ComfyUI-H3-Continuum Introduction
ComfyUI-H3-Continuum helps you create longer MiniMax H3 videos with audio by generating them in chunks. Instead of committing to an entire sequence at once, you can review a short section, keep it or try it again, then continue. Saved results, called Takes, let you return to an earlier choice when the saved generation data is compatible with your current settings.
This approach is useful when you want to:
- Build a scene one action at a time and check its direction as you go.
- Give each chunk a different prompt while maintaining a shared subject or setting.
- Try alternative endings without discarding an earlier chunk.
- Organize separate creative ideas as Projects and trials.
V3.9 is the current workflow. V3.8X2 remains available as a separate, retained workflow; opening it does not convert its graph to V3.9. Start with the V3.9 workflow, or watch the 12-minute getting-started video for a visual introduction to chunks, prompts, review controls, and Reference Images.
How ComfyUI-H3-Continuum Works
Think of a Continuum Run as a sequence of short scenes. If you choose 2 chunks at 5 seconds per chunk, the target video is 10 seconds long. In Review Each Chunk mode, the first Queue normally produces only the first 5-second section. You can inspect it before asking ComfyUI to make the next section.
A Take is a saved candidate for a generated section. After Chunk 1, you might continue to Chunk 2, then decide its action needs changing. You can retry that section while keeping compatible earlier work. You can also select a saved Take and branch from it to explore a different continuation. Some final sections that use Last Image can be generated as a combined unit, so a review unit is not always exactly one logical chunk.
Continuing from saved work is subject to compatibility checks. Keep the Run identity and earlier-chunk inputs consistent, including the model, LoRAs, seed, prompts, source media, dimensions, sampling settings, and continuation settings. A fixed seed alone does not guarantee reuse. Read the Sampler’s status and reuse counts to confirm what was retained and what was generated again. Older or unsafe saved history may remain viewable without being reusable.
The optional Project / Take Manager adds organization around this process. A Project can represent a film or scene; a trial is one independent generation history; and a named savepoint points to a particular Take and the chunks leading to it. Savepoints do not copy the underlying media and are not backups. Manager prepares a selection, but you still press ComfyUI’s normal Run / Queue button to generate anything.
ComfyUI-H3-Continuum Features
Prompt each section
Write your instructions in Sequence Prompt before the first run. Choose a format that matches the way you plan your scene:
- Fixed uses one prompt throughout.
- Timeline uses chunk headers or time ranges. Text before the first header is shared across chunks.
- List uses separate prompt blocks divided by
---. Repeat important shared details in each block. - Auto can detect prompt formats, though selecting Timeline or List explicitly can make an early test easier to understand.
For example, with two 5-second chunks, a Timeline prompt could be:
subject_definitions: The woman in the reference image.
retention_analysis: [reference generation] Keep her face and red coat.
overall_soundscape: Soft wind and flowing water.
[0-5s]
She walks to the railing and rests her hand on it.
[5-10s]
With her hand still on the railing, she turns to look at the river.
You can use [Chunk 1] and [Chunk 2] instead of time ranges. Time ranges should begin at the start of the video and match the chunk duration chosen in the Sampler. For more examples, see the prompt guidance and V3.9 reference-prompt resource.
Choose how to generate and review
For a first test, set Chunks = 2, Seconds per Chunk = 5, Run = Review Each Chunk, and Progress = On — Resume and Takes available. Confirm that Total Length shows 10 seconds, leave Regenerate From = Auto, and read the Ready to Queue status before pressing Run. The template begins at Draft — 0.30 MP; you can use Size Source = First Image for image-led sizing or Manual for exact width and height.
After a review result appears, choose what should happen on the next Queue:
| Review action | What it does |
|---|---|
| Use it and continue | Keeps the accepted work and generates the next review unit. |
| Try this chunk again | Keeps earlier chunks and makes another Take of the current unit. |
| Use it and finish the rest | Keeps accepted work and generates the remaining units without another review pause. |
| Use This Take | Selects a saved result without new sampling, although decoding or saving may still occur. |
| Continue From Here | Branches from the selected Take and generates the next review unit. |
| Back to Settings / Return to Review | Changes views without generating or deleting Takes. |
Choose Generate Full Video when you do not want to pause for review. To extend a completed sequence, keep the compatible settings, increase Chunks to the new final target count, and Queue again. Chunks is not a “number of extra chunks” setting.
Regenerate a later section
Regenerate From = Chunk N lets you keep compatible earlier chunks and rebuild from a chosen point. With Progress on, choose Generate Full Video to rebuild through the target in one Queue, or Review Each Chunk to pause after the next review unit. Set Variation Nonce, then Queue. Restore Regenerate From = Auto before ordinary continuation.
Nonce 0 requests another automatic variation after a completed regeneration; a positive value fixes the requested variation, so repeating it with identical inputs does not request a new one. Review retries manage their variation separately. Regeneration may replace several later chunks, not just Chunk N, because those chunks depend on the changed continuation. It cannot edit an arbitrary frame inside a chunk. See the regeneration guide.
