H3 Continuum Second Pass:
This node performs a context-aware second refinement pass for MiniMax H3/Continuum video generation. It applies a low-denoise refinement schedule to complete physical video decode groups, helping improve the visual quality of externally upscaled video latents while preserving the structure established during the first pass. When refinement context is supplied, the node can maintain the physical group’s First, Last, and Reference conditioning so that transitions and related frames remain consistent. The node is designed to refine video while preserving the first-pass audio exactly, making it suitable for workflows where the video has been upscaled or otherwise prepared for a final quality pass. The refinement behavior is controlled by the supplied sigma schedule and a seed that is independently derived for each physical group.
H3 Continuum Second Pass Input Parameters:
model
The diffusion model used to perform the second-pass video refinement. It determines how the supplied video latents are denoised and reconstructed, so it must be compatible with the H3/Continuum latent format and the rest of the connected workflow. The model affects the visual characteristics, stability, and quality of the refined result. This input is required, accepts one MODEL value, and has no default, minimum, maximum, or selectable options specified by the node.
clip
The CLIP conditioning component used by the refinement process to interpret text or other conditioning information associated with the model workflow. It must be compatible with the selected model and the conditioning setup used to create the original video. A mismatched CLIP component can lead to incorrect conditioning or workflow errors. This input is required, accepts one CLIP value, and has no default, minimum, maximum, or selectable options specified by the node.
sampler
The sampling method used to carry out the low-denoise second-pass refinement. Different compatible samplers can affect motion stability, texture detail, smoothness, and the way the refinement follows the existing latent content. Use a sampler intended for the connected H3/Continuum model and the supplied sigma schedule. This input is required, accepts one SAMPLER value, and has no default, minimum, maximum, or selectable options specified by the node.
sigmas
The sigma schedule controlling the refinement denoising levels. Sigma values determine how strongly the node modifies the existing video latents during the second pass: a low-denoise schedule generally preserves more of the original result while adding detail or correcting visual issues, whereas a more aggressive schedule can change the image more substantially. The schedule must be compatible with the sampler and model. This input is required, accepts one SIGMAS value, and has no default, minimum, maximum, or selectable options specified by the node.
video_latents
The video latent data to refine, normally consisting of externally upscaled H3 video latents produced by an earlier stage of the workflow. The node processes these latents by physical decode group, so the data should correspond to the supplied assembly plan and retain the expected H3/Continuum structure. The quality, dimensions, frame organization, and latent format of this input directly affect the refined video output. This input is required, accepts one or more LATENT values as provided by the workflow, and has no default, minimum, maximum, or selectable options specified by the node.
audio_latents
The matching first-pass audio latent data associated with the video latents. The node carries this audio through the second pass and preserves the audio output bit-exactly from the first pass rather than regenerating or modifying it. The audio data should match the same assembly plan and physical groups as the video input so that audio and video remain synchronized. This input is required, accepts one or more LATENT values as provided by the workflow, and has no default, minimum, maximum, or selectable options specified by the node.
assembly_plan
The original H3/Continuum assembly plan describing how the video and audio material is organized into physical decode groups and how those groups should be assembled. This plan is essential for processing the complete groups consistently and for returning an updated plan that corresponds to the refined video. It should be the matching plan used for the supplied first-pass latents, rather than a newly created or unrelated plan. This input is required, accepts one H3_CONTINUUM_ASSEMBLY_PLAN value, and has no default, minimum, maximum, or selectable options specified by the node.
refine_seed
The base random seed used for second-pass refinement. The node derives refinement behavior independently once per physical group, allowing each group to receive deterministic processing based on the same base seed while avoiding unintended reuse of one identical random operation across all groups. Changing this value can change the refined visual details while leaving the general workflow configuration unchanged. The default value is 0, the minimum is 0, and the maximum is 18446744073709551615 (0xFFFFFFFFFFFFFFFF). The input is required and accepts an integer value.
refine_context
Optional context information for preserving the physical group’s First, Last, and Reference conditioning during refinement. Supplying the matching Continuum refine context can help the second pass maintain relationships between boundary frames, reference frames, and the rest of each physical group. Leave it unconnected only when the workflow is designed to perform refinement without this additional conditioning context. This optional input accepts one H3_CONTINUUM_REFINE_CONTEXT value and has no default, minimum, maximum, or selectable options specified by the node.
video_vae
An optional VAE used when the refinement workflow requires video latent decoding or related video-specific latent processing. It must be compatible with the H3/Continuum video latents and model workflow. When connected, it provides the node with the video decoding component needed by the underlying refinement implementation; when not connected, the node uses the behavior available without an explicitly supplied video VAE. This optional input accepts one VAE value and has no default, minimum, maximum, or selectable options specified by the node.
