ComfyUI-H3-Multishot Introduction
ComfyUI-H3-Multishot turns short MiniMax-H3 generations into a longer scene designed to feel like one continuous camera take. MiniMax-H3 normally generates clips of roughly 10–15 seconds. This extension creates successive shots, carries visual and audio context between them, and assembles the results into a longer video.
It is useful when you want a character to keep speaking across several generated clips, a camera movement to continue through a join, or the lighting and appearance of a scene to remain consistent. It also gives you ways to preview the first shot, reuse character reference images, manage longer renders, and correct some of the gradual changes that can appear in a chain.
A seamless result still depends on the script. The extension can carry context across a boundary, but the next shot’s description must make sense as a continuation of what the previous shot showed.
How ComfyUI-H3-Multishot Works
Think of each H3 generation as a short section of a longer performance. For the next section, the extension supplies information from the end of the previous one so H3 has a starting point for the picture and sound. It then joins the sections into a master video.
There are two main approaches to continuity:
- Pinned context carries part of the previous clip’s picture and audio into the next generation. This helps motion, speech, and sound continue at the join rather than restart as a similar-looking or similar-sounding new take. The full workflow’s
context_pinmode uses ComfyUI-H3-Motion-Context. - Frame or plate chaining uses a visual boundary from one shot to guide the next. For example,
flf_chainrelies on its boundary plates for continuity rather than also feeding a memory-bank clip from shot 1 into later shots.
The assembled video is not simply a sequence of unrelated prompts. Each shot affects the next, which is why both the continuity mode and your wording matter. If shot 1 ends on a close-up of one person, opening shot 2 by immediately describing a completely different arrangement can produce a confused image. A clearer script begins from the existing framing, allows a small natural action—such as a breath or a shift in gaze—and then moves into the new action.
For a scene with several shots, write one prompt per shot and separate prompts with ---. Keep descriptions of a character’s appearance and the room consistent, but give each shot a distinct physical action. Repeating the same action in different words can make later shots look like near-copies.
ComfyUI-H3-Multishot Features
Workflows for different creative tasks
- CORE workflow: A starting point for multishot rendering that uses this pack and ComfyUI’s built-in nodes, without additional third-party node packs.
- Full
H3_Seamless_Chain_v2workflow: Includes optional lanes for features such as an LLM script writer,context_pin, speed and memory controls, and an on-canvas script preview. Some of those lanes require their corresponding third-party packs. H3_Extend_Takeworkflow: Sets up an extend take: you provide one premise and a target duration, and the writer divides a continuous speech into windows at sentence boundaries.- Keyframes workflow: A separate task for anchoring one clip at chosen frame positions with individual anchor strengths. It is not a multishot workflow.
Older H3_Multishot_AIO and H3_Multishot_MEMORY workflows are retired, although existing copies can still work. Their multishot lanes were incorporated into v2.
Shot length, previews, and longer takes
The MASTER CONTROLS panel sets resolution, frames per shot, and sampling steps. For a first render, the author recommends leaving the workflow’s supplied values unchanged and testing with optional switches off and the manual memory reserve at 0. The automatic reserve planner measures what the render needs; a hand-set reserve overrides that measurement.
preview_first_shot is on by default. It lets you inspect shot 1’s framing and voice before waiting for the whole scene. Completed output is written under output/video/H3CHAIN/ as a 24 fps video with a paired audio file.
For an extend take, set take_seconds above 0. With take_seconds = 0, the normal shot-by-shot behavior remains. The window setting can follow frames_per_shot, use an explicit window length, or choose fit this card (VRAM auto) when your priority is avoiding memory pressure rather than specifying an exact window.
Scripts, dialogue, and character references
The optional LLM writer can turn a premise into a shot script. Match its num_frames value to the sampler’s frames_per_shot: if it assumes a shorter shot, dialogue may be rushed or garbled; if it assumes a longer one, you may hear unwanted pauses. The writer reports dialogue that appears too long for the available speaking time.
You do not need an LLM writer to use the extension. You can provide your own ----separated script to the sampler; the CORE workflow already follows this manual approach.
Reference photographs can help keep characters recognizable. Auto Refs can select available photos, while manual reference inputs let you choose them yourself. When several people have references, reference_subjects groups pictures by person in picture order: 3,3, for example, means three pictures for the first person and three for the second. Leaving the field empty treats all pictures as one person. The samplers also offer voice_ref_2 and voice_ref_3 so additional subjects can retain distinct voices across a scene.
