ComfyUI-Easy-Media

中文文档 | Changelog

A ComfyUI custom node package for streamlined media loading and video pipeline assembly. Provides intuitive nodes that simplify media resource editing and loading with user-friendly parameters, making it easier to build and configure video processing workflows.

cd Your_ComfyUI_Path/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Media.git

After installing, open ComfyUI and find the bundled example workflows in the Templates panel on the left sidebar — look for entries under ComfyUI-Easy-Media.

The package includes the easy-media-multitrack-workflow Skill. Give Codex your assets and requirements in natural language to create a workflow from the bundled template or modify an existing one, automatically configuring the timeline, prompts, continuity modes, and sampling settings.

Use $easy-media-multitrack-workflow to create three 5-second tasks from these reference images and prompts.
Use Shot for the first segment and Context for the next two, then generate a new workflow JSON.

By default, the Skill only generates JSON. When explicitly asked to run the workflow, it can upload assets and submit it to ComfyUI. See the Skill documentation for installation and detailed usage.

Introduced in v1.3.0, the MultiTrack Project Pipeline connects timeline arrangement, segment generation, context continuity, and final video assembly. MultiTrack Project currently supports MiniMax H3; MultiTrack Editor remains independent of any model and can still feed other model workflows.

Use the workflow generation Skill to configure this pipeline through natural language, either from a blank template or by modifying an existing workflow.

MultiTrack Editor ── TRACKS_INFO ──→ MultiTrack Project ── PROJECT_NAME ──→ MultiTrack Project Video Combine ── VIDEO ──→ SaveVideo
                                          ↑
                                    model_loader
                              (optional model_loader_2nd)

Select the MiniMax format, set the target dimensions and frame rate, enter prompts for each task segment, and add reference images, video, or audio as needed. See the MultiTrack Editor overview below for tracks, task modes, and media editing.

The new continuity mode determines how a task follows the previous segment. It is configured separately from task modes such as text-to-video, image-to-video, and reference-to-video:

The first segment starts in Shot mode. Set subsequent segments individually or select multiple tasks to change them together. For example, “Shot → Context → Context → Shot” creates three connected segments followed by a new shot. This setting affects generation, rather than adding a crossfade during assembly. Context does not guarantee seamless continuity across arbitrary scene or prompt changes.

Connect the editor's TRACKS_INFO to tracks_info, then connect a model_loader containing the H3 model, CLIP, video VAE, and audio VAE. The node automatically expands the tasks in sequence, so no manual segment-index loop is needed:

Read task and load media on demand → Encode prompts and reference conditioning → First pass
    → [Optional: upscale video latent → Second pass] → Decode audio/video → Trim context overlap
    → Save segment and context → Process next task → Output project name

Each segment uses its own prompts, references, and continuity mode. With Context enabled, the first pass uses the previous segment's context at the corresponding resolution; the second pass uses its final high-resolution result to preserve the join. Each segment is saved before the next one runs. You can choose a starting segment and count, or save a first-pass preview and resume the second pass later. See MultiTrack Project details for settings, upscaler dependencies, and the context implementation's source and adaptations.

Connect PROJECT_NAME to the combine node to preview saved segments, choose a version for each, and assemble the final video. Auto combine produces a complete video in one workflow run. Turn it off to review results first, then click Combine to run only assembly and downstream saving without re-running the model. If a segment has multiple generated versions, select two for synchronized comparison before choosing the final version. See MultiTrack Project Video Combine for details.

Tips: The advantage of the multi-track editor is its decoupling design — it is used solely for media editing and loading, and is not bound to any model. Users can freely choose any model node to process the media data output by the multi-track editor.

easy multiImagesLoader provides the same resolution choices for an ordered list of up to 25 images. Add images through the media selector or drop image files onto the grid; the node resizes each image and outputs an IMAGE list in grid order.

• Task segments are the core of this node; workflows can be designed for automatic looping based on the number of task track segments
• When adding video clips to the video track, corresponding task segments will be automatically added with matching duration
• Selecting a task segment allows you to set image, task type, and user prompt / system prompt (defaults exist based on task type, or you can write your own)
• The MultiTrack Info Output node outputs video dimensions, total frame count, frame rate, and task count
• The MultiTrack Task Output node outputs user prompt & system prompt for corresponding segments; users can decide whether to connect LLM nodes for prompt expansion or use images in segments for reverse inference

The editor passes time ranges, prompts, media locations, and track settings through TRACKS_INFO. Downstream output nodes load the required media when processing each task, reducing the cost of decoding and assembling an entire long timeline in advance. Metadata probing, frontend thumbnails, and waveform previews are separate from loading complete media for workflow execution.

