H3 Continuum Sampler V3.8:
H3ContinuumSamplerV38 is a production video-generation sampler for creating coherent H3 Continuum video and audio latent sequences from prompts, optional images, video references, audio, and continuity settings. It extends the V3.7 sampler while adding a shared resolution workflow, review-by-chunk generation, render-history take selection, optional bundled references, experimental timeline-based video references, and post-generation reliability diagnostics. You can use it for complete video runs or generate and review one chunk at a time before continuing. The node preserves continuity between generated sections, supports rerolling from a selected chunk, can reuse saved run information, and reports diagnostic information about frame grids, audio grids, compatibility, and the relative load of visual conditioning. Its resolution system can derive the canvas from the first image, use exact manual dimensions, or retain legacy aspect-based behavior for older workflows.
H3 Continuum Sampler V3.8 Input Parameters:
model
The diffusion or video-generation model used to create the output. It determines the learned visual behavior, quality, style capabilities, and compatibility of the sampling process. Connect a model supported by the H3 Continuum workflow. The input type is MODEL; no minimum, maximum, or default value is specified by the node interface.
clip
The CLIP text encoder used to interpret sequence_prompt and convert text into conditioning information for generation. A compatible CLIP object is required so the sampler can understand the words in your prompt. The input type is CLIP; no minimum, maximum, or default value is specified by the node interface.
video_vae
The video variational autoencoder used to convert between video data and latent representations. It affects how generated video frames are encoded and decoded and must be compatible with the selected model. The input type is VAE; no minimum, maximum, or default value is specified by the node interface.
sampler
The sampling algorithm used to progressively produce the video latent sequence from noise and conditioning. Different samplers can affect motion smoothness, detail, stability, and generation time. The input type is SAMPLER; no minimum, maximum, or default value is specified by the node interface.
sigmas
The noise schedule supplied to the sampler. It controls how the generation process moves from noisy information toward the final result, influencing detail, denoising strength, and overall sampling behavior. The input type is SIGMAS; no minimum, maximum, or default value is specified by the node interface.
sequence_prompt
The text prompt that describes the requested video sequence. Its interpretation depends on prompt_mode, allowing one prompt, a list of prompts, or time-based prompt changes. Clear descriptions of the subject, action, setting, camera behavior, and visual style generally provide more useful guidance than isolated keywords. The input type is STRING; the interface does not specify a default value.
prompt_mode
Controls how sequence_prompt is interpreted. Auto lets the node choose the appropriate interpretation, Fixed uses one prompt throughout the sequence, List interprets the text as multiple prompt entries, and Timeline allows prompt changes over time. Choose Fixed for consistent wording across a clip, List for separate prompt sections, and Timeline when different parts of the video need different descriptions. The available options are Auto, Fixed, List, and Timeline; the source does not specify a default value.
chunks
The number of physical generation groups into which the sequence is divided. More chunks can make long sequences easier to manage and review, while also increasing the number of continuity boundaries and processing stages. This value is an integer; the interface does not provide its minimum, maximum, or default value, although reroll controls expose chunk positions through Chunk 1 to Chunk 16.
chunk_seconds
The target duration of each generated chunk in seconds. It affects the amount of video produced per physical group and therefore influences total duration, memory use, and continuity handling. Larger chunks can reduce the number of boundaries but may require more resources; smaller chunks offer finer review and reroll control. The input type is FLOAT; no minimum, maximum, or default value is specified by the node interface.
aspect
The legacy aspect-ratio selection used for saved-workflow and API compatibility. Its options are Auto from First Image, Landscape 16:9, Portrait 9:16, and Square 1:1. In the V3.8 interface, the frontend migrates this value to size_source, and it is normally hidden from the main interface. It is primarily relevant when using legacy sizing behavior rather than the newer resolution controls. The available options are Auto from First Image, Landscape 16:9, Portrait 9:16, and Square 1:1; the source sets the default to Auto internally.