An existing clip can serve as an H3 Reference Video for a new generation. This guides the scene’s appearance—such as framing, room contents, and color temperature—but does not make the subject copy the clip’s movements. Short reference windows cost less processing time than long ones.
Continuity, finishing, and memory controls
The samplers offer different continuity choices. Use context_pin when you want the pinned picture and audio continuation supplied by Motion-Context. first_frame is an alternative when Motion-Context is unavailable. In flf_chain, boundary plates carry continuity; the memory bank is automatically disabled in that mode so footage from the opening plate does not reappear in later shots.
Long chains can gradually become brighter, sharper, or more detailed because each shot draws on generated material from earlier shots. Available controls include:
H3ChainNormalize, a post-chain color and texture leveller.master_normalize = luma+contrast, which levels brightness and contrast in the finished chain.pin_renorm, which helps hold the pinned material at a consistent level.chain_gain_control = flatten, intended to reduce accumulating texture in longer chains.
These controls reduce drift; they do not guarantee that a very long scene will remain identical from beginning to end.
For demanding renders, low_ram_master stages completed shots on disk instead of retaining the entire chain in system RAM until assembly. A remote text encoder can move text encoding to another computer, leaving more GPU memory for video generation. H3 TAE Decode provides fast draft previews for judging composition and motion, but its softened detail makes it unsuitable for final output. Optional speed boosters can reduce render time; assess their effect on people and audio before using them for a final take.
H3Retake offers a different finishing tool: it regenerates a selected stretch of a completed clip while keeping the material outside that window.
ComfyUI-H3-Multishot Models
The main model choice described in the documentation is between two MiniMax-H3 checkpoint types:
| Checkpoint | Choose it when you want… | What to expect |
|---|---|---|
| ref2va | Character, voice, image, or video references across a scene | Supports reference conditioning. It can be guided toward a supplied frame, but is less exact at landing on that frame. |
| fl2va | A shot to start precisely from a supplied frame, or to use both a first- and last-frame target | Stronger frame placement. It does not use the reference rows needed for the reference-video feature. |
fl2va is not a smaller checkpoint than ref2va at the same quantization level. Its work per sampling step can be lower because it does not carry reference rows, but the files are the same size at corresponding quantization levels.
The extension’s loader supports safetensors and GGUF formats, including the GGUF architecture support H3 needs. The workflows also use an H3 text encoder and video and audio VAEs. If you use the optional local Ollama script writer, its configured language model is separate from the H3 video model: the full workflow points to qwen3:14b, but its model_name must match a model available to your chosen OpenAI-compatible endpoint.
What’s New with ComfyUI-H3-Multishot
The original documentation describes the following releases and changes:
| Version | Changes that matter for your workflow |
|---|---|
| 2.7.2 | Publishing-format cleanup only. No rendering behavior changed; it was intended to make the release available through ComfyUI-Manager. |
| 2.7.1 | Removed an incompatible code-execution mechanism from the pack’s interior-anchor handling. Also documented a context_pin alignment problem on ComfyUI 0.34 whose fix is in ComfyUI-H3-Motion-Context, not this pack. |
| 2.7.0 | Added references for second and third subjects’ voices, fixed opening-plate imagery reappearing in flf_chain, introduced H3ChainNormalize, improved refresh_pin splice alignment, and added memory-sampler controls. A second-sampler schedule option was introduced as experimental. |
| 2.6.3 | Made frames_per_shot control extend-take window length when window = auto; added a separate VRAM-fit option. Remote encoder addresses are checked earlier, and stale models are cleared before the first remotely encoded shot. |
| 2.6.2 | Aligned reference prompts more closely with MiniMax’s documented format, improved handling of multiple referenced characters, and included H3Retake. |
| 2.6.1 | Fixed a reference-video error caused by a change in ComfyUI. Auto Refs also gained a path that warns and continues without photos when no matching character folder is found. |
| 2.6.0 | Introduced extend takes, simplified Auto Refs routing, and improved how the writer describes characters when reference photos are attached. |
| 2.5.5 | Added automatic GPU-memory headroom handling, disk-staged master assembly, remote text encoding, fast draft decoding, and optional speed boosters. |
| 2.2.5–2.2.2 | Improved memory-reserve learning, warned before costly upscales, fixed dialogue-length guidance and several long-render failures, added grouped multi-person references and folder-based script selection, and corrected misleading reference descriptions. |
| 2.1 | Added the prompt-source and script-picker nodes, plus the optional unload_model_after control on the LLM writer. |
Compatibility note: The source documentation specifically warns that Motion-Context 0.4.0 does not work with this pack’s context_pin path and says to use 0.3.1 for that configuration. Because the context_pin fix for ComfyUI 0.34 belongs to Motion-Context, check that pack’s current documentation before starting a long render. The Motion-Context repository is the appropriate source for its current compatibility guidance.
