ComfyUI-BFSNodes Introduction
ComfyUI-BFSNodes is a collection of ComfyUI nodes for making, guiding, editing, and comparing AI-generated images and videos. It includes tools for preparing reference images, planning a longer video as individual shots, marking people or objects in a guide, and keeping visual references organized as a workflow grows.
A useful way to think about the extension is as a set of helpers around your generation model. Some nodes prepare the material you give the model. Others control how that material is used during generation. Still others help you review the result. You can use the parts that fit your project; you do not need to use every node in one workflow.
The extension includes support for LTX-Video Edit Anything workflows and a range of MiniMax H3 image and video workflows. These are distinct paths. Choose nodes that match the model and workflow you are using rather than assuming that an LTXV node can be exchanged for an H3 node.
How ComfyUI-BFSNodes Works
ComfyUI connects nodes in a graph: one node produces an image, video, mask, model, or other output, and another node uses it. ComfyUI-BFSNodes adds specialized steps to that graph.
For example, suppose you want a generated video to retain the look of a character in a reference picture:
- Prepare the picture. A mask can identify the character, while a preparation node places them on a clean background or adjusts their apparent size.
- Give the model the reference. A conditioning node tells a compatible model how to use the picture. Depending on the workflow, it may also receive guide frames that describe the desired motion or composition.
- Generate and review. The sampler produces the result, and comparison nodes help you inspect where it differs from the guide.
For longer videos, the Shot Planner follows a similar idea at a larger scale. It divides a video into manageable shots, lets you choose a prompt and reference for each, and sends the shots through the rest of your workflow. Shot Join then brings the results back together with their intended timing and audio. Think of it as editing a sequence scene by scene rather than asking a model to handle the whole sequence at once.
A mask is simply a selection showing which part of an image or video matters. A guide supplies visual information such as pose, motion, shape, or layout. A reference is a picture you want the model to consider for appearance or identity. Which information takes priority depends on the particular node, model, and settings.
ComfyUI-BFSNodes Features
LTX-Video reference editing
LTXV Edit Anything (Apply) is intended for Edit Anything LoRAs trained for the LTX-Video 2.3-style video_to_video_ref_adaln workflow. It combines up to three ways of using a reference:
- Sequence conditioning places a reference image, and optionally guide video frames, into the model’s visual conditioning sequence.
- Role embedding helps the model distinguish the reference’s role when the LoRA contains the required weights.
- AdaLN reference conditioning carries an overall appearance or style signal from the reference image when the LoRA contains the required projector weights.
In everyday terms, these paths can help the model understand what picture to look at, what purpose that picture serves, and what overall look to carry into the result. The available paths depend on the LoRA checkpoint.
For the basic workflow, connect an LTXV model loader to a standard Load LoRA node, then to LTXV Edit Anything (Apply), and finally to your sampler. Give the Apply node the same LoRA checkpoint, your positive and negative conditioning, the LTXV VAE, the target video latent, and a reference image. Guide frames are optional. The node returns a model and conditioning for sampling, an updated latent, and a preview of the resized reference.
Suggested starting settings are guide_strength: 1.0, ref_strength: 1.0, role_strength: 1.0, adaln_scale: 1.0, both enable_adaln and enable_role_embedding on, and resize_mode: pad_to_fit. If the result seems to ignore the reference, try raising role_strength. If it copies the reference’s colors or textures too strongly, try lowering adaln_scale. Change one setting at a time so you can see what made a difference.
A standard LoRA loader alone does not provide the additional reference conditioning required by Edit Anything LoRAs that contain role-embedding or AdaLN weights. Use the Apply node after the standard loader for those checkpoints.
LTXV Apply Neutral Mask replaces areas outside your mask with white, neutral gray, or black. This can remove distracting surroundings from a reference or guide frame. LTXV Resize Reference By Mask uses the selected object’s mask to size it on a clean canvas, which is useful when the object appears too small or too large relative to the intended shot.
Nodes that insert clean appearance references also expose a reference-only temporal offset. Leave it at 0 for standard frame-zero placement. Experimental values of -1 or -2 move the appearance reference earlier in latent time and may reduce a brief reference-composition artifact at the start of a video; greater separation may weaken identity. Guide videos, masks, and target-video timing are not moved by this setting. In LTX Multiple Controls, the setting is named identity_temporal_offset_latents and affects only identity_image.
Shot planning, joining, and comparison
BFS Shot Planner presents a timeline with thumbnails, a cut-detection curve, draggable shot boundaries, and a reference gallery and prompt for each shot. You can divide footage at detected camera cuts, at fixed lengths, or manually. Its shot output is a ComfyUI list, allowing downstream nodes to process one shot at a time.