preset
Selects the target resolution when size_source is First Image. Draft — 0.30 MP is intended for faster, lower-resource tests, Balanced — 0.60 MP provides more detail with greater processing cost, Native 768 uses the H3 native short edge, and Custom uses the custom_mp value. The available options are Draft — 0.30 MP, Balanced — 0.60 MP, Native 768, and Custom; the source sets the default to Draft — 0.30 MP.
custom_mp
Sets a custom target size in megapixels when preset is Custom and size_source is First Image. The node preserves the connected first image's aspect ratio while selecting a canvas near this target pixel count. The source defines a default, minimum, and maximum through H3_CUSTOM_MP_DEFAULT, H3_CUSTOM_MP_MIN, and H3_CUSTOM_MP_MAX, but their numeric values are not included in the provided context. The control uses a step of 0.01.
continuity
Controls how strongly neighboring chunks are connected. Auto — conservative lets the node choose a cautious continuity strategy, Balanced — 22 frames uses a moderate continuity window, Fast — 5 frames uses a shorter window for speed, and Strong — 39 frames (Experimental) uses a larger experimental continuity window. Stronger continuity can help preserve visual relationships across chunk boundaries but may increase resource use and processing time. The available options are Auto — conservative, Balanced — 22 frames, Fast — 5 frames, and Strong — 39 frames (Experimental); no default is stated in the provided interface.
base_seed
The starting random seed for generation. The seed influences the initial noise pattern and therefore affects the visual result, motion variation, and repeatability. Reusing the same seed and settings can help reproduce a result, while changing it can produce a different take. The input type is INT; no minimum, maximum, or default value is specified by the node interface.
audio_continuity
Enables or disables continuity handling for generated audio across chunks. When enabled, the node attempts to maintain more consistent audio transitions between generated sections; disabling it may be useful when audio continuity is not required or when testing. This is a boolean control; the interface does not specify its default.
diagnostics
Selects the amount of post-generation reliability information. Basic provides normal status information, Detailed Report performs fuller checks of frame and audio grids, Core compatibility, and relative visual conditioning load, and Off disables these diagnostics. Diagnostics inspect the result without changing the generated tensors. The available options are Basic, Detailed Report, and Off; the normalization code recognizes these modes, but the provided context does not state the default.
reroll_from_chunk
Selects where a reroll should begin when regenerating part of a saved run. Auto allows the saved-run logic to determine the appropriate boundary, while Chunk 1 through Chunk 16 explicitly select a restart point. Rerolling from an earlier chunk regenerates that chunk and the dependent remainder, so it can alter more of the final sequence. The available options are Auto, Chunk 1 through Chunk 16; no default is explicitly stated in the interface, although the source method uses Auto when no value is supplied.
reroll_nonce
A numeric variation value used to request a new reroll without necessarily changing the main configuration. A nonzero value can force a new execution path when saved run storage is enabled, while a stable value of zero may allow normal caching behavior. The input type is INT; the source uses 0 as the default in cache-change handling, but no range is specified.
strict_compatibility
Controls whether compatibility checks are enforced strictly. Enabling it is appropriate when you need the workflow to reject incompatible model, Core, or related components instead of continuing with a warning or compatibility path. This is a boolean control; its default is not specified by the provided interface.
debug
Enables additional debugging behavior or information during execution. It can help investigate workflow problems, compatibility issues, or unexpected results, but may produce more verbose status information. This is a boolean control; its default is not specified by the provided interface.
show_preview
Controls whether the node displays a preview during generation. Enabling previews can make progress easier to monitor, while disabling them may reduce display overhead in automated or resource-constrained workflows. This is a boolean control; its default is not specified by the provided interface.
run_storage
Controls whether the node stores run information for reuse and review. Off disables saved-run storage, while Save + Auto Resume stores state and enables features such as review generation, rerolling, and automatic resumption. Review Each Chunk requires Save + Auto Resume. The available options are Off and Save + Auto Resume; no default is explicitly stated in the interface.
run_name
The name assigned to a stored run. A clear name helps you identify the correct generation when using automatic resume, rerolling, review actions, or render-history controls. It has type STRING; the source does not specify a default value.
