, subject counting, looping, and stitching. An example workflow named H3 Subject Refine - Existing Video takes an existing video, recovers its prompt from metadata, refines it, and outputs video with its original audio.
Settings that have a noticeable visual or performance effect include:
| Setting | What it changes |
|---|---|
denoise | About 0.35–0.55 aims to add detail while keeping the shot recognizable. Higher values rebuild more of the subject and can cause it to drift relative to the body or scene. |
steps | Match this to the turbo LoRA applied to the H3 model: for example, use 8 steps with an 8-step LoRA. The author suggests 8 if you do not know which to use. |
crop_factor | Around 1.5 magnifies the subject more. Try 2.0–2.5 when the blended edge needs more surrounding context. |
canvas_size | 768 is the documented default. Larger canvases increase processing cost, especially over many frames. |
confidence | Controls detection for the video. Inspect the track preview: green indicates detection and red indicates an interpolated position. |
subject_order | left_to_right can help distinguish multiple similar subjects standing near one another. |
mask_source | union uses the original and refined subject outlines so neither is unnecessarily clipped. |
Segmentation-based masks allow more precise pasting. A bounding-box-only detector can still track a subject, but the paste mask falls back to an ellipse. If a subject is hidden for part of the clip, those occluded frames retain their original pixels.
This workflow uses additional components: ComfyUI-Pixaroma for its loop nodes and VideoHelperSuite. If you separately use H3-FaceRefine, CoachBate also offers an off-by-default h3_facerefine compatibility patch for that pack.
Upscale and Frame Interpolator TensorRT: Finish a video in one node
CoachBate Upscale and Frame Interpolator TensorRT first enlarges the input frames using TensorRT, then uses ComfyUI’s FILM frame interpolation to create frames between them. Finally, it encodes an MP4 with the supplied audio. It replaces a separate Video Combine step in this workflow and returns the saved video filenames and the exact output frame rate.
You can choose an upscale model and FILM model, final size, interpolation multiplier, codec, pixel format, encoding quality, and output name. Important controls include:
| Setting | Effect |
|---|---|
rife_multiplier | Controls the number of output intervals per source interval. Despite the control’s name, interpolation here uses FILM; 1 bypasses FILM. |
source_fps | Sets the incoming frame rate used to calculate playback. The documented default is 24. |
playback_mode | preserve duration raises output FPS to accommodate added frames; slow motion keeps source FPS and extends playback time. |
codec | Choose H.264/AVC or H.265/HEVC MP4. |
pixel_format | Choose broadly compatible 8-bit yuv420p or 10-bit yuv420p10le. |
crf | Controls encoding quality: a lower number generally means higher quality and a larger file. The documented default is 16. |
save_output | Save in ComfyUI’s output location, or write a temporary preview when off. |
pingpong | Append frames in reverse order for back-and-forth playback. |
trim_to_audio | End the finished MP4 at the shorter of the video and audio streams. |
Custom width and height appear when resize_to is set to custom. The node needs ComfyUI-Upscaler-Tensorrt, ComfyUI’s native frame-interpolation nodes, and a FILM checkpoint in the configured frame_interpolation model folder. It does not require a separate RIFE node pack.
ComfyUI-CoachBate Models
ComfyUI-CoachBate does not provide one interchangeable set of built-in art models. Its model-dependent video features use models from other projects. Choose them according to the subject or processing task:
| Model or model type | When to use it |
|---|---|
| MiniMax H3 model, text encoder, and video/audio VAEs | The base components for H3 Subject Refine. See Comfy-Org/MiniMax-H3. |
| H3 turbo LoRA | Sets the appropriate generation-step workflow for your H3 model. See Kijai’s MiniMax H3 LoRAs. Match the node’s steps setting to the LoRA you use. |
| Subject-specific H3 LoRA | Helps H3 render the subject you are refining. The refinement pass uses the model you supply, including its applied LoRA. |
| YOLO segmentation model | Detects and tracks a subject and supplies a subject-shaped paste mask. Examples listed by the author include CockAndBallYolo8x.pt, face_yolov8m-seg_60.pt, and PitHandDetailer-v1-seg.pt. Choose a detector appropriate to your subject. See the YOLO model collection. |
| YOLO bounding-box model | Can track a subject if a suitable segmentation model is unavailable, but refinement uses an elliptical fallback paste mask. The author points to the same model collection’s bbox/ folder. |
| TensorRT upscale model | Enlarges frames in the combined upscale-and-interpolation node. Choose it based on the upscale result you want. |
| FILM checkpoint | Generates intermediate frames after upscaling. It is needed when interpolation is enabled, but is not loaded when the effective multiplier is 1 for every interval. |
The documented H3 detector locations are models/ultralytics/segm/ for segmentation models and models/ultralytics/bbox/ for bounding-box models. ComfyUI-Impact-Subpack registers those folders, and the CoachBate nodes also look in them directly when that subpack is absent.
