ComfyUI-H3-Continuum 3.8.3 — V3.8X2
V3.8X2 is the product and workflow label for package3.8.3. It keeps the V3.8 production sampler and adds optional Reference Images 4–9 plus the built-in Decode Cache Helper. The older V3.8.0 package remains available from tagv3.8.0.
Official V3.8X2 workflows
The two names contain the same official graph; they are distribution aliases, not different configurations. Both include the built-in Decode Cache Helper with its savedAuto / RAM 256MB / Disk 8GB / reset_token 0 settings and the nine-reference helper connection. Each ZIP contains only the correspondingly named JSON. Older V3.8X workflows remain available for existing projects.
Nine Reference Images: memory warning
Reference Images 1–3 connect directly to the Sampler. Optional images 4–9 connect through H3 Continuum Reference Images and then to Reference Images (Optional). Using all nine is a high-resource configuration, not the normal minimum requirement.
The acceptance run completed on an RTX 5060 Ti 16 GB / 64 GB system only after the nine source references were kept near 0.30 MP each. With 512 × 608 output, 3 × 5 seconds, Spectrum Off, LoRA Off, Sage Attention, and Balanced 22-frame continuity, Sampling used about 15.0–15.5 GiB of 16,311 MiB VRAM. There was little 16 GB headroom.
When using many references, especially all nine, resize copies of the source images to about 0.30 MP before loading or connecting them. The Sampler still performs its normal conditioning resize, but smaller source files reduce input and preprocessing pressure. Larger references, larger output sizes, extra wrappers, or concurrent GPU work may require a higher-spec GPU and more system RAM. Nine references are optional and are not guaranteed to improve quality over three.
Built-in Decode Cache Helper
Important: this is not an unconditional new-generation speedup
Manual cache clear: normally no action is needed. Press the Helper's "キャッシュをクリア" (Clear cache) button to discard that Helper's cache on the next Queue and run native Decode again. It does not delete immediately or queue automatically. The internal reset_token remains serialized for workflow/API compatibility and increments once per click (wrapping to 0 at the INT limit). Creating, loading, or cloning a node does not change it. Without JavaScript, or in API workflows, change reset_token through the traditional numeric input.
The Helper saves time only when unchanged latents are decoded again. It does not accelerate Sampling. A first MISS runs native Decode and stores the result; hashing and storage can make that run slower. Generating different latents with a new random seed each time normally produces MISS, so this feature should not be expected to speed up that usage.
Concrete reuse scenarios:
- Continue by adding chunks: when existing chunk latents are retained and decoded again for output, cached existing chunks can HIT; newly generated chunks initially MISS.
- Fix / partial regeneration: unchanged chunk latents can HIT when submitted again; regenerated chunks whose latent contents changed MISS.
- Complete / Resume output: decoding retained latents again without resampling can benefit. Merely playing an existing video file is not a cache use case.
- Retry from Decode onward, or change Finalize/output settings: reuse applies when the same latents and decode-related metadata reach the Helper and Decode runs again.
- Regenerate with a fixed seed: reuse requires identical resulting latent contents. The same seed alone does not guarantee HIT.
Auto or RAM. In its report, hit means reuse, miss means native Decode, and off / bypass mean no cache reuse.
Cache reuse is process-local and does not survive Python restarts. Eviction and VAE identity changes can also cause MISS. Changing mode, RAM/Disk budgets, or reset_token clears this implementation's Helper cache. Keeping reset_token = 0 does not reset it on every Queue.
V3.8X2 ships H3DecodeCacheHelper inside this Continuum package. It remains a separate public node under MiniMax H3/Continuum/Helpers; no second custom-node addon is needed. The nine public node IDs are the seven V3.8X IDs plus Reference Images and this helper. Sampler, Finalize, Assembly Plan, Core Decode, and saved Sampler widget/socket contracts are unchanged.
Connect Sampler video_latents and audio_latents plus their native Video/Audio VAE inputs to the Helper, then connect Helper images and audio to Finalize. Keep the Sampler assembly_plan directly connected to Finalize. The official V3.8X2 workflow preserves its saved graph and settings. It contains no Core VAE Decode nodes, so returning to Core direct Decode requires adding those two Core nodes and reconnecting Finalize, or loading a prior Core-direct V3.8X workflow. Do not simply remove the Helper and expect automatic rewiring.
Auto caches full physical Video decodes on private process-local disk and small Audio decodes in bounded RAM. Off delegates native Decode and clears the Helper cache. Cache failure falls back to native Decode; native Decode, OOM and queue interruption are not retried. Files are not reused across Python restarts. Keep only one provider of the H3DecodeCacheHelper node ID: the previous standalone addon must be disabled or removed during the migration, after a verified built-in installation.