H3 Continuum Second Pass Output Parameters:
refined_video_latents
The video latents after the low-denoise second-pass refinement. These latents contain the visually refined version of the supplied video and can be passed to later decoding, assembly, or export nodes. The result is organized according to the physical decode groups described by the assembly plan, allowing the downstream workflow to preserve the intended group structure. This output is returned as one or more LATENT values.
audio_latents
The audio latents carried forward from the first pass. This output is preserved bit-exactly, meaning the node does not intentionally alter the audio data during video refinement. Use it with the refined video latents and the updated assembly plan to maintain the original audio content and synchronization. This output is returned as one or more LATENT values.
updated_assembly_plan
The assembly plan updated to correspond to the refined video result. It provides downstream nodes with the plan needed to interpret and assemble the refined physical groups correctly. Because the plan is produced alongside the refined latents, it should generally be passed forward together with refined_video_latents and the preserved audio_latents. This output is one H3_CONTINUUM_ASSEMBLY_PLAN value.
status
A text status message describing the outcome or state of the second-pass operation. You can use it to monitor whether processing completed and to inspect information reported by the refinement workflow. This output is a STRING value intended for workflow displays, logging, or troubleshooting.
H3 Continuum Second Pass Usage Tips:
- Use the video latents produced or prepared for the matching H3/Continuum assembly plan. The node expects the latent data and plan to describe the same physical groups.
- Connect the original first-pass audio latents whenever you want the refined video to retain the exact first-pass audio. The node is intended to refine video without regenerating that audio.
- Use the same compatible model, CLIP component, sampler family, and latent format that belong to the H3/Continuum workflow. Mixing components from unrelated model families can cause connection errors or produce unreliable results.
- Start with the low-denoise sigma schedule intended for second-pass refinement. This generally preserves the composition and motion established in the first pass while allowing the node to improve visual detail.
- Connect the matching
refine_contextwhen continuity at First, Last, and Reference frames matters, especially for multi-group videos, transitions, or workflows that use reference conditioning. - Keep
video_vaeconsistent with the video model and latent format when the workflow requires explicit video decoding support. Do not substitute an unrelated image-only VAE unless the surrounding workflow specifically confirms compatibility. - Change
refine_seedwhen you need a different refinement variation while keeping the model, sampler, sigma schedule, and source latents unchanged. Keep the seed fixed when comparing settings or reproducing a preferred result. - Pass
updated_assembly_plandownstream together with the refined video and preserved audio. Treating the plan as optional metadata can cause later assembly stages to interpret the physical groups incorrectly. - For difficult or memory-intensive videos, refine complete physical groups with matching dimensions rather than combining fragments from different plans. The node’s stated operation is based on complete physical decode groups.
H3 Continuum Second Pass Common Errors and Solutions:
model must contain exactly one value
- Explanation: The node receives zero model values or more than one model value. Its list-based input handling requires exactly one model object.
- Solution: Connect one compatible model output to
modeland remove duplicate links or list items that provide additional model values.
clip must contain exactly one value
- Explanation: The node did not receive exactly one CLIP value. This usually occurs when the input is disconnected or when multiple CLIP values are passed through a list.
- Solution: Connect one CLIP component compatible with the selected H3/Continuum model.
sampler must contain exactly one value
- Explanation: The sampler input contains zero or multiple sampler values, so the node cannot select one sampling method for refinement.
- Solution: Connect exactly one sampler and remove any extra sampler entries from the workflow.
sigmas must contain exactly one value
- Explanation: The node did not receive exactly one sigma schedule. Without one schedule, it cannot determine the denoising levels for the second pass.
- Solution: Connect one valid
SIGMASoutput generated for the selected sampler and refinement workflow.
assembly_plan must contain exactly one value
- Explanation: The node requires one assembly plan but received none or received multiple plans. Multiple plans can make the relationship between video groups, audio groups, and conditioning context ambiguous.
- Solution: Connect the original matching assembly plan as a single value and remove unrelated plan outputs.
refine_seed must contain exactly one value
- Explanation: The seed input does not contain exactly one integer value. The node cannot derive deterministic per-group refinement behavior without a single base seed.
- Solution: Enter one integer in the range
0through18446744073709551615, or connect one integer seed output.
refine_seed outside the allowed range
- Explanation: The supplied seed is lower than
0or higher than18446744073709551615, which exceeds the range declared by the node. - Solution: Use an integer within the allowed range. The default value
0is valid.
video_latents and audio_latents do not match the assembly plan
- Explanation: The supplied video or audio latent groups do not correspond to the physical groups described by the assembly plan. This can prevent correct processing or cause synchronization and assembly problems.
- Solution: Use the video latents, audio latents, and assembly plan produced by the same first-pass workflow and keep their group ordering consistent.
refine_context is incompatible with the assembly plan
- Explanation: The optional refinement context does not describe the same physical groups or First/Last/Reference conditioning as the supplied assembly plan.
- Solution: Connect the refine context generated for the same source workflow and group structure. Disconnect it only when a context-free refinement workflow is intentionally being used.
video_vae is incompatible with the video latents
- Explanation: The supplied VAE may not support the latent format or video model used by the workflow, which can lead to decoding or refinement failures.
- Solution: Use the video VAE associated with the H3/Continuum model and latent format, or leave the optional input disconnected when the workflow does not require an explicit VAE.
SIGMAS schedule is incompatible with the sampler or model
- Explanation: The sigma schedule may have been created for a different sampler or model family, or may not represent a valid low-denoise refinement schedule.
- Solution: Generate the sigma schedule with a compatible scheduler and connect it to the sampler and model used by this node.