BFS Shot Unpack exposes an individual shot’s guide frames, references, prompt, and length for use with your chosen model. BFS Shot H3 Conditioning provides MiniMax H3-specific options, including an aligned guide, native reference video, or an aligned guide with CLIP video context. The MiniMax H3 Reference to Video — Guide 0 CLIP Context (BFS) option lets Qwen inspect a video guide at 2 fps without using native guide video-reference latents; motion guidance can be supplied separately through Add Guide.
BFS Shot Join trims generated shots to their intended lengths, combines them, can soften joins with cross-fades, and returns audio. When comparison is enabled, it can also output the original guide aligned frame by frame with the result. BFS Video Compare puts two videos in one player with horizontal or vertical wipes, overlays, side-by-side views, difference views, and a per-frame difference display you can click to navigate.
The shot workflow offers two run modes: auto loop processes all shots in one run; queue loop processes one shot per run, stores progress on disk, and re-queues automatically. Shots can also use SAM 3 masks for cropping and pasting, continue from a previous result, or receive shot and segmentation suggestions from a connected Qwen3-VL model. See the Shot Planner and Shot Loop guide for workflow details.
MiniMax H3 guides and side panels
MiniMax-H3 Downscaled Latent Guide (BFS) is for H3 LoRAs trained with target_grid_stride_v1. It provides a smaller, aligned image or video guide; connect both its MODEL and positive CONDITIONING outputs to sampling. MiniMax-H3 Guide Target — Image / Video (BFS) helps establish an exact output canvas. For example, the documented 1024 × 768 output with a factor of 4 gives a 256 × 192 guide preview. The H3 guide documentation covers setup, compatible workflows, and validation.
BFS H3 Duet creates a split-screen-style generation in which a source clip or reference is pinned beside the video during generation, then cropped away before decoding. BFS Shot H3 Duet applies this approach shot by shot. BFS H3 Side Panel and BFS H3 Side Panel Crop expose the underlying steps separately. Options include an aligned latent guide, panel_static over a selected person or face, and a BFS Setting Picture for same-world swaps. See the H3 Duet guide.
Experimental native H3 temporal motion transfer uses temporal alignment to help a generated video follow a source’s movement while retaining native image references. It requires a matching experimental LoRA and has backend and GPU limitations; consult the native temporal guide and its illustrated Flash Partitions explanation before relying on it in a production workflow.
Illustration of H3 temporal weight decomposition
Selecting and refining subjects
BFS Segment Picker uses SAM 3.1 to find objects in an image and lets you select them on the node’s canvas. Its automatic search starts with person, animal unless you change concepts. You can search for missed objects with Find or add keep/exclude points. It can output the selected mask, masks for individual objects, a tinted marked image, a numbered overlay, and object information. The default mark is magenta at mark_opacity: 0.3; mark_color and opacity can be changed to make the selection easier to inspect.
For a more precise selection, Brush adds area, Remove cuts out an object found at a point, and Eraser removes painted area. Holding Shift while painting reverses the brush or eraser action. These edits can be switched off individually. When you edit a selected object’s own mask, the edit belongs to that object and can be reset.
BFS Video Segment Picker (SAM 3) extends selection across a video. Choose a frame, choose a colored subject slot, and click the subject. The slots correspond to Picture 1–4: red, blue, green, and yellow. Propagate tracks the subjects through the video, working shot by shot rather than tracking across detected cuts. A whole-video scrubber shows masks and highlights stretches where tracking was lost, so you can add a corrective frame. Optional Molmo2 assistance can describe and re-anchor a subject after cuts.
The picker offers clean_masks for small errors where subjects touch or cross. 0 turns cleaning off; the documented usual range is 0.5–2. Its hybrid output can depict selected subjects as a normal map, an OpenPose-style skeleton with face details, flat color, shading, blur, pixelation, gray detail, or edges. Some modes require corresponding MoGe or DWPose inputs. hybrid_grow changes the margin around a subject, while free_after controls whether models are unloaded from GPU memory when the node finishes.
Character swaps and reference control
BFS H3 Character Swap Conditioning uses marked subjects and Pictures 1–4 as references for MiniMax H3 video generation. Its keyframe pins mode combines a source video with guides at selected frames. Its normal map guide and DWPose guide modes use the Video Segment Picker’s hybrid guide. The node provides the model, conditioning, latent, prompt, keyframe video, and reference pictures needed by the workflow. BFS Keyframe Planner helps choose and export keyframes; accompanying mask and tracking nodes help follow a subject at those frames.