Troubleshooting ComfyUI-H3-Multishot
The full workflow will not queue because nodes are missing
ComfyUI checks every node class in a workflow before a render begins, including nodes in optional lanes.
- Try the CORE workflow if you want a multishot render without third-party packs.
- If you want the full workflow, identify the missing node in ComfyUI’s message.
- Check the full workflow’s optional requirements: the bundled JoyEcho writer, ComfyUI-H3-Motion-Context, RES4LYF, the optional optimization packs, and ComfyUI-Custom-Scripts for the canvas preview.
- Alternatively, follow the bundled
INSTALL.mdguidance to remove or replace the optional nodes. For example, without RES4LYF, change the scheduler frombeta57tobeta; without Motion-Context, usecontinuity=first_frame.
The script writer reports model 'qwen3:14b' not found
The full workflow is pointing to an Ollama model that the endpoint does not have.
- Check the writer’s
base_urlandmodel_name. - If you use local Ollama, compare
model_namewith the exact tags shown byollama list. The original workflow’s model can be obtained withollama pull qwen3:14b. - If you use another OpenAI-compatible endpoint, enter its URL and the exact model tag it provides.
- If you prefer not to use an LLM, use manual prompts and feed a
----separated script to the sampler, as in the CORE workflow.
On a single GPU, a local writer can compete with H3 for memory. Turning on the writer’s unload_model_after option frees its Ollama model after the script is written.
Speech is rushed, garbled, or followed by a long pause
- Confirm that the writer’s
num_framesmatches the sampler’sframes_per_shot. - Read any dialogue-budget warnings produced when the script is generated.
- Shorten lines that overrun the shot, or add enough natural speech to a long speaking window.
- For a continuous monologue, consider the extend-take writer style, which divides speech across windows at sentence boundaries.
Later shots repeat imagery or visibly change character
- Make sure each shot introduces a new physical action, rather than restating the previous shot’s action.
- Keep appearance and room descriptions consistent, especially when a face must carry into the next shot.
- If you use
flf_chain, use a version containing the 2.7.0 fix that disables its competing memory bank. - For gradual sharpening or color drift over a long chain, try the chain-levelling controls described above and review the finished master, not only individual shots.
A remote text encoder fails or the first shot is unexpectedly slow
Check the remote encoder’s address first. An empty address, the unchanged OTHER-PC placeholder, or an unreachable host should now produce an early, specific error. If the address is valid but the first shot is slow, check the console for messages about clearing leftover models; versions from 2.6.3 include a pre-render sweep for the remote-encoding path.
Frequently asked questions
Do I need an LLM to make a multishot video?
No. You can write the shot script yourself, with --- between shots. The CORE workflow does not require the optional writer.
Will a video reference copy my actor’s movements?
No. It guides the look and setting of the scene, not pose or motion.
Should I set a large manual VRAM reserve before my first render?
No. The author recommends starting at 0 so the automatic planner can measure the render. Adjust memory controls in response to a specific warning or spill.
Is a fast draft preview suitable for delivery?
No. H3 TAE Decode is for quickly judging composition and motion; use the normal VAE decode for the final video.
Can the extension prevent all drift in an arbitrarily long scene?
No. Levelling and pin controls help, but generated detail and sound can still change over repeated joins. Preview and listen to long chains before treating them as finished.
Learn More about ComfyUI-H3-Multishot
- The author’s five-minute MiniMax-H3 guide — a shorter route from an idea to a talking-video render.
- Every setting explained — a deeper reference when you are ready to tune continuity, memory, and output settings.
- ComfyUI-H3-Multishot on GitHub — the extension’s workflows, documentation, and release information.
- ComfyUI-H3-Multishot on Civitai — the extension’s model page and release context.
- ComfyUI-H3-Motion-Context — read this when using
context_pin, particularly for guidance about continuous audio, pinned frames, and version compatibility.
