Getting Started: Chunks, Prompts & Reference Images
3.9.1 — what changed (2026-10-04)
This patch fixes Second Pass conditioning, Take reuse and Review audio:
Second Pass: preserve audio-only/mixed keyframes and inherit the verified conditioning and Reference assignments of the actual Review output, including a single middle chunk.
Take reuse: include sampler closures and MODEL CFG/wrappers in compatibility checks to prevent reuse under different generation settings.
Review Driving Audio: cut the source PCM once at the physical group's natural time, with the same start position for Exact ON/OFF.
Audio resampling: prefer ComfyUI Core's standard API, with the legacy fallback for older Core versions. The Reference Encode Cache limitation for direct VAE weight changes is documented below.
V3.9 Performance Fix — October 1, 2026
Reduced unnecessary preparation time when using Fixed prompts without Reference Images. First Image is still supported. In matched tests, the earlier slowdown compared with V3.8 was no longer observed under these conditions.
Update Continuum from GitHub main, then fully restart ComfyUI. No workflow changes are required for this fix.
What’s New in V3.9
V3.9 introduces per-chunk Reference Images.
Reference Images 1–9 now keep fixed IDs @R1@R9), and the new Reference Images V3.9 node lets you choose which images are active for each chunk. All references are connected to the V3.9 Sampler through one Reference Images input.
For the supplied 2 × 10 s workflow, put each time header on its own line and describe only that chunk's action. For example, after enabling the matching image loaders and checking their chunk assignments:
[0-10s]
@R1 walks forward through a quiet hallway. The camera follows smoothly.
[10-20s]
@R2 enters from the right. The camera turns to follow @R2.
### V3.8X2 Compatibility
V3.8X2 is still supported as-is.
You can drag & drop your existing V3.8X2 workflow and continue using it normally with the V3.8 Sampler.
The new V3.9 Reference system uses different wiring and is not automatically applied to old workflows. To use V3.9 features, either load the included V3.9 workflow or build/rewire your own workflow with the V3.9 Sampler and Reference Images V3.9 node.
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/tree/main#v39-reference-images
What’s New in V3.8X2
V3.8X2 is a small update to V3.8X. The existing sampler, Review, Resume, Retry, Takes, and continuation behavior are unchanged.
Up to 9 Reference Images
Reference Images 4–9 can now be added through the new Reference Images helper. Using many or large reference images can use substantial VRAM and system RAM, so resizing source images first is recommended.
Built-in Decode Cache Helper
A Decode Cache Helper node is included for repeated decoding of unchanged latents. It can help with reuse and resume cases, but it does not speed up new Sampling and may not be useful for normal fresh generations.
Use the included V3.8X2 workflow.
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
What’s New in V3.8X
V3.8X is the new main line of H3 Continuum, focused on more reliable Review, Resume, Retry, and Continue workflows.
Current-chunk Review preview
Validated prefix reuse
Atomic crash-safe chunk saving
Reliable Resume, Retry, and Takes
No unnecessary regeneration of accepted chunks
Same 7 public nodes and sampler interface
Use the included V3.8X workflow.
V3.8.0 remains available for existing setups.
For a new Git installation:
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.gitFor an existing Git checkout:
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin mainRestart ComfyUI after cloning or pulling. If ComfyUI Manager installed the node, use its Update action instead of
What’s New in V3.8
Please refer to the GitHub repository for usage instructions.
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
V3.8 focuses on workflow control, recovery, and ease of use. It does not introduce a new image-quality enhancement; equivalent settings continue to use the established V3.7 Production sampling foundation.
Review Each Chunk — inspect each generated chunk before continuing.
Keep, Retry, or Finish — use the current chunk, generate it again, or finish all remaining chunks.
Resume & Takes — reopen saved progress and choose from previous Takes.
Partial Regeneration — regenerate from a selected chunk without rebuilding the accepted beginning.
Simplified Size Controls — choose First Image sizing or exact Manual dimensions.
Built-in Input Bypass — quickly enable or disable Image, Audio, and Video inputs.
Three Reference Audios — use up to three ordered audio references.
Simplified Public Surface — seven supported V3.8 Continuum nodes.
Spectrum / Turbo Workflow — the supplied Spectrum graph can also be configured for LightX2V Turbo.
Known limitation: Issue #13 remains open. V3.8 does not claim to improve or solve cumulative image-quality drift during long continuation.
What’s New in V3.7
V3.7 focuses on faster, more reliable high-resolution Second Pass refinement. The new Conditioning Adapter rebuilds First/Last images and continuation context at the target resolution, while RefineSchedule provides exact Full, Tail, Partial, and External refinement ranges.
Using Tail 6 instead of Tail 10 reduced Second Pass sampling from 100.27s to 64.18s in our reference test—about 36% faster—while preserving the original first-pass audio bit-exact.
Production defaults remain unchanged, and existing V3.6 workflows remain compatible. The new Still Image Guide is included as an Experimental feature because hard anchors may cause abrupt trajectory changes.
What’s New in V3.6.1
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
V3.6 introduces a new Masked AV Continuation backend for faster and cleaner long-form generation.
Instead of adding the previous chunk again as a separate Reference block, V3.6 keeps the