At 736×416 (0.306MP), 3×5 seconds, Euler/simple 6 steps, Turbo FL2V v1.2, a warm backend and three Video HITs, one measured configuration produced a mechanistic estimate of about 47 seconds / 24.5% less Full Run Total. One direct A/B pair observed 50.887 seconds / 26.46%, and two Complete/Resume re-decode pairs averaged 46.309 seconds / 79.826% less Total. These are narrow observations, not repeated universal averages or an unconditional 20%+ speed promise. The first MISS can be slower, new Sampling is not accelerated, and a changed latent or VAE identity produces a MISS.
Historical V3.8X workflow: JSON | ZIP — Published releases
✅ V3.8X on main includes the September 8, 2026 Review hotfix: Review Each Chunk continuation, stale Regenerate From state, completed-sequence extension, and Render History queue handling are repaired. Update with git pull --ff-only origin main (or ComfyUI Manager Update), restart ComfyUI, and hard-refresh the browser if the old UI remains. Existing saved Takes are preserved. Issue #13 remains a separate open long-continuation quality issue and is not part of this hotfix.
H3 Continuum is a Production Sampler for generating, reviewing, partially regenerating, and resuming long-form MiniMax H3 video without restarting the entire work. V3.8 has two product layers: Main / Production and Advanced.
Install or update first
For a new Git installation: ``bash
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git
`
For an existing Git checkout:
`bash
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
`
Restart ComfyUI after cloning or pulling. If ComfyUI Manager installed the node, use its Update action instead of mixing Manager updates with a second Git checkout.
To install the historical V3.8.0 package into a separate checkout instead:
`bash
cd ComfyUI/custom_nodes
git clone --branch v3.8.0 --single-branch https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git ComfyUI-H3-Continuum-v3.8.0
`
Use only one Continuum checkout at a time. ComfyUI Manager Update tracks main; it does not select historical tags.
Download the current V3.8X2 workflow: JSON or ZIP containing the same JSON. This is one Spectrum-capable graph, also usable with LightX2V Turbo; both Spectrum and all Turbo LoRA entries are saved disabled. Install its external Spectrum, rgthree, and KJNodes nodes before opening it. ComfyUI-Easy-Use is not required. See Spectrum and Turbo setup below.
Prompt and skill downloads
- LLM system prompt ZIP — system instructions and reference material for Continuum prompt authoring.
- Continuum prompt skill ZIP — general chunk-aware prompt authoring for Codex and other compatible agents.
- Continuum Dance Director skill ZIP — long-form choreography prompts for one primary dancer, with body and camera continuity across chunks.
These are optional prompt-authoring resources, not ComfyUI custom nodes. Extract each ZIP and follow its included instructions. They do not change the sampler or guarantee generated motion or image quality.
You can ask an AI to read this manual
This README is intentionally detailed. You do not have to memorize it. Give its GitHub URL or this file to a web-enabled AI, describe your target video, GPU/VRAM, available First Image/audio, and whether you want to review every chunk. Ask it to use the exact V3.8 labels below.
Example request:
`text
Read the current ComfyUI-H3-Continuum V3.8 README. I have a 16 GB GPU and want
to create a 30-second I2VA video, reviewing every chunk. Tell me exactly which
Continuum inputs and controls to use, what must stay fixed between Queue runs,
and what output length I should expect after each action. Do not use older V3.7 UI.
`
AI products, plans, and repository-reading behavior change frequently. If a URL is not read correctly, paste or upload this README instead. Do not send private workflows, local paths, tokens, or credentials to a public AI service. The README and the UI remain the source of truth; an AI summary can still be wrong.
V3.8 supported surface
This local package exposes nine searchable nodes: the original seven V3.8 nodes plus Decode Cache Helper and Reference Images.
- H3 Continuum Sampler V3.8 — the Main sampler
- H3 Continuum Finalize — decoded Video/Audio assembly with optional seam handling
- H3 Continuum Load Image
- H3 Continuum Load Audio
- H3 Continuum Load Video
- H3 Continuum Second Pass — the Advanced bridge for external latent processing or upscaling
- H3 Continuum Reference Audios — an ordered Reference Audio 1/2/3 bundle helper
- H3 Continuum Reference Images — optional additional Reference Images 4–9
- Decode Cache Helper — the existing optional decoded-result cache
The current frontend uses Show Advanced Settings / Hide Advanced Settings to change presentation without changing generation values. Saved drafts that still contain the former frontend-only H3 Continuum View property are migrated to the matching collapsed/expanded view and the legacy property is removed. If the frontend extension is unavailable, the complete Python-defined interface remains visible and executable.