reference_size
Controls the size used when processing reference images. Match Output aligns reference processing with the output canvas, while Max Identity prioritizes the maximum available identity-reference size. Larger reference processing may preserve more identifying detail but can require more memory. The available options are Match Output and Max Identity; no default is specified by the provided interface.
project_id
An identifier used to associate the run with a project. It can help separate stored runs, histories, and related workflow data when you work on multiple projects. The input type is STRING; no default value is specified.
video_reference_size
Selects the processing size for a video reference. Efficient - 0.4 MP reduces resource use, Balanced - 0.6 MP provides a middle setting, and Match Output follows the output resolution. Higher settings may retain more reference detail while increasing memory and processing requirements. The available options are Efficient - 0.4 MP, Balanced - 0.6 MP, and Match Output; the V3.8 execution path uses Efficient - 0.4 MP as its reference-video default when no other value is supplied.
continuation_backend
Selects the backend used for continuation processing. Standard uses the normal continuation path, while Compatibility is intended for workflows that require the compatibility behavior. The available options are Standard and Compatibility; no default is specified by the interface.
generation_mode
Selects between complete execution and review-based execution. Full Run performs the normal production generation, while Review Each Chunk generates at most one new physical group per queue operation and requires run_storage to be Save + Auto Resume. Review mode lets you inspect progress before advancing or changing the current chunk. The available options are Full Run and Review Each Chunk; the source sets the default to Full Run.
review_action
Determines what happens during a review-mode queue operation. Continue / Next accepts the current review and advances, Regenerate Current performs a one-shot regeneration of the current group, and Finish Remaining generates the rest of the run. The source notes that smart regeneration requires Regenerate From = Auto, although the exposed input is named reroll_from_chunk. The available options are Continue / Next, Regenerate Current, and Finish Remaining; the source sets the default to Continue / Next.
take_group
Selects a render-history take group. It is an advanced internal value normally managed through visible Previous/Next Take controls rather than direct editing. It has a default of 0, minimum 0, maximum 16, and step 1. Selecting a group changes which stored take is considered for related history actions, but selecting it alone does not change the canonical branch.
take_revision_id
Identifies the immutable render-history revision selected for a take. It is an advanced internal value managed by verified Run Storage data and the visible take controls. An empty string is the source default. The input type is STRING; no other range applies.
take_action
Defines the one-shot action associated with the selected render-history take. Automatic lets the system decide the appropriate behavior, Use This Take selects the chosen take, and Continue From Here uses it as the starting point for continuation. Selecting a take does not by itself change the canonical branch; you normally queue after selecting an action. The available options are Automatic, Use This Take, and Continue From Here; the source sets the default to Automatic.
size_source
Determines how the output width and height are resolved. First Image preserves the connected first image's aspect ratio and applies the selected resolution preset, Manual uses width and height exactly, and Legacy Aspect retains compatibility with older aspect-based workflows. Manual mode is generally appropriate for text-to-video workflows or workflows without a first image. The available options are First Image, Manual, and Legacy Aspect; the source sets the default to Legacy Aspect.
width
Sets the exact output width when size_source is Manual. The source defines a default through H3_MANUAL_WIDTH_DEFAULT, a minimum of H3_CANVAS_MIN, a maximum of H3_CANVAS_MAX, and a step of H3_CANVAS_MULTIPLE; the numeric values are not included in the provided context. Use a value divisible by 32, as stated by the node tooltip. In other sizing modes, this value may be resolved or replaced by the selected first-image or legacy sizing rules.
height
Sets the exact output height when size_source is Manual. The source defines a default through H3_MANUAL_HEIGHT_DEFAULT, a minimum of H3_CANVAS_MIN, a maximum of H3_CANVAS_MAX, and a step of H3_CANVAS_MULTIPLE; the numeric values are not included in the provided context. Use a value divisible by 32, as stated by the node tooltip. In other sizing modes, the height is resolved from the selected sizing method.
