H3 Continuum Sampler V3.9:
H3ContinuumSamplerV39 generates a longer MiniMax H3 video as a sequence of chunks while providing controls for continuity, resolution, audio, saved progress, and review between chunks. Its V3.9 change is a simpler way to guide appearance across that sequence: one companion node holds up to nine fixed reference images and assigns each image to all chunks or selected chunks. This lets you keep a character, object, or visual setting available where it matters without reconnecting separate image sockets for every reference. The assignments guide generation; they do not guarantee an exact visual match or replace the sequence’s opening image. (github.com)
H3 Continuum Sampler V3.9 Input Parameters:
The interface supplied for this node lists types and options, but does not expose a default or numeric range for every inherited control. Where those values are not established by the supplied source or cited project guide, they are marked as unspecified rather than inferred.
model
Required MODEL. Connect a compatible MiniMax H3 generation model, including any intended upstream model adjustments. This is the model that creates the sequence; an incompatible model can prevent sampling. No default or range is specified. (github.com)
clip
Required CLIP. Connect the compatible text encoder so the node can interpret your sequence prompt. No default or range is specified. (github.com)
video_vae
Required VAE. Used to encode connected visual conditioning such as images; it is still a required connection even when a pure text-to-video setup does not use image conditioning. It is not the final video decoder. No default or range is specified. (github.com)
sampler
Required SAMPLER. Supplies the sampling algorithm used during generation. Choose one supported by your H3 workflow; the supplied source does not establish a default or enumerate algorithms. (github.com)
sigmas
Required SIGMAS. Supplies the noise schedule and effective sampling-step count. This influences how generation proceeds, so keep it consistent when comparing otherwise identical runs. No default or range is specified. (github.com)
sequence_prompt
Required STRING. Write the scene or sequence instructions here; prompt_mode determines how the text is interpreted. For routed image references, you can refer to fixed slots with tags such as @R1, which the routing system maps to active references for the relevant chunk. No default or length limit is specified. (github.com)
prompt_mode
Required choice: Auto, Fixed, List, or Timeline. Select how sequence_prompt should be parsed; Auto is a useful starting choice when you do not need to force a format. The interface does not establish a default. (github.com)
chunks
Required INT. Sets the planned total number of chunks, not the number of additional chunks to generate on the next queue. The project guide states a range of 1–16; the V3.9 routing call uses 3 when this value is absent from a direct runtime call, but that does not establish the widget’s default. More chunks allow a longer sequence and create more boundaries to manage. (github.com)
chunk_seconds
Required FLOAT. Sets the duration of each chunk; planned duration is chunks multiplied by chunk_seconds. The project guide says Continuum accepts 4–30 seconds for compatibility, while 5–15 seconds is the validated practical range. Longer chunks can take more time and memory; no widget default is specified. (github.com)
aspect
Required choice: Auto from First Image, Landscape 16:9, Portrait 9:16, or Square 1:1. This is the legacy aspect control retained for saved-workflow and API compatibility; use size_source for the current sizing workflow. The supplied V3.8 implementation sets its default to Auto from First Image. (github.com)
preset
Required choice: Draft — 0.30 MP, Balanced — 0.60 MP, Native 768, or Custom; default: Draft — 0.30 MP. Used when size_source is First Image: Draft favors speed and lower memory, Balanced targets more detail, Native 768 uses a 768-pixel short-edge target, and Custom uses custom_mp. It does not set dimensions in Manual mode. (github.com)
custom_mp
Required FLOAT. Sets the target megapixel area for a connected First Image while preserving its aspect ratio; it matters only when preset is Custom and First Image sizing is in use. The supplied source refers to default, minimum, and maximum constants but does not provide their numeric values. (github.com)
continuity
Required choice: Auto — conservative, Balanced — 22 frames, Fast — 5 frames, or Strong — 39 frames (Experimental). Controls how much previous video context is carried across chunk boundaries. Balanced is the project guide’s starting point for standard audiovisual continuation; Strong is explicitly experimental. The interface does not establish a default. (github.com)
base_seed
Required INT. Provides the base value used to derive per-chunk seeds. Keep it unchanged when resuming a run or comparing takes so a seed change does not confound the comparison. No default or range is specified. (github.com)
audio_continuity
Required BOOLEAN. When enabled, previously generated audio provides context for the next chunk; it does not import an external audio track. The project guide recommends enabling it for normal audiovisual generation. The supplied interface does not establish its default. (github.com)
diagnostics
Required choice: Basic, Detailed Report, or Off. Controls the amount of diagnostic reporting; the V3.8 source uses Basic when a diagnostics value is absent from its runtime arguments. Detailed reporting can help investigate frame, audio, compatibility, or conditioning issues; diagnostics inspect results rather than alter generated tensors. (github.com)
reroll_from_chunk
Required choice: Auto or Chunk 1 through Chunk 16. Auto lets saved progress determine a compatible resume point; an explicit chunk requests regeneration starting there rather than from the beginning. The supplied V3.8 cache-key method defaults to Auto, but the interface does not separately show widget metadata. Use this with saved progress and verify the chosen boundary before queuing. (github.com)
reroll_nonce
Required INT. Controls variation for an explicit regeneration boundary: the project guide describes 0 as automatic variation selection with interrupted-run resume, and 1 or higher as a fixed variation value. The supplied V3.8 cache-key method defaults to 0; a widget minimum and maximum are not shown. (github.com)
strict_compatibility
Required BOOLEAN. Exposes an inherited compatibility control. Its precise validation rules, execution effect, and default are not present in the supplied implementation, so do not treat it as a general quality switch without checking the version-specific workflow.