The official V3.8X2 workflow uses the Helper path:
`text
H3 Continuum Sampler V3.8
-> Decode Cache Helper
-> H3 Continuum Finalize
-> Create Video
-> Save Video
`
The Helper delegates cache MISS entries to native Core Video/Audio Decode. A manually wired direct Core Decode path remains supported, but it has no decoded-result cache. Save and upscaling remain external. Hi-Res Fix is not part of the V3.8X2 standard workflow; use Second Pass as the bridge when an external latent processor or upscaler is needed. External processors may change only Video LATENT spatial geometry while preserving physical groups, B/C/T, finite values, first-pass Audio, and the original Assembly Plan. Finalize accepts public IMAGE/AUDIO outputs and does not require a specific decoder class. SageAttention, Sol-Attn, and Spectrum remain external MODEL wrappers. See the V3.8 Open Integration Contract.
Complete V3.8 UI reference
This section covers every user-facing control and socket on the seven public V3.8 nodes. Names such as Power Lora Loader (rgthree), Spectrum, SageAttention, Core VAE Decode, and Core Save Video belong to ComfyUI or another extension. They may be used in a supplied workflow, but they are not Continuum controls.
H3 Continuum Sampler V3.8: required graph inputs
| Input | What it does | Normal use |
|---|---|---|
| model | MiniMax H3 diffusion model, including any upstream MODEL wrappers or selected LoRA | Required |
| clip | MiniMax H3 text encoder | Required |
| video_vae | Encodes connected image conditioning; Continuum does not use it for final decoding | Required by the node; unused by pure T2VA |
| sampler | ComfyUI sampling algorithm | Required |
| sigmas | Noise schedule and effective step count | Required |
| Sequence Prompt | Complete text, list, timeline, or JSON sequence prompt | Required |
The Sampler returns six outputs: a list of video_latents, a list of audio_latents, one assembly_plan, a text status, the selected driving_audio when used, and a refine_context for the Advanced Second Pass. Raw latent lists normally go through Core Video/Audio Decode before Finalize.
H3 Continuum Sampler V3.8: optional media inputs
| Input | Role | Important distinction |
|---|---|---|
| first_frame | First Image for I2VA or FL2VA | Can also supply the output aspect ratio when Size Source = First Image |
| last_frame | Optional final-image constraint for FL2VA | Can trigger a Long Terminal Merge; the final pair may become one atomic review unit |
| reference_image_1–reference_image_3 | Ordered appearance, identity, subject, or scene references | They never become the implicit size source |
| Reference Images (Optional) | Additional images 4–9 bundled by H3 Continuum Reference Images | Empty inputs are skipped; encoding remains in the Sampler |
| Video Guide Frames | A video loader's IMAGE frame batch, applied as a persistent guide | It does not carry the source video's audio; frames are interpreted at 24 fps |
| Driving Audio + Driving Audio VAE | Original audio timeline used as native guide conditioning | The selected source audio becomes final audio; generated audio and Audio Seam are bypassed |
| Reference Audio (Optional) + Reference Audio VAE (Optional) | Legacy single conditioning-only audio reference | Generated audio remains final audio |
| Audio References (Optional) | Ordered bundle from H3 Continuum Reference Audios | Do not connect this together with the legacy single Reference Audio path |
| Still Image Guide (Optional) | Compatibility socket inherited from the V3.7 guide contract | Experimental; not part of the V3.8X2 official workflow |
First Image, Last Image, and Reference Images
!First Image enabled and Last Image bypassed
- First Image establishes the opening image and is the normal visual starting point for I2VA/FL2VA. It is also the only media input that can drive
Size Source = First Image.
Last Image constrains the sequence ending. Leave it OFF for T2VA and ordinary I2VA. With a connected Last Image, extending or regenerating the sequence may rebuild the terminal pair.
Reference Images 1–9 guide identity or appearance throughout generation. Inputs 1–3 remain on the Sampler; optional inputs 4–9 are on H3 Continuum Reference Images, connected to the Sampler's Reference Images (Optional) socket. The helper only bundles images: resizing, hashing, and VAE encoding remain in the existing Sampler path. Empty slots are skipped; Picture numbers follow active slot order after any First/Last Image. Existing three-loader templates still work unchanged.