What’s New with ComfyUI-CoachBate
The original README records the following dated changes, from newest to oldest:
| Date | Changes and why they matter |
|---|---|
| 2026-09-14 | Added the H3 Subject Refine nodes and one-subject/multiple-subject blueprints, making targeted video refinement possible. Added Video Prompt From Metadata and an optional H3-FaceRefine compatibility patch. |
| 2026-09-01 | Added CoachBate settings for editable model-path replacements, the Workflows+ panel toggle, and independently controlled compatibility patches. Model replacement changes can take effect without a restart; compatibility-patch changes require one. |
| 2026-08-02 | Improved Workflows+ folder/file ordering and row alignment. The panel now lists only JSON and PNG workflow files, opens listed PNG workflows correctly, and resets its temporary search and folder state when closed. |
| 2026-07-14 | Expanded Batch Prompter with one-at-a-time queueing, random order, corrected prompt numbering, and more reliable stopping at the last prompt. Updated Video Combine’s defaults and saved-filename cleanup. |
| 2026-07-13 | Improved Strip API Key Metadata with folder processing, path handling, and reliability fixes. Reworked Text Preview and Edit’s text area and added Copy Text. Updated confirmation dialogs and corrected Shot Loader’s next-shot display in decrement mode. |
| 2026-07-05 | Added folder rename, move, and soft-delete to Workflows+, along with a clearer folder-search indicator and a more prominent sidebar position. |
| 2026-07-04 | Added folder-scoped Name and Contains searches to Workflows+. |
| 2026-07-03 | Added workflow rename, move, duplicate, bulk selection, and soft-delete to Workflows+. |
| 2026-07-02 | Introduced Workflows+ with large-library browsing, search, sorting, pins, recent workflows, and drag-and-drop loading. |
| 2026-06-24 | Updated Batch Prompter to start from ComfyUI’s normal Queue button and queue independent jobs for non-blank lines. It also began ignoring muted, bypassed, or unconnected Batch Prompter nodes when deciding whether to run a batch. |
Troubleshooting ComfyUI-CoachBate
A workflow opens with an unavailable model
- Check that Workflow Model Path Auto-Fix is enabled under Settings → CoachBate → Model Path Replacements.
- Confirm that the model actually appears among the choices in the affected widget. Auto-Fix cannot select a file that is not available there.
- If you want to substitute a particular local file, add a Source → Target rule and save it.
- Open the workflow again and inspect the widget to confirm it selected the model you intended.
Workflows+ search does not find a workflow
- Check whether a folder is highlighted as the active search scope. Select Show All to search the full library.
- Check Search subfolders if the workflow may be nested inside another folder.
- For a filename search, remember that
black catlooks for a contiguous phrase. Tryblack AND catif the words may be separated. - If you remember text inside the workflow rather than its name, switch to Contains and press Enter. Allow time for a large scan, or cancel it if needed.
Batch Prompter runs only once or stops sooner than expected
- Make sure the Batch Prompter is active, not muted or bypassed, and that its output is connected.
- Check for a second eligible Batch Prompter in the workflow. With two or more, queuing falls back to one normal run.
- Check
starting_numberandmax_prompts; blank lines do not count as prompts. - If
randomizeis on withmax_prompts = 1, one randomly chosen prompt is the expected result.
Shot Loader cannot advance to a usable shot
- Confirm that
json_pathpoints to the intendedshotlist.json. - Check that each shot has
shot_id,video_filename_prefix,duration_seconds, andvideo_prompt. - Check
statusvalues. Shots marked"DONE"are skipped, and the node reports an error if all shots are marked DONE. - If an image warning appears, check the referenced start or end image path. A missing image produces a warning and an image strength of
0.0.
H3 refinement looks inconsistent or changes the subject too much
- Inspect the track preview. Green frames were detected; red frames use an interpolated track position.
- Adjust
confidenceif detection is unreliable, and choose a detector suited to the subject. - Try
denoisearound0.35–0.55to retain more of the original shot. - Match
stepsto your applied turbo LoRA. If the pasted boundary needs more surrounding context, try increasingcrop_factortoward2.0–2.5. - For several similar subjects, try
subject_order = left_to_right.
Frequently asked questions
Does Workflows+ replace or alter ComfyUI’s built-in Workflows tab?
No. It appears as its own sidebar tab alongside the built-in one.
Can Batch Prompter change something other than an image prompt?
Yes. It outputs an ordinary string, so a connected workflow can use each line as a filename, path component, label, or another text value.
Will a deleted workflow disappear immediately and permanently?
A first deletion moves it to _trash. Deleting it from _trash, or emptying the trash, permanently removes it.
Why does the upscale-and-interpolation node output a filename rather than frames?
That node encodes the finished MP4 itself. Its outputs are saved video filenames and the frame rate used for encoding, not an image batch.
Can H3 Subject Refine work without a segmentation detector?
A bounding-box detector can still track the subject, but the paste mask falls back to an ellipse instead of a subject-shaped segmentation mask.
Learn More about ComfyUI-CoachBate
- ComfyUI-CoachBate source and project information: The extension’s project page.
- ComfyUI-H3-FaceRefine: Background on the face-refinement approach adapted for H3 Subject Refine.
- MiniMax H3 model resources and H3 LoRAs: Resources for understanding the models used by the H3 workflow.
- YOLO segmentation models: Detector choices for different subjects.
- VideoHelperSuite and ComfyUI-Pixaroma: Related tools used by the documented video and looping workflows.