first_frame
An optional starting image for the generated video. It can establish the appearance of the opening frame and, when used with size_source set to First Image, supplies the aspect ratio used for resolution calculation. The input type is IMAGE; no default is specified.
last_frame
An optional ending image that guides the final part of the generated sequence. It can help define where the motion should finish and improve control over the beginning-to-end visual transition. The input type is IMAGE; no default is specified.
reference_image_1
An optional image reference used to guide visual identity, appearance, or other image-based conditioning. It is one of the original single reference-image sockets and is considered before bundled reference images. The input type is IMAGE; no default is specified.
reference_image_2
An optional second image reference for additional visual guidance. It can provide another view or example of the desired subject and is processed alongside the other reference inputs. The input type is IMAGE; no default is specified.
reference_image_3
An optional third image reference for additional visual guidance. It can supplement the first two references when more visual information is needed. The input type is IMAGE; no default is specified.
reference_video_1
An optional video-reference frame sequence supplied as IMAGE data, typically from a Video Adapter at 24 fps. With video_reference_mode set to Follow Timeline, successive physical output intervals use successive reference intervals; with Repeat Reference, the legacy shared-prefix guide behavior is used. The source explicitly states that reference audio is not included and that Follow Timeline is experimental rather than a hard motion-copy control. No default is specified.
driving_audio
An optional audio source that drives or accompanies the generated sequence. It can provide timing or audio content for an audio-aware generation workflow. The input type is AUDIO; no default is specified.
audio_vae
The audio variational autoencoder used when audio processing requires encoding or decoding. It must be compatible with the audio workflow and selected model components. The input type is VAE; no default is specified.
reference_audio_1
An optional single legacy audio reference. It provides audio-based conditioning or reference information where supported. The node also accepts bundled audio_references; the source specifically warns not to connect both this legacy input and the bundled input together. The input type is AUDIO; no default is specified.
reference_audio_vae
The audio VAE associated with reference_audio_1. It is used to process the legacy single reference audio input and must be compatible with that audio workflow. The input type is VAE; no default is specified.
guide
An optional Still Image Guide. The V3.7-compatible behavior prepares this guide at the resolved output width and height and injects it only into its owning physical sampling group. This provides localized visual guidance rather than applying the guide indiscriminately to every group. The input type is H3_CONTINUUM_STILL_IMAGE_GUIDE; no default is specified.
audio_references
An optional ordered bundle of additional audio references created by an H3 Continuum Reference Audios node. It is intended for multiple reference entries and should not be connected together with the legacy reference_audio_1 input. The input type is H3_CONTINUUM_AUDIO_REFERENCES; no default is specified.
image_references
An optional bundle containing Reference Images 4 through 9. Existing individual slots reference_image_1 through reference_image_3 remain first, so this bundle extends the available ordered image references rather than replacing the original sockets. The input type is H3_CONTINUUM_IMAGE_REFERENCES; no default is specified.
video_reference_mode
Controls how the optional reference_video_1 sequence is applied. Follow Timeline selects successive reference intervals for successive physical output intervals and is marked experimental; Repeat Reference keeps the legacy shared-prefix guide behavior. The available options are Follow Timeline and Repeat Reference; the provided source does not state a default.
H3 Continuum Sampler V3.8 Output Parameters:
video_latents
The generated video in latent form. This is the main visual result and is intended to be connected to compatible decoding, preview, refinement, or video-output nodes. Latents are compressed model-space data rather than ordinary image frames, so they normally need a compatible video VAE or downstream decoder before they can be viewed or exported.
audio_latents
The generated or processed audio in latent form. It carries the audio result in the representation expected by compatible audio decoding or refinement nodes. The value may be empty or contain limited information when the workflow does not generate audio or when no applicable audio conditioning is provided.
assembly_plan
A structured H3 Continuum assembly plan describing how generated physical groups and decoded chunks are arranged into the final sequence. It is useful for downstream assembly, inspection, continuation, or refinement because it preserves information about the organization of the generated sections. The output type is H3_CONTINUUM_ASSEMBLY_PLAN.
