H3 Continuum

ComfyUI-H3-Continuum 3.9.8 — V3.9

Generate long-form MiniMax H3 video and audio in chunks, review each result, retry a section, and continue from compatible saved Takes. V3.9 is the current workflow; V3.8X2 remains available separately.

Download v3.9.8 · Release notes · Checksums · 日本語

Getting started video

Watch: H3 Continuum V3.9 — Chunks, Prompts & Reference Images — 12:26, with English narration, subtitles and chapters.

The video covers First Image setup, chunk settings, Review/Take retries, Timeline prompts and Reference Images. For screenshots and every control, open the illustrated user guide.

Install or update

Use one Continuum installation in ComfyUI/custom_nodes. Models and third-party custom nodes are not bundled.

Fixed-version ZIP

1. Download the custom-node ZIP above and optionally verify its checksum.
2. Extract its package folder into ComfyUI/custom_nodes. When updating, replace the existing package; do not leave duplicate copies.
3. Restart ComfyUI and hard-refresh the browser (Ctrl+F5).
4. Open a V3.9 workflow below and select your installed models and input files.

Keep your saved workflows, Run/Take folders and settings when updating. Workflow ZIPs contain graphs, not the custom-node installer.

Git or ComfyUI-Manager

For a new Git installation:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

For an existing Git installation:

cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main

If Manager owns your installation, update that copy through Manager. Restart ComfyUI and refresh the browser after updating. Registry review, Manager's selected version and GitHub Releases are separate; the fixed-version links above identify the exact v3.9.8 package.

The package declares Python 3.10+ and ComfyUI 0.32.0+. The minimum declaration is not a GPU acceptance result for every Core version.

Required model files and extensions

Download the weights separately from Comfy-Org / MiniMax-H3. The table follows the saved V3.9 template selections; folders are relative to ComfyUI/.

| Component | File / download | Folder | | --- | --- | --- | | H3 diffusion model | minimax_h3_fl2va_pruned_int8_convrot.safetensors | models/diffusion_models/ | | Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | models/text_encoders/ | | Video VAE | minimax_h3_video_vae_int8_convrot.safetensors | models/vae/ | | Audio VAE | minimax_h3_audio_vae_fp32.safetensors | models/vae/ |

Select the downloaded files in the loaders; set the text encoder's type to minimax. Reselect paths if the saved minimax/ subfolder differs from your installation. The Video VAE loader selects INT8, although its older embedded download metadata names FP16; choose the VAE you actually installed. Other model variants must match the generation task and loader support. Quantized variants require compatible Core/GPU support; these selections are not a universal hardware guarantee. The retained V3.8X2 graph instead saves DasiwaMinimaxH3_dasiwaHybridV2_int8.safetensors and FP16 Video VAE. That custom model is not bundled or supplied by the table above; reselect a compatible H3 model you have installed.

The complete supplied graph also needs these external node packages:

• Spectrum — optional acceleration; saved disabled. Its presence does not imply every version is reuse-compatible; see the compatibility table below.
• KJNodes — supplies the template's H3 SageAttention patch. Follow its attention dependency requirements, or replace/bypass that patch if unavailable.
• rgthree-comfy — supplies Power Lora Loader (rgthree).

Optional Turbo: download one task-matched LightX2V Turbo LoRA into models/loras/; the standard workflows start with all entries disabled. The separate Turbo Sample enables one LoRA. Follow that LoRA's sampler/step instructions; listed files and setup explain the choices. Models and extensions are not installed by Continuum.

Choose a workflow

| Workflow | Download | Use | | --- | --- | --- | | V3.9 | JSON / ZIP | Current Sampler and fixed Reference Image slots | | V3.9 Turbo Sample | JSON / ZIP | Auto, 2 x 8s dance and Thai rap; setup and limits | | V3.8X2 | JSON / combined ZIP | Retained workflow; combined ZIP also contains V3.9 |

They are also available in Templates → ComfyUI-H3-Continuum. V3.8X2 graphs and their Reference wiring are not automatically converted to V3.9. Keep the original graph and open the dedicated V3.9 graph for new work.