debug
Required BOOLEAN. Exposes an inherited debug control. The supplied source does not specify which additional information it emits or its default; use it for investigation only after confirming the behavior in your installed version.
show_preview
Required BOOLEAN. Exposes an inherited preview control. The supplied source does not establish its exact preview behavior, performance cost, or default.
run_storage
Required choice: Off or Save + Auto Resume. Saved storage supports compatible progress reuse and is required by the supplied V3.8 tooltip for Review Each Chunk; Off does not provide that review storage path. No widget default is established by the supplied interface. (github.com)
run_name
Required STRING. An optional override in practice: a stable name helps you identify saved runs and take history, while a blank value uses an automatic run identity according to the project guide. No widget default or length limit is specified. (github.com)
reference_size
Required choice: Match Output or Max Identity. Controls the sizing choice for image-reference conditioning, not the final canvas dimensions or the size of your source files. The supplied source does not quantify the resulting size or establish a default. (github.com)
project_id
Required STRING. Exposes an inherited project identifier. Its exact effect on storage or generation and its default are not established by the supplied source; keep a value consistent when continuing a project rather than assuming it is merely a display label.
video_reference_size
Required choice: Efficient - 0.4 MP, Balanced - 0.6 MP, or Match Output. Selects a size policy for video-reference conditioning; the listed megapixel targets distinguish the first two choices, while Match Output follows output sizing. The supplied runtime facade names Efficient as its fallback when a value is absent, but does not establish a widget default. (github.com)
continuation_backend
Required choice: Standard or Compatibility. Standard is the current continuation route; Compatibility restores an older reference-context route for comparison. This can change how continuity conditioning is applied. The interface does not establish a default. (github.com)
generation_mode
Required choice: Full Run or Review Each Chunk; default: Full Run. Full Run follows normal production execution. Review Each Chunk generates at most one new physical group per queue so you can inspect progress, and requires run_storage set to Save + Auto Resume. (github.com)
review_action
Required choice: Continue / Next, Regenerate Current, or Finish Remaining; default: Continue / Next. In review mode, Continue advances after accepting the current review, Regenerate Current requests another attempt, and Finish Remaining requests completion. The latter two are one-shot actions. (github.com)
take_group
Required INT; default: 0, minimum: 0, maximum: 16, step: 1. This is an internal Render History selection for a physical group. Use the visible Previous/Next Take controls instead of editing the number directly, so the selection stays aligned with verified saved history. (github.com)
take_revision_id
Required STRING; default: an empty string. Holds the immutable revision selected by Render History. The visible Take controls manage this value from verified Run Storage data; typing an arbitrary ID does not create a take. No length limit is specified. (github.com)
take_action
Required choice: Automatic, Use This Take, or Continue From Here; default: Automatic. Applies a one-shot Render History decision. Selecting a take alone does not change the canonical branch; choose the intended action and queue normally. (github.com)
size_source
Required choice: First Image, Manual, or Legacy Aspect; default: Legacy Aspect for compatibility. First Image preserves the connected opening image’s aspect ratio using preset; Manual uses exact width and height. Legacy Aspect is retained for saved-workflow and API compatibility rather than normal new setup. (github.com)
width
Required INT. In Manual sizing, sets the exact output width and must be a multiple of 32; the project guide states an accepted range of 32–16,384. It is a normal choice for text-to-video or when no First Image supplies the desired aspect. The supplied source names a default constant without exposing its numeric value. (github.com)
height
Required INT. In Manual sizing, sets the exact output height and must be a multiple of 32; the project guide states an accepted range of 32–16,384. It works with width to determine the canvas shape. The supplied source names a default constant without exposing its numeric value. (github.com)
first_frame
Optional IMAGE. Supplies the opening image for image-to-video generation and is the image whose aspect ratio can drive size_source = First Image. A reference image does not substitute for it as an implicit size source. No default or image-size limit is specified. (github.com)
last_frame
Optional IMAGE. Supplies a desired ending-image constraint for a first/last-image workflow. Leave it disconnected for ordinary text-to-video or first-image-only generation; it can affect how the terminal part of a sequence is rebuilt during review or regeneration. No default or image-size limit is specified. (github.com)