Guide the video with images and audio
The V3.9 workflow has Image 1–9 Reference slots. You can share connected images across All chunks, or use Per chunk assignments to specify where each reference applies. A tag such as @R1 always names Image 1, even if other slots are empty. The supplied template starts with Reference loaders 1–6 bypassed and per-chunk assignments off; slots 7–9 need added and connected loaders.
First Image, Last Image, Reference Images, Reference Audio, and Driving Audio serve different purposes. In particular, Reference Audio guides generation, while Driving Audio preserves a source track. For a Reference-only experiment, turn First/Last Image off and choose Manual size. A subject can remain visible after its Reference assignment ends because the video itself carries visual continuity; assignments do not guarantee exact identity or motion. See the V3.9 Reference guide.
Organize Projects, trials, and Takes
The optional Project / Take Manager V1.2 is off by default. Enable it in Settings → H3Continuum → Project Manager, then open H3 Continuum → Project / Take. It uses the existing V3.9 Sampler and does not require another graph node.
To preserve Chunk 1 and explore two versions of Chunk 2, generate and select the Chunk 1 Take in Render History, attach its Run to a Project trial, and choose Name selected Take. Select that savepoint and click Prepare from savepoint, check the Sampler, then Queue. Prepare the same savepoint and Queue again for another continuation. Previous Takes remain available.
Use New trial and Prepare new trial for an independent idea that starts at Chunk 1. A new trial does not copy the previous trial’s prefix. Typing a name alone neither creates an item nor switches the Sampler: use its action button and confirm the result. You can rename, archive, restore, and refresh management entries without deleting the underlying videos or Takes. Save your workflow separately, because Manager does not restore models, prompts, references, or other graph settings. The illustrated Manager walkthrough explains the controls and safeguards.
Save and refine results
With Progress on, Continuum retains Run/Take data for compatible continuation. The optional Decode Cache Helper stores decoded results separately from sampling history. Temporary storage settings apply to Finalize and Decode Cache, not saved Runs/Takes or finished videos; only change or clean that location when no jobs are running or pending.
Second Pass can optionally refine video at the same resolution or after a compatible external upscaler. It does not include an upscaler. Its audio, Plan, and context connections must match the documented workflow. See the Second Pass wiring and upscaling guide.
ComfyUI-H3-Continuum Models
Continuum works with separately obtained MiniMax H3 models; it does not bundle model weights. The supplied V3.9 template names these components:
| Component | Saved V3.9 selection | Role |
|---|---|---|
| H3 diffusion model | FL2VA INT8 model | Generates the video-and-audio content for the template’s task. |
| Text encoder | Qwen3VL 32B H3 NVFP4/AWQ | Processes your text instructions; select encoder type minimax. |
| Video VAE | H3 Video VAE INT8 | Helps turn generated video data into viewable frames. |
| Audio VAE | H3 Audio VAE FP32 | Helps turn generated audio data into a playable track. |
Choose files you actually have in the workflow loaders. Other MiniMax H3 diffusion variants must suit the generation task and loader. Quantized variants depend on compatible ComfyUI Core and GPU support; the template selections are not a guarantee for every machine.
The retained V3.8X2 graph saves a different custom hybrid H3 model selection and an FP16 Video VAE. That custom hybrid model is not included with Continuum or in the V3.9 model list. If you use V3.8X2, select a compatible model available to you.
An optional, task-matched LightX2V Turbo LoRA can be used with the appropriate sampler and step settings. Standard workflows leave Turbo LoRA entries disabled; the Turbo Sample workflow demonstrates a separate enabled example. The supplied graph also references external KJNodes, rgthree-comfy, and optional Spectrum nodes.
Spectrum is an optional accelerator, not another H3 model. It predicts some intermediate sampling work from earlier actual model evaluations. That can reduce expensive evaluations, but it is approximate: the image, motion, or audio may differ even with the same seed. The standard Continuum workflows save Spectrum disabled. If you compare it, keep the seed, prompts, checkpoint, references, sampler, schedule, size, duration, and LoRAs identical, and judge both speed and the resulting media. Continuum only reuses saved chunks when its compatibility checks accept the particular configuration; Spectrum’s presence alone does not make every version reuse-compatible.
What's New with ComfyUI-H3-Continuum
The following changes are documented in the supplied release history. V3.9 remains the workflow name across the 3.10.x package updates.