If an older workflow connects directly to the editor's media outputs, insert MultiTrack Task Output (easy multiTrackTaskOutput):

• Process one segment at a time: Connect TRACKS_INFO and set task_index = 0, 1, 2… to retrieve media for the corresponding task range.
• Retrieve all media at once: Set task_index = -1 to output the entire timeline's media, replacing the editor's former full-media output behavior. This restores the resource cost of loading the full timeline.
• Read only dimensions, frame rate, total frames, and task count: Use MultiTrack Info Output; complete media loading is unnecessary for this metadata.
• Use the project pipeline: MultiTrack Project calls Task Output internally, so no extra Task Output node is needed between the editor and the project.

• Task track: Adds Shot / Context / Character Swap Context continuity modes. Select multiple task segments to change their task mode, continuity mode, and reference image size together.
• Lock audio: In MultiTrack Project, the input audio constrains visual generation, and the delivered video uses the original task audio instead of regenerating it. This differs from using audio only as a reference.
• Reuse audio: Share the same reference audio across tasks without copying it under every segment. Up to 15 seconds are used, without truncating it to a shorter task duration.

• Only the most common open-source model generation types are listed; theoretically any video model pipeline can use the multi-track editor as a preprocessing tool

Note: Some models support automatic download via the built-in Easy-Media model download interface. Model files will be placed in the ComfyUI/models/ directory.

easy multitrackProject manages segment-by-segment generation for MiniMax H3 projects. The editor's dimensions are the first-pass dimensions; dual sampling scales up from this size for the second pass. Project files are stored in ComfyUI/output/easy_media/projects/<project_name>/, including segment media, project records, and context latents for continuation.

1. Read and encode the task: Load its prompts and media, build conditioning for text-to-video, first/last frames, last-frame-only, or multimedia references according to the task mode, and create audio/video latents. When audio is locked, encode it and apply sampling constraints.
2. First pass: Both sampling_mode = single and dual generate at the editor's configured dimensions. The first-pass size is not reduced based on upscale_by.
3. Upscale and second pass (dual only): Multiply each configured dimension by upscale_by, then align to the nearest multiple of 32 using round(dimension * upscale_by / 32) * 32. Upscale the first-pass video latent to this size, recombine it with the audio latent, and sample again. Rebuild reference conditioning when dimensions change. Setting upscale_by = 1.000 skips resizing but still allows a second pass.
4. Decode and save: Decode video and audio. For Context segments, trim the repeated guide prefix and extra trailing frames required by the temporal grid before saving the segment.
5. Continue to the next segment: Save context and project records, then process subsequent tasks automatically. Shot starts a new shot; Context inherits the previous segment. Finally, output PROJECT_NAME for the combine node.

Alignment uses Python round, matching the H3 latent upscaler; it is not always upward, and exact ties round to the even integer. For example, 1344 × 768 at 1.250 becomes 1680 × 960 before alignment and 1664 × 960 for the second pass. Both upscaling paths use this size, and project records and combined exports retain the resulting dimensions. Single-pass generation and first-pass-only previews keep the editor dimensions; audio-only projects (32 × 32) do not upscale. Existing workflows retain their saved multiplier rather than automatically adopting the new default.

To customize presets, copy h3_sample.json.example to presets/h3_sample.json and edit it. The current built-in dual-sampling presets use separate first- and second-pass sigma schedules. Do not assume that light always splits sigmas and leaves the first pass incomplete. A custom preset with split_step explicitly splits one sigma schedule across both passes. Existing custom files take precedence, so review your configuration after upgrading.

upscale_model selects H3 latent upscaler weights, which serve a different purpose from the H3 generation model in model_loader_2nd. Direct latent upscaling avoids the intermediate video VAE decode/encode round trip, but the second pass still runs at the target resolution; its VRAM requirements do not decrease proportionally. Easy Media includes the inference node and runtime; only the checkpoint is downloaded separately.

The standalone easy minimaxH3LatentUpscaler node exposes enable_temporal_chunking and force_unload. Both default to enabled; MultiTrack's internal latent-upscale path also enables both explicitly. Temporal chunking lowers peak memory for long latents, while forced unload releases the upscaler from VRAM after inference.

For example, download minimax_h3_latent_upscaler_3d_fp16.safetensors from the model repository above, place it in the specified directory, restart ComfyUI, and select it under upscale_model. Its documentation specifies a 1–4× scaling range. Follow the selected model's supported range even if the project parameter accepts larger values.

Context conditioning is based on NikoDemon80 / ComfyUI-H3-Motion-Context, adapted inside Easy Media for project loops and dual sampling. The current hard-continuity implementation requires native H3 audio/video keyframe support in ComfyUI (0.34.0+). Check ComfyUI compatibility when upgrading.