For H3 image generation, BFS H3 Image Character Swap Conditioning treats the marked guide image as Picture 1 and connected character references as Pictures 2–4. BFS H3 Image Character Swap • Auto Subjects reads the Picker’s selected colors and determines which reference slots are active, so you do not need to enter a subject count manually. You can keep its generated prompt or replace it with a complete prompt_override. The Picker’s colored hatching and contour mode offers another way to mark subjects while leaving source pixels visible between the marks.
BFS Optional Reference Image can load an image or return no image for [none]. Its resizing preserves aspect ratio; megapixels and scale_method are adjustable, with 4 MP and Lanczos as defaults. The H3 encoder still applies its own reference-size limits.
Other experimental controls include Reference Attention Binding, which adjusts how strongly individual Pictures influence specified areas; SAM 3.1 click-to-Picture bindings; and a Token / Word Weight Editor that can emphasize words or (phrase:1.3)-style phrases. Their behavior and limits are documented in the reference binding guide and cast and token controls guide.
Workflow and model utilities
The repository also includes Reserved Region Frame Composer, Frame Ranged Face Loader, and Face Sequence Batch for video composition. BFS Workflow Loop, BFS XY Sweep, and BFS Frame Loop Start/End help repeat part of a workflow or compare settings systematically.
BFS Model Surgery is an advanced, reversible weight-scaling control for a cloned MODEL. It can apply rules to selected layers; an empty rule set leaves the model unchanged. Because it scales model weights rather than only a LoRA’s contribution, use it for deliberate experiments, not as a routine reference-strength slider. BFS H3 PDD Surgery is a separate experimental control for native H3 PDD output heads. Settings of 1 preserve the original model; setting offsets to zero does not recreate a pre-distillation checkpoint.
ComfyUI-BFSNodes Models
ComfyUI-BFSNodes supplies workflow nodes; the referenced generation models, LoRAs, and segmentation checkpoints are separate components. Choose the combination appropriate to your task:
| Model or component | When you would use it |
|---|---|
| LTXV/LTX-Video model and compatible VAE | For the LTXV reference-image and Edit Anything video workflows. |
Edit Anything LoRA trained for video_to_video_ref_adaln | When using LTXV Edit Anything (Apply) to bring in the LoRA’s reference-conditioning paths. The readme names edit_anything_reference_v0.1_r128_12000.safetensors as the first expected public reference checkpoint, not as a checkpoint bundled with the nodes. |
| MiniMax H3 video model and matching workflow components | For H3 shot conditioning, guides, duet, temporal experiments, and video character swaps. Some specialized guides require a matching LoRA. |
| MiniMax H3 Image model, image VAE, and matching text encoder | For the H3 image character-swap workflow, which requires ComfyUI 0.45.0 or newer according to the provided documentation. |
| SAM 3.1 checkpoint | For selecting objects with BFS Segment Picker; its documented checkpoint is sam3.1_multiplex_fp16.safetensors. SAM 3-based video selection is used in the Video Segment Picker. |
| Optional Molmo2 or MolmoPoint-8B components | For supported subject-description, re-anchoring, or pointing assistance in Video Segment Picker. SAM 3 still produces the masks. |
These components do different jobs. A generation model creates the image or video; a LoRA adapts a compatible model for a particular conditioning strategy; SAM helps select subjects. Adding one does not replace the others.
What’s New with ComfyUI-BFSNodes
The following milestones are described in the supplied readme. Features marked experimental should be evaluated with their linked documentation and example workflows.
| Version | Change and why it matters |
|---|---|
| 1.46.0–1.47.0 | Added downscaled MiniMax H3 latent guides, followed by experimental source-phase reference layouts and identity reference controls. These give compatible H3 LoRAs more ways to align a lower-resolution guide or reference with the target. |
| 1.48.0 | Introduced Shot Planner and Shot Loop, making it possible to organize long videos into prompted, referenced shots and rejoin the results. |
| 1.50.0 | Added H3 Duet and side-panel workflows for synchronized, training-free split-screen generation. |
| 1.84.0 | Added panel_static and a setting picture option for H3 side-panel swaps. |
| 1.115.14–1.115.22 | Expanded experimental native temporal backend support, including sol-attn compatibility, an optional one-pass Kitchen extension, and automatic selection of verified prebuilts for specified Linux and Windows GPU configurations. Platform validation remains experimental. |
| 1.115.25 | Added a reference_images preview output to BFS H3 Describe References, making it easier to check Picture 1–N order without changing what the encoder receives. |
| 1.115.27–1.115.31 | Added an experimental native shot-loop example, Reference Attention Binding, SAM 3.1 Picture bindings, phrase weights, and an example workflow with independently switchable SAM and shot-loop branches. |
| 1.116.0 | Added BFS Segment Picker for choosing SAM 3.1 objects directly on an image. |
| 1.118.0 | Added hand-editing tools for image selections and fixed a canvas-sizing issue when the side panel is open. |
| 1.119.0 | Added the Video Segment Picker, H3 character-swap conditioning, and supporting keyframe and workflow-loop tools. |
| 1.121.1–1.121.2 | Improved picked-frame brush repairs, then made brush edits apply directly to a selected object’s mask, with a reset control. |
The readme also describes H3 image character-swap tools and model-surgery utilities, but does not assign them a version in the supplied context.