Saved workflows with the earlier direct image 4/5 inputs migrate those links into a bundle after graph loading. Migration preserves the original source links and only removes a legacy input after its new link is verified. If a source is unavailable or its destination is occupied, the legacy link is retained with a warning. Old API prompts remain accepted; supplying both legacy and bundle images for the same slot is an explicit input conflict.
!Three optional Reference Image loaders
Every H3 Continuum Load Image has Enable Image. ON loads through ComfyUI Core. OFF uses native node bypass and makes that optional path behave as unconnected without deleting the node or cable. First, Last, and up to nine Reference Image inputs can therefore be connected without forcing every input to be active. An empty image bundle returns no reference and does not change the zero-to-three-image sampling/reuse contract.
| Public loader | Controls | Outputs |
|---|---|---|
| H3 Continuum Load Image | Enable Image, image file | IMAGE, MASK |
| H3 Continuum Load Audio | Enable Audio, audio file | AUDIO |
| H3 Continuum Load Video | Enable Video, Video file, Force Rate | images, audio |
Video, audio, and their bypass switches
!Continuum Video and Audio loaders with Enable controls
H3 Continuum Load Video: Enable Video, Video, and Force Rate. Force Rate = 0 uses source FPS; a positive value drops/duplicates frames to that rate while preserving nominal duration and audio. Connect images to Video Guide Frames when visual video guidance is wanted. Connect audio to Driving Audio only when the source audio should guide and become the final output.
H3 Continuum Load Audio: Enable Audio and the audio file. Use it for Driving Audio or Reference Audio according to the socket you connect.
The node titled Audio Switch in the example image is ComfyUI Core's If/Else Switch (ComfySwitchNode); Fast Groups Bypasser (rgthree) belongs to the rgthree extension. Neither is a Continuum node or a requirement of Continuum. The Continuum loaders already provide their own native ON/OFF controls.
Audio Continuity is different from both audio inputs: it passes the previously generated audio context into the next generated chunk. It does not import a source track. Driving Audio replaces final generated audio; Reference Audio only conditions generation.
Main Sampler controls
| Displayed control | Meaning | Recommended starting point |
|---|---|---|
| Prompt Format | Auto, Fixed, List, or Timeline interpretation of Sequence Prompt | Auto unless you need to force one parser |
| Continuity | Prior Video context retained at each chunk boundary | Balanced — 22 frames for Standard audiovisual continuation |
| Base Seed | Base for deterministic per-chunk seed derivation | Keep the value unchanged when continuing, comparing, or selecting Takes |
| Control After Generate | What ComfyUI does to Base Seed after a Queue | fixed for Review Each Chunk, resume, and controlled comparisons |
| Audio Continuity | Carries generated-audio context across boundaries | true for normal audiovisual generation |
| Chunks | Planned total chunk count, from 1 to 16 | Enter the final total, not the number to add next |
| Seconds per Chunk | Duration shared by every chunk | 5–15 seconds is the validated range; see duration notes below |
| Total Length | Read-only Chunks × Seconds per Chunk | Check only; it never edits either input |
| Size Source | First Image or Manual output geometry | First Image for I2VA/FL2VA; Manual for T2VA or exact dimensions |
| Resolution | First-Image sizing preset | Draft — 0.30 MP for tests, Balanced — 0.60 MP for more detail, or Native 768 for the H3-native short edge |
| Custom MP | Custom First-Image pixel budget | Visible only when Resolution = Custom |
| Width / Height | Exact output canvas in 32-pixel steps | Visible/editable only in Manual mode; 32–16384 is accepted |
| Run | Generate Full Video or Review Each Chunk | Use Review when you want a human decision after each physical group |
| Progress | Saved raw chunks, safe resume, and Take history | Review turns it On automatically; enable it for resume/regeneration |
| Ready to Queue | Plain-language summary of the next Queue | Read it before pressing ComfyUI's top-right blue Run button |
| Show Advanced Settings / Hide Advanced Settings | Shows or hides technical controls | Visibility only; it does not change their values |
Continuity = Balanced — 22 frames with Continuation Backend = Standard and Audio Continuity = true uses the current Masked AV route. Fast — 5 frames, Strong — 39 frames (Experimental), and Auto — conservative safely resolve to the older Reference Context route while generated-audio continuity is on. With Audio Continuity off, Standard uses the Video-only masked route.