The templates reference external Spectrum, rgthree and KJNodes nodes. Install, replace or remove missing optional dependencies before Queue. In the standard workflows, Spectrum and Turbo LoRA entries are saved disabled; ComfyUI-Easy-Use is not required.

In the V3.9 template, select your own First Image file before Queue. Reference loaders 1–6 start bypassed and all per-chunk assignments start off. Slots 7–9 require you to add and connect loaders. See template setup and screenshots.

First generation

Start with a small two-chunk test before a long sequence.

1. Select your H3 model, text encoder, sampler and appropriate VAE files.
2. Write Sequence Prompt; use Prompt Format = Fixed for one shared prompt.
3. For image-to-video, use Size Source = First Image and choose a resolution. The template starts at Draft — 0.30 MP. For exact Width/Height, use Manual.
4. Set Chunks to 2, Seconds per Chunk to 5, and check Total Length = 10 seconds.
5. Keep Base Seed unchanged and Control After Generate set to fixed.
6. Set Run to Review Each Chunk and Progress to On — Resume and Takes available.
7. Leave Regenerate From = Auto, read Ready to Queue, then press ComfyUI's top-right Run button.

Chunks is the final target count, not the number to add on the next Queue. Review normally generates one review unit at a time: the first output in this example is 5 seconds, even though the target is 10 seconds. With Last Image terminal merging, the final two logical chunks may form one unit.

For uninterrupted generation, select Generate Full Video instead. Keep Progress on if you want saved Takes, compatible resume and later regeneration.

Review, continue and retry

Selecting a Review action does not start sampling. Select it, then press ComfyUI's Run button (Queue) to execute it.

| Action | Next Queue | | --- | --- | | Use it and continue | Keep accepted chunks and generate the next review unit | | Try this chunk again | Keep earlier chunks and create another Take of the reviewed unit | | Use it and finish the rest | Keep accepted chunks and generate all remaining units | | Back to Settings / Return to Review | Navigate between settings and review; neither queues work nor deletes Takes | | Use This Take | Select a stored result without new sampling; decoding/saving may still run | | Continue From Here | Branch from the selected Take and generate the next review unit |

For the 2 × 5-second example, continue after Chunk 1 to reuse it and generate Chunk 2; the completed result is 10 seconds. At completion, retrying the current unit creates another Chunk 2 Take.

Keep the same model/LoRAs, seed, prompts, media, dimensions, sampling schedule, continuation settings and Run identity while continuing. Keeping only the seed fixed is insufficient to guarantee reuse. Read the status/reuse counts before treating a result as a successful continuation.

To extend a completed sequence, keep those conditions and Regenerate From = Auto, increase the final Chunks target, and Queue again. Illustrated Review, Take and extension steps.

Regenerate from a chosen chunk

1. Open settings and choose Regenerate From = Chunk N.
2. Keep Progress on and earlier-chunk generation inputs unchanged.
3. Choose Generate Full Video to rebuild N through the target in one Queue, or Review Each Chunk to pause after the next review unit.
4. Set Variation Nonce and press ComfyUI's Run button.
5. Restore Regenerate From = Auto before ordinary continuation or Take actions.

Nonce 0 requests another automatic variation after a completed regeneration; a compatible interrupted regeneration resumes its existing variation. A positive Nonce fixes that variation: repeating it with identical inputs does not request a new result. Review retry manages its own variation separately.

Earlier compatible chunks and previous Takes are retained. Later chunks follow the changed continuation, so Full Run from Chunk 2 may regenerate more than Chunk 2 when the target contains additional chunks. Terminal-merged pairs are atomic; arbitrary frames inside a chunk cannot be edited. Regeneration details and compatibility.

Prompts and Reference Images

• Fixed: one shared prompt; List: per-chunk prompts; Timeline: time-based instructions.
• V3.9 Reference Images uses fixed Image 1–9 slots and one Sampler connection. All chunks shares connected images; Per chunk lets you tick assignments.
• @R1 always means Image 1, even if other slots are empty. Use tags only for active References; resolved picture numbers can differ by physical group.
• First/Last Image, Reference Images, Reference Audio and Driving Audio have different roles. Driving Audio preserves a source track; Reference Audio conditions generation.
• For a Reference-only test, turn First/Last Image off and choose Manual size. A terminal-merged final pair needs the same effective Reference set in both chunks.
• Pre-resize large source images: conditioning resize does not eliminate source-loading RAM use. A subject may persist through video continuity after its assignment ends.