| Package version | What changed and why it matters |
|---|---|
| 3.10.2 | A hotfix preserves finalized Takes and named savepoints after interrupted or failed continuation by using separate storage revisions for new work. It also improves handling of regeneration requests, supports Manager operations on HTTP LAN pages through a UUID fallback, and keeps accepted Second Pass video on CPU rather than retaining earlier outputs in VRAM. It does not automatically repair previously interrupted history. Release details. |
| 3.10.1 | Project / Take Manager now displays the actual reason for a refused operation. Read-only Takes, integrity problems, busy Runs, and capacity limits are no longer all described as another-tab conflicts. Release details. |
| 3.10.0 | Introduced optional Project / Take Manager V1.2 for named Projects, independent trials, and fixed Take savepoints. Existing generation still works with Manager off; no workflow rewiring is required. Release details. |
| 3.9.12 | The Reference Plan Inspector stops showing stale verified information after Reference assignments change or a workflow is reloaded. A new backend result supplies the next verified plan. Release details. |
| 3.9.11 | Fixed a cause of false read-only Review history when unchanged audited Spectrum code was imported repeatedly. Supported saved Takes can continue without loosening checks for unknown code. Release details. |
| 3.9.10 | Improved Spectrum reuse checks for line endings and supported Python profiles. The Reference Images assignment table also shows the proper 1–16 range when Chunks comes from an external input; that display change does not generate extra chunks. Release details. |
| 3.9.9 | Addressed a supported KJ Preview Override configuration that could incorrectly make a new Run read-only, and made saved reasons for unavailable reuse visible. Release details. |
These releases include targeted tests, but their results do not promise the same visual quality, speed, memory use, or compatibility for every GPU, model variant, and third-party node combination. Published package details are on the v3.10.2 release page.
Troubleshooting ComfyUI-H3-Continuum
A chunk did not continue from my saved result
- Check the Sampler’s status and reuse counts to see whether earlier work was actually reused.
- Confirm you selected the intended Run and Take.
- Compare the model, LoRAs, seed, prompts, references and other source media, dimensions, sampling settings, and continuation settings with the saved work.
- If the history is marked read-only or incompatible, restore the supported source conditions or begin a new independent Run at Chunk 1. Do not edit reuse flags or delete history to force continuation.
I selected a Review action, but nothing happened
Review actions prepare the next operation. After choosing Use it and continue, Try this chunk again, or Continue From Here, press ComfyUI’s normal Run / Queue button. Back to Settings and Return to Review only change the view.
Manager cannot name a Take or prepare a trial
- Select the intended Target Sampler, Project, and trial.
- For Name selected Take, first select a Take in Render History and confirm the Manager trial belongs to the same Run.
- For Prepare new trial, select a trial created with New trial, not an attached existing Run.
- If a job is queued or running, wait for it to finish before preparing another action. Connected identity inputs or conflicting Run identities may also prevent preparation.
- Read Manager’s refusal message. Refresh its data after changes made in another tab.
The Manager walkthrough and screenshots show the panel, savepoint controls, and existing Review view.
A savepoint is listed, but its source is missing
A named savepoint points to original Run metadata and raw data; it is not a backup. Keep the source Run and its files. If that source is missing or corrupt, the savepoint cannot recreate it.
Reference Images do not seem to affect a chunk
- Check that the intended image loader is connected and active, rather than bypassed.
- Confirm All chunks or the correct Per chunk assignment is selected.
- Match prompt tags to the fixed slot numbers:
@R1means Image 1. - For a Reference-only test, turn First/Last Image off and use Manual size. Consult the Reference guide for assignment examples.
An image assignment guides generation; it does not guarantee an exact likeness or prevent a subject from persisting into later chunks.
I am running out of memory
Start with fewer chunks, lower resolution, and fewer or smaller source images. Large images still consume memory when loaded even if conditioning later resizes them. Long final decoding or many high-resolution References may exhaust RAM or VRAM, and disk-backed Finalize cannot prevent allocations by other nodes. Use the storage guidance to understand which temporary files its setting affects.
Frequently asked questions
Can I try a new Chunk 2 without losing Chunk 1?
Yes. Keep a compatible Chunk 1 Take, then retry the current unit or prepare a savepoint at Chunk 1 and Queue a new continuation. Earlier compatible work and previous Takes remain available.
Does “Chunks = 2” add two more chunks to an existing video?
No. Chunks is the final target count. Increase it beyond the completed count to extend a sequence.
Does a Project save my complete workflow?
No. Projects and savepoints organize references to Runs and Takes. Save the workflow separately to retain its models, prompts, references, and other graph settings.
Can I switch between V3.8X2 and V3.9 automatically?
No. V3.8X2 Reference wiring is not automatically converted. Keep that graph for older work and open the dedicated V3.9 workflow for new work.
Will Spectrum always produce the same video faster?
No. Spectrum is an approximate accelerator, so results may change. Its supported sampling path and Continuum reuse compatibility depend on the actual configuration.
Learn More about ComfyUI-H3-Continuum
- Getting-started video: a visual tour of First Image setup, chunks, prompts, review, and Reference Images, with English narration, subtitles, and chapters.
- Illustrated user guide: screenshots, workflow setup, and explanations of the controls.
- V3.9 Reference Images guide: slot assignments and prompt-reference behavior.
- Project / Take Manager guide: illustrated steps for Projects, trials, savepoints, and preparation.
- Second Pass guide: connections, output scope, and compatible external upscaling.
- README history and changelog: earlier prompt examples and change history.
- GitHub Issues: existing reports and a place to report a reproducible problem. Include your workflow, Sampler status, and any
auto-resume disabled:reason so others can understand what happened.