status
A text report describing the execution state and relevant reliability information. Depending on the diagnostics and run settings, it can include basic progress or completion information, detailed checks of frame and audio grids, Core compatibility, and relative visual conditioning load. The diagnostics inspect the result and do not change the generated tensors.
driving_audio
The driving-audio output corresponding to the audio source used or prepared by the node. It is provided as AUDIO data for downstream synchronization, preview, or export workflows. When no driving audio is connected or applicable, the output may contain no usable audio content.
refine_context
A structured context object for downstream refinement or continuation operations. It carries the information required by compatible refine workflows to understand the generated result and continue processing from it. The output type is H3_CONTINUUM_REFINE_CONTEXT.
H3 Continuum Sampler V3.8 Usage Tips:
-
Use
size_source = First Imagewhen you want the output to preserve the aspect ratio offirst_frame; selectDraft — 0.30 MPfor quick tests,Balanced — 0.60 MPfor a practical quality and resource balance,Native 768for the native H3 short edge, orCustomwhen you need a specific pixel budget. -
Use
size_source = Manualwhen there is no first image or when exact dimensions are important, especially for text-to-video workflows. Setwidthandheightto values divisible by 32 and within the configured canvas limits. -
Start with
continuity = Auto — conservativeorBalanced — 22 framesfor general work. UseFast — 5 frameswhen speed and lower resource use matter more than strong transitions, and useStrong — 39 frames (Experimental)only when you specifically need stronger cross-chunk continuity and can accept additional cost or experimental behavior. -
Use
generation_mode = Review Each Chunktogether withrun_storage = Save + Auto Resumewhen you need to inspect a long sequence incrementally. UseContinue / Nextto accept a chunk,Regenerate Currentto retry the current section, andFinish Remainingwhen you are satisfied with the review. -
Choose
run_storage = Save + Auto Resumebefore relying on rerolls, review mode, automatic resumption, or render-history take controls. Give each run a meaningfulrun_nameand, when appropriate, a stableproject_id. -
Use
reroll_from_chunk = Autofor normal smart rerolls. Select a specific chunk only when you intentionally want to regenerate from that point and understand that later dependent chunks may also change. -
Use
reference_image_1throughreference_image_3for the first reference images andimage_referenceswhen you need the extended ordered set of References 4 through 9. Do not provide overlapping legacy and bundled audio-reference inputs at the same time. -
Select
video_reference_mode = Repeat Referencefor the established shared-prefix behavior. TreatFollow Timelineas experimental and use it when you specifically want different video-reference intervals assigned to successive physical output intervals. -
Keep
reference_video_size = Efficient - 0.4 MPwhen memory is limited, useBalanced - 0.6 MPfor a compromise, and chooseMatch Outputwhen preserving output-scale reference detail is more important than resource use. -
Use
diagnostics = Detailed Reportwhen investigating quality, compatibility, frame-grid, or audio-grid problems. UseBasicfor ordinary production status andOffonly when you do not need the additional reliability report. -
Keep
audio_continuityenabled when the generated sequence should maintain smoother audio transitions across chunks. Disable it only when audio continuity is unnecessary or when testing a video-only workflow.
H3 Continuum Sampler V3.8 Common Errors and Solutions:
First Image sizing requires a connected first image
- Explanation: The selected
size_sourceisFirst Image, butfirst_frameis missing, so the node cannot preserve a first-image aspect ratio or resolve the requested preset size. - Solution: Connect an image to
first_frame, or changesize_sourcetoManualand provide validwidthandheightvalues.
Invalid manual canvas dimensions
- Explanation:
widthorheightis outside the allowed canvas range or does not follow the required multiple-of-32 step. - Solution: Use dimensions within the node's displayed minimum and maximum values and choose widths and heights divisible by 32.
Custom MP is only used with the Custom preset
- Explanation: The requested custom megapixel value is not being applied because
presetis not set toCustom, or the value is outside the configured custom megapixel range. - Solution: Set
presettoCustom, then choose acustom_mpvalue within the control's displayed minimum and maximum.