reference_video_1
Optional IMAGE. Accepts video frames, displayed as Timeline Video Frames in the supplied V3.8 schema. The tooltip specifies 24 fps frames from a Video Adapter and notes that source audio is not included; connect audio separately when needed. Its behavior also depends on video_reference_mode. No default or frame-count limit is specified. (github.com)
driving_audio
Optional AUDIO. Supplies source audio to guide the sequence and become its selected final audio, rather than merely influencing newly generated audio. Connect it when the finished work should follow an existing recording or soundtrack. No default or duration limit is specified. (github.com)
audio_vae
Optional VAE. Supplies audio encoding for the driving-audio path. Use a compatible audio VAE when that path needs it; the supplied source does not establish a default or all connection-validation conditions. (github.com)
reference_audio_1
Optional AUDIO. Supplies the legacy single audio reference for conditioning: it can guide generated audio but does not replace the final audio the way driving_audio does. Do not use it together with the bundled audio_references path. No default or duration limit is specified. (github.com)
reference_audio_vae
Optional VAE. Supplies audio encoding for the legacy reference-audio path. It is distinct from audio_vae, which accompanies driving audio; no default is specified. (github.com)
guide
Optional H3_CONTINUUM_STILL_IMAGE_GUIDE. Accepts a prepared still-image guide. The project guide describes this inherited socket as experimental compatibility functionality; the supplied V3.9 code does not define its placement or strength controls. No default is specified. (github.com)
audio_references
Optional H3_CONTINUUM_AUDIO_REFERENCES. Accepts an ordered bundle from H3 Continuum Reference Audios for audio conditioning. It is separate from driving audio and should not be connected together with the legacy single reference_audio_1 input. No default or bundle-size setting is specified in this node’s interface. (github.com)
video_reference_mode
Optional choice: Follow Timeline or Repeat Reference. Follow Timeline selects successive physical output intervals from the supplied video frames; Repeat Reference keeps the older shared-prefix guide. The supplied tooltip calls Follow Timeline experimental and warns that it is not a hard motion-copy control. The project guide identifies Repeat Reference as the behavior for older requests that omit this setting; the supplied V3.9 interface does not separately show widget default metadata. (github.com)
reference_images
Optional H3_CONTINUUM_REFERENCE_IMAGES_V39. Connect the output of H3 Continuum Reference Images V3.9 here, then assign its fixed image slots 1–9 to all or selected chunks in that companion node. This is V3.9’s single public image-reference connection; older direct reference-image sockets are not part of this interface. No default is specified. (github.com)
H3 Continuum Sampler V3.9 Output Parameters:
video_latents
LATENT, output 1. Contains generated video in a latent representation rather than viewable frames. Pass it through the compatible video decoding path before inspecting pictures or finalizing the sequence; the project guide describes the raw output as a list of physical groups. (github.com)
audio_latents
LATENT, output 2. Contains generated audio in a latent representation rather than playable sound. Decode it with the matching audio path and keep it paired with the same run’s assembly plan so the pieces can be placed on the intended timeline. (github.com)
assembly_plan
H3_CONTINUUM_ASSEMBLY_PLAN, output 3. Carries the instructions needed to assemble decoded physical groups into the requested sequence. The V3.9 code also checks this plan for a verified reference-routing record before returning a normal result; use the plan from the same sampling run as the latents, not one from an earlier attempt. (github.com)
status
STRING, output 4. Provides human-readable run and diagnostic information. Check it for sizing fallbacks, progress or review information, and reported compatibility concerns before deciding whether to resume, regenerate, or finalize. Its exact wording depends on the run; the V3.9 code may also update a Second Pass warning when output-scope conditioning has been verified. (github.com)
driving_audio
AUDIO, output 5. Returns the selected source-audio path when driving audio is used. It is an audio value for downstream handling, not the generated audio latents in output 2; connect it where your assembly workflow expects the source track. (github.com)
refine_context
H3_CONTINUUM_REFINE_CONTEXT, output 6. Provides context for the advanced Second Pass refinement path. V3.9 can project this context against the verified assembly plan, so a downstream refinement workflow has the appropriate output scope; it is not itself decoded video or audio. (github.com)
H3 Continuum Sampler V3.9 Usage Tips:
- For a quick workflow check, start with