Choose First Image sizing or Manual sizing
!First Image sizing with Draft 0.30 MP
Size Source = First Image preserves the connected First Image aspect ratio. Resolution then chooses its pixel area:
| Preset | Target |
|---|---|
| Draft — 0.30 MP | Fastest and lowest-memory First Image starting point |
| Balanced — 0.60 MP | More detail with greater memory and processing cost |
| Native 768 | 768 px short-edge target with a 1344 px long-edge cap |
| Custom | Custom MP selected by the user; separate from the Native 768 cap |
!Size Source First Image tooltip
Size Source = Manual ignores the preset and uses exact Width/Height values. Both must be multiples of 32. Manual is the normal choice for T2VA and also supports a 32×32 diagnostic canvas when only generated audio is being checked.
!Manual Size Source with explicit Width and Height
If First Image mode is selected but no usable First Image reaches the Sampler, V3.8 uses the stored Manual Width/Height as a safe fallback and reports the dimensions in status. Select Size Source = Manual to inspect or edit those values; Width/Height are hidden in the First Image facade. Reference Images and Video Guide never become implicit size sources. The historical Auto / Landscape / Portrait / Square values are migration/API compatibility values, not current Main choices.
Size selection is built into the V3.8 Sampler. A separate megapixel, empty-latent, or image-size node is not required for the normal Continuum workflow.
Choose a chunk duration
- 5 seconds is a valid H3 duration and the fastest practical choice for workflow checks, Review UI checks, and quick Takes.
- 8–10 seconds usually reduces the number of handoffs while keeping each Queue easier to review and rerun.
- 15 seconds is the upstream MiniMax H3 native maximum per generation and minimizes the number of boundaries, but it increases per-chunk time and memory. Longer is not automatically higher quality.
- Continuum accepts 4–30 seconds for compatibility and experimentation. Values above 15 seconds are not the same as an upstream native-duration recommendation and can be expensive at high resolution.
For an audio-only listening diagnostic, use T2VA, Size Source = Manual, and a very small canvas such as 32×32, then play the saved video and ignore its picture. This is useful for hearing music or boundary continuity; it is not evidence of normal-resolution video quality.
Advanced Sampler controls
| Control | When it appears | Meaning |
|---|---|---|
| Continuation Backend | Advanced open | Standard is the V3.8 path; Compatibility restores the older Reference Context route for comparison |
| Regenerate From | Advanced open and Progress On | Auto resumes the longest compatible prefix; Chunk N reuses earlier chunks and rebuilds N through the target |
| Variation Nonce | Explicit Regenerate From = Chunk N | 0: automatic variation selection with interrupted-run resume; 1 or higher: a fixed variation value. Base Seed and the other generation settings stay unchanged |
| Run Name (Optional Override) | Advanced open and Progress On | Stable name for resume and Render History; blank uses the Sampler's automatic identity |
| Reference Image Size | Advanced open and a Reference Image connected | Match Output is practical; Max Identity preserves more reference detail and may use more memory |
| Video Guide Size | Advanced open and Video Guide connected | Efficient — 0.4 MP, Balanced — 0.6 MP, or Match Output |
Frontend-managed IDs, selected Take IDs, one-shot review actions, diagnostics, preview, and legacy compatibility values are deliberately not editable as ordinary node widgets. The corresponding settings or action buttons below are the supported interface.
Review-state controls
When Run = Review Each Chunk, Progress is On, and a review unit is ready, the settings view changes to this action view:
| Button | What the next top-right ComfyUI Run does |
|---|---|
| Use it and continue | Accepts the current result, reuses it, and generates one next physical group |
| Try this chunk again | Keeps the earlier accepted prefix and creates another Take of the current review unit |
| Start again from Chunk 1 | Selects a fresh branch from Chunk 1 with current settings and automatic variation; press Queue afterward. Saved Takes are kept. The selection resets to Auto after Queue |
| Use it and finish the rest | Accepts the current result and generates every remaining group without further review pauses |
| Back to Settings | Shows the normal settings; does not queue, reset, or discard the review |
| Return to Review | Returns from settings to the pending review; does not undo edited values |
| Render History — N Takes | Opens or closes stored Take browsing; N is Take count, not chunk count |
| Previous Take / Next Take | Changes the selected stored Take only |
| Use This Take | Makes the selected Take canonical on the next Queue; no new sampling is required |
| Continue From Here | Branches from the selected Take and generates the next review unit |
The action button selects intent and gains a ✓; it does not start generation. Press ComfyUI's top-right blue Run button afterward. Full step-by-step examples begin in Review, continue, and revise a chunked video.
H3 Continuum Reference Audios
H3 Continuum Reference Audios bundles up to three standalone audio references behind one Audio References (Optional) Sampler socket. Connect the references without gaps and use