Reference guide · Prompt examples · V3.9 prompt skill · General prompt resources

Storage, decoding and Second Pass

Progress saves Run/Take data for compatible reuse. The optional Decode Cache Helper caches decoded results; it is separate from saved sampling history. Finalize assembles the output; downstream Core nodes create and save the video.

Set scratch storage under Settings → H3Continuum → Storage → Manage temporary storage. The folder is on the ComfyUI server and applies to Finalize and Decode Cache. It does not move saved Runs/Takes or finished videos. Change or clean it only with empty running/pending queues. Moving folders on one physical drive does not reduce that drive's writes. Storage modes, cleanup and limits.

Second Pass is an optional bridge for external refinement/upscaling. It is separate from chunk continuation and ordinary Review retries. The V3.8X2 standard workflow does not require Hi-Res Fix. V3.8 Second Pass does not support partial Review sequences; V3.9 can inherit validated conditioning for its actual output groups. Integration contract.

Compatibility and v3.9.8 scope

| Area | Supported boundary | | --- | --- | | Saved history | Existing Takes stay readable. Old v5 or unsafe history is not repaired or promoted into reusable Sampling v6 history; start a new Run when required. | | Unknown MODEL wrappers | Fresh generation remains available; reuse requires observable, matching contracts. No blanket approval by class name or version. | | Attention / Spectrum | Finite audited profiles support Core PyTorch/Kitchen and specified Spectrum implementations, including upstream v0.2.29. Exact scope. | | KJ Model Preview Override | Only the pinned KJ/Core source profile: VAE / Audio VAE disconnected, Tiny VAE none, Core NoPreviews/Auto/Latent2RGB. KJNodes 1.5.2 alone is insufficient; the audited local source differs from upstream. Supported hashes. |

Published v3.9.8 passed new-Run, separate-Queue Chunk 1→2 GPU continuation with PyTorch + KJ and Comfy Kitchen INT8 + audited local Spectrum + KJ. Preview images remained held, original Chunk 1 data/contracts stayed unchanged, and all four video/audio outputs decoded successfully. These were 512×512, 24fps, 2×5-second, 8-step Turbo Core-API tests on an RTX 5060 Ti 16 GB. They do not establish support for other KJ/Spectrum builds or preview-VAE paths. Full conditions and verification limits.

Known limitations

• Reference Plan Inspector: stale display after repeated Sampler configuration remains a known issue; the two expected test failures do not mean it is fixed.
• Mobile: reduced UI work is verified; mobile freeze resolution is unconfirmed.
• Resources: nine high-resolution References or long final decoding can exhaust VRAM/RAM. Disk-backed Finalize does not prevent allocations in other nodes.
• Refinement: high-ratio latent resize can cause artifacts; longer Hi-Res/Hybrid cases and doubled outlines remain limited or under investigation (Issue #27).
• Continuity: exact identity, motion, lip sync and frame-by-frame reproduction are not guaranteed. Some long continuations have known quality limits.
• Acceptance: browser Run-button/save-reload, the reporter's exact full graph, other provider builds, subjective quality and speed/VRAM benchmarks were not established by the targeted v3.9.8 GPU recheck.

Detailed documentation

| Topic | Read more | | --- | --- | | Setup, screenshots and complete controls | Illustrated user guide | | Current Reference Images | V3.9 Reference guide | | Updating from V3.8X2 | Migration and retained contracts | | History, earlier screenshots and implementation notes | README history / CHANGELOG | | Current release and GPU verification | v3.9.8 notes / package validation | | Bug reports | GitHub Issues — include the workflow, status text and auto-resume disabled: reason |

Historical Releases and tags remain available. Documentation updates on main do not replace the published v3.9.8 tag or package assets. This condensed README changes documentation only; public nodes, official workflow JSON/ZIP files, Sampling v6 and existing Runs/Takes are unchanged.

License

MIT. Third-party components retain their own licenses and notices.