Review Each Chunk requires Save + Auto Resume
- Explanation: Review mode depends on stored run state so the node can remember generated groups, review progress, and continuation decisions.
- Solution: Set
run_storagetoSave + Auto Resume, provide a stablerun_name, and queue the workflow again.
No saved run available for resume or reroll
- Explanation: A review, reroll, or continuation action requested stored state, but no compatible saved run was found for the selected project or run name.
- Solution: Run the workflow once with
run_storage = Save + Auto Resume, verifyproject_idandrun_name, and then perform the review or reroll action.
Invalid reroll chunk
- Explanation:
reroll_from_chunkrefers to a chunk that is outside the configured run range or is not valid for the current number ofchunks. - Solution: Select
Autoor choose a chunk number that exists in the current run. Make sure the configured chunk count matches the saved run you are attempting to reroll.
Smart Regenerate requires Regenerate From = Auto
- Explanation: The review action requests an automatic current-chunk regeneration while
reroll_from_chunkis set to a specific chunk. - Solution: Set
reroll_from_chunktoAutobefore usingreview_action = Regenerate Current, or use an explicit reroll action appropriate for the selected chunk.
Do not connect both legacy and bundled reference audio inputs
- Explanation: The node provides both
reference_audio_1andaudio_references, but the source explicitly treats them as alternative input paths. - Solution: Disconnect one of them. Use
reference_audio_1for the legacy single reference oraudio_referencesfor the ordered bundle.
Incompatible model, CLIP, VAE, sampler, or sigma inputs
- Explanation: One or more connected model components do not match the H3 Continuum workflow or each other. This can prevent generation or produce compatibility diagnostics.
- Solution: Use model, CLIP, video VAE, sampler, and sigma nodes intended for the same H3 Continuum setup. Enable
strict_compatibilitywhen you need incompatible combinations to be rejected clearly.
Reference video timeline is unavailable or invalid
- Explanation:
Follow Timelinerequires a usable reference-video frame sequence with timing that can be mapped to physical output intervals. Missing, malformed, or incompatible frames can prevent timeline following. - Solution: Connect valid 24 fps
IMAGEframes from the Video Adapter, verify the reference duration, or switch toRepeat Referencefor the legacy shared-prefix behavior.
Output exceeds available memory
- Explanation: High resolution, long chunks, strong continuity, large reference sizes, multiple references, or detailed diagnostics can increase memory requirements beyond available hardware resources.
- Solution: Reduce the resolution preset or use a lower
custom_mp, shortenchunk_seconds, selectFast — 5 frames, useEfficient - 0.4 MPfor video references, reduce unnecessary references, or process the run in review mode.
Missing or incompatible audio VAE
- Explanation: Audio processing requires an appropriate
audio_vae, but the connected VAE is missing or incompatible with the audio workflow. - Solution: Connect the audio VAE required by the selected H3 Continuum audio setup, or remove optional audio inputs when creating a video-only result.
Invalid review take selection
- Explanation:
take_grouportake_revision_iddoes not match a verified revision in Run Storage, or a take action was selected without a valid stored take. - Solution: Use the visible Previous/Next Take controls, confirm that Run Storage is enabled, select a verified revision, and queue after choosing
Use This TakeorContinue From Here.
Generation cache did not update after changing review or reroll settings
- Explanation: Review and saved-state reroll behavior depends on information outside the ordinary input cache. A change may not produce the expected new execution when the run state or nonce is unchanged.
- Solution: Confirm that
generation_mode,run_storage, and the reroll settings are correct, use a nonzeroreroll_noncewhen a new reroll is required, and verify that the intended saved run is selected.
Frame or audio grid diagnostics report a mismatch
- Explanation: The generated frame or audio structure does not align with the expected timing, chunk, or assembly grid. This is reported by reliability diagnostics and does not itself alter the tensors.
- Solution: Review the
statusoutput, checkchunks,chunk_seconds, audio settings, and reference timing, then retry with simpler chunking or with compatible model and VAE components.