In addition to attaching the previous segment's context conditioning, the project pipeline applies these adaptations:

• First-pass hard continuity: Preserves native audio/video keyframes and existing multimedia references, while copying the previous segment's audio/video tail into the current starting latent. Separate video and audio masks lock and gradually release the copied region, allowing new content to emerge around the join.
• Separate low- and high-resolution context: In dual sampling, the first pass inherits the previous segment's first-pass context. After upscaling, the second pass copies the previous final high-resolution video tail into the current high-resolution latent, rather than merely enlarging the low-resolution join. Context second-pass sampling freezes the current audio to retain continuity established in the first pass.
• Context-specific second-pass sigmas: Built-in presets provide sigmas_2nd_context = 0.50, 0.30, 0.14, 0.06, 0.0 for second passes with previous-segment context. Explicit custom second-pass sampler or sigma inputs prevent this substitution.
• Synchronized trimming and clean continuation sources: By default, the project takes the previous segment's last 22 frames as context and reserves 34 extra generation frames to satisfy H3's temporal grid. After decoding, it removes the repeated prefix and excess tail, retaining the task's required frame count. It then re-encodes context from the actual delivered audio/video range so subsequent segments do not inherit discarded tail frames.
• Memory-bounded high-resolution context: Before the next segment starts, the high-resolution continuation latent is reduced to the 22-frame video/audio tail, detached from the full sampling result, and moved to CPU. Project artifacts store the same context-only high-resolution payload after second-pass completion. The complete low-resolution first-pass latent remains saved so a 1st pass only run can resume its second pass later; runtime context handoffs use a separate trimmed copy.
• Automatic handoff and resuming across runs: Context passes automatically between segments in one run. When starting from a later segment, the project loads context from the previous segment's active saved version, without manually wiring Save / Load Latent nodes.

Note: Context continuation requires the previous segment's saved latents; an MP4 alone is insufficient. After changing editor dimensions, scaling factor, or the previous segment's version, check the downstream context chain. Regenerate first-pass checkpoints and context latents created with the old reduced-first-pass sizing before resuming with the new sizing behavior. Existing later segments are not automatically regenerated when an earlier segment changes.

For example, to regenerate only segment 3, set segment_start_number = 3 and segment_count = 1. Choose project_save = new to retain the old result for comparison. If segment 3 uses Context, the project must contain compatible context from segment 2. After accepting the new segment 3, regenerate any later Context segments that depend on it.

easy multitrackProjectVideoCombine reads saved video segments from a project. Preview them sequentially on the timeline without first creating a merged file. Use the project selector to switch projects or refresh saved results.

• Auto combine (enabled by default): After project generation finishes, reads the current segments, concatenates them in timeline order, and outputs VIDEO and FILENAME_PREFIX. Connect an output node such as SaveVideo to save the final video.
• Manual combine: With Auto combine disabled, the project still generates and saves individual segments, while the combine node updates the preview without sending a merged result downstream. Review and select your clips, wait for the ComfyUI queue to empty, then click Combine. Only this node and its downstream nodes are queued; upstream encoding and sampling are not repeated.
• Saving requirements: Connect a downstream video save output before combining manually. The combine node provides a temporary merged video; the save node controls the final path and filename. Use FILENAME_PREFIX as a naming prefix if desired.

After regenerating with project_save = new, click a timeline segment to open its video file list:

1. Select one video: Use that version for project preview and assembly.
2. Select two videos: Enter synchronized comparison to review results from different seeds, prompts, or sampling settings. Each segment allows at most two selected versions at a time.
3. Keep one after comparing: Deselect the unwanted version, then combine manually. Selecting two is only for comparison: assembly still uses the primary version (the first in the current selection list), not the comparison layout or both versions together.

Segments display their Shot / Context mode to help review joins. Selecting an older version changes the assembled clip but does not repair already-generated downstream context. If versions end with different motion or audio, regenerate the affected later segments. Deleting a segment version or an entire project also deletes associated files and context; make sure they are no longer needed for resuming.

Scope: This node is for video projects. Audio-only projects with editor dimensions set to 32 × 32 cannot use it to combine video.

Integrated and enhanced the video saving node from the SaveVideoRGBA node package. Supports video export with customizable output path, filename prefix, frame rate, and other parameters.

Load video files from a list of file paths (or URLs) and concatenate them into a single video output.

The trim_frame_count parameter defaults to -1, which keeps all frames of the merged video. When set to a value greater than 0, the node calculates the duration based on the merged video's frame rate and uses FFmpeg to trim the final video.

1. Create a config.yaml file in the ComfyUI-Easy-Media directory and add the following content to enable frontend development mode:

WEB_VERSION: dev

2. Navigate to the frontend directory and compile the development code for debugging:

cd frontend && bun install && bun run dev

3. After modifying the code, compile for production:

bun run build:release

• OmniShotCut
• Qwen3-ASR
• Whisper
• VoxCPM2
• Bernini S2V
• H3 Motion Context
• Drift Control for h3 motion context
• MiniMax H3 Latent Upscaler
• SelfLift for H3 Sampling
• Nvidia-RTX-VSR for video

• WhatDreamsCost-ComfyUI
• ComfyUI-PromptRelay
• ComfyUI-Licon-MSR
• ComfyUI-RH-Bernini