Troubleshooting ComfyUI-BFSNodes
My LTXV result ignores the reference image
- Confirm that the workflow uses an LTXV model and compatible VAE.
- Confirm that the Edit Anything LoRA is loaded through the standard LoRA loader before LTXV Edit Anything (Apply).
- Select that same checkpoint in the Apply node’s
lora_nameinput and check thatref_imageis connected. - Start with the recommended settings above. If the reference still has too little influence, test a higher
role_strength.
The extra role and AdaLN paths work only when the checkpoint contains their respective weights. A plain LoRA load does not replace the Apply node for LoRAs that need those paths.
The result copies too much color or texture from my reference
- Reduce
adaln_scalein LTXV Edit Anything (Apply). - Preview the prepared reference. If its background is distracting, use LTXV Apply Neutral Mask to replace areas outside the subject with a clean color.
- Change one control at a time and compare the outputs.
The reference composition appears briefly at the start of a video
Try the appearance reference’s temporal offset at -1, then compare it with the default 0. -2 offers stronger separation but may weaken identity. This is an experimental adjustment to the appearance reference, not a retiming control for your guide video or target frames.
My H3 downscaled guide is not affecting the sample
Check that your LoRA is compatible with the documented target_grid_stride_v1 guide approach. Then connect both the Downscaled Latent Guide node’s MODEL and positive CONDITIONING outputs to the sampler. See the H3 guide workflow and compatibility notes.
A subject disappears or the mask becomes inaccurate after a camera cut
Open BFS Video Segment Picker and inspect the Whole video scrubber for lost stretches. Add a frame from the affected shot to the strip, pick the subject there, and propagate again. For small local errors, use the brush repair on a picked frame. If subjects overlap, test a modest clean_masks value and inspect the result; 0 disables that cleaning.
Frequently asked questions
Do I need every model mentioned here?
No. Use the model and supporting components for the workflow you choose. For example, an LTXV Edit Anything workflow does not require the H3 Image model.
Can I use any LoRA with LTXV Edit Anything (Apply)?
The node is intended for Edit Anything LoRAs trained for its conditioning strategy. Its role-embedding and AdaLN paths depend on the relevant weights being present in the LoRA.
Do I have to plan shots to use the H3 nodes?
No. Shot planning is for workflows that benefit from handling a longer video in sections. Other H3 guides and conditioning nodes can be used in their corresponding workflows without the Planner.
Will an experimental guide or alignment node guarantee the same result as a guide-trained video LoRA?
No. For example, Add Guide for MiniMax H3 Image (BFS) has passed structural checks, but the supplied documentation does not guarantee visual equivalence to a video-trained guide LoRA.
Learn More about ComfyUI-BFSNodes
- Shot Planner and Shot Loop: timeline, run modes, and joining a sequence.
- H3 Duet and Side Panel: split-screen generation and its component nodes.
- MiniMax H3 guides: guide modes, compatible workflows, and validation.
- Native H3 temporal motion transfer and the illustrated Flash Partitions explanation: experimental motion alignment, performance, and limitations.
- Native H3 automatic prompting: reference descriptions, prompt stages, and source-derived video settings. See also the visual example workflow.
- Native H3 shot-loop example and shot-loop guide: an experimental loop without the Shot Planner.
- Reference Attention Binding and its example workflow: controls for individual Picture references.
- Cast and token controls: SAM 3.1 Picture bindings, word and phrase weights, and workflow limits.
- Animated TARA parameter guide: a visual explanation of temporal-control settings.
- ComfyUI-Molmo2-H3: the referenced source for the optional MolmoPoint-8B loader used with Video Segment Picker.
For the project’s licensing terms, see LICENSE.