preset = Draft — 0.30 MPwhen using a First Image and a short, validatedchunk_secondsvalue such as 5. Increase resolution or duration after the connections and routing behave as expected. (github.com) - For text-to-video or exact delivery dimensions, set
size_source = Manualand enterwidthandheightin multiples of 32. For image-to-video, connectfirst_frameand chooseFirst Imageto preserve its aspect ratio; reference images will not determine that ratio. (github.com) - Put recurring character or object images in H3 Continuum Reference Images V3.9 and connect its single bundle to
reference_images. Assign only the chunks that need each image, then inspect the displayed reference plan to check where the references were routed. (github.com) - Set
run_storage = Save + Auto Resumebefore usinggeneration_mode = Review Each Chunk. Keepbase_seedand other generation settings stable while evaluating takes, then use the visible Take controls and an explicittake_actionwhen you intend to change the accepted branch. (github.com) - Use
driving_audiowhen an existing recording should become the final track; use a reference-audio path when audio should influence generation without replacing the generated track. Choose eitherreference_audio_1oraudio_references, not both. (github.com)
H3 Continuum Sampler V3.9 Common Errors and Solutions:
The first message below is an exact error string from the supplied V3.9 class. The other headings describe recognizable problems or likely validation failures; their wording is not asserted to be a verbatim error message.
V3.9 Reference Routing has no verified output Plan
- Explanation: V3.9 expected a verified reference-routing record in the returned assembly plan but did not find one.
- Solution: Check that
reference_imagescomes from H3 Continuum Reference Images V3.9, then retry with matching V3.9 components. If it persists, retain the complete error and run status for troubleshooting; changing prompts alone will not repair a missing plan record.
Required model or sampler connection is missing
- Explanation: Generation needs the required model, text encoder, video VAE, sampler, and sigma connections.
- Solution: Inspect
model,clip,video_vae,sampler, andsigmasfor disconnected or incompatible upstream nodes before queuing. (github.com)
Reference Images connection has the wrong type
- Explanation: The V3.9
reference_imagessocket expects the V3.9 companion node’s typed bundle, not an individualIMAGEcable or an older reference-image bundle. - Solution: Connect H3 Continuum Reference Images V3.9 to
reference_imagesand place your images in that companion node. Rewire older workflows that used direct reference-image sockets. (github.com)
Review Each Chunk cannot proceed with storage Off
- Explanation: The supplied mode tooltip says review requires
Save + Auto Resume; review decisions depend on saved state. - Solution: Set
run_storage = Save + Auto Resume, keep the run identity consistent, and queue again. (github.com)
First Image sizing does not match the expected image
- Explanation:
size_source = First Imageusesfirst_frame, not a reference image or timeline-video frame, to establish the aspect ratio. Without a usable First Image, the documented fallback uses stored Manual dimensions and reports them instatus. - Solution: Connect the intended image to
first_frame, or selectManualand inspectwidthandheight; readstatusto confirm the dimensions actually used. (github.com)
Manual width or height is rejected
- Explanation: Manual dimensions must follow the node’s canvas limits and 32-pixel increments.
- Solution: Set each dimension to a multiple of 32 within the documented 32–16,384 accepted range, then retry. (github.com)
Requested chunk cannot be regenerated
- Explanation: An explicit
reroll_from_chunkmay not match the configured chunk count or compatible saved progress. - Solution: Check the planned
chunkstotal, choose an existing boundary, and confirmrun_storageand the saved run identity before regenerating; useAutowhen you want the compatible resume point selected for you. (github.com)
Audio reference paths conflict
- Explanation: The supplied schema warns against connecting the legacy single reference-audio input and bundled audio references together.
- Solution: Disconnect one path. Keep
reference_audio_1for a single legacy reference, or useaudio_referencesfor the bundled path; usedriving_audioseparately only when source audio should guide and become the final track. (github.com)
Timeline video does not include its source audio
- Explanation:
reference_video_1receives IMAGE frames; its supplied tooltip explicitly says source audio is not included. - Solution: Connect the source audio to
driving_audioas well when you intend to use that recording, and check the selected audio in the downstream assembly workflow. (github.com)
Output fails during decoding or final assembly
- Explanation:
video_latentsandaudio_latentsare not final media, and decoded groups must be assembled using their matchingassembly_plan. - Solution: Use compatible video and audio decode paths, then pass their decoded results together with the plan from the same run to Finalize. Do not combine outputs from different attempts or takes. (github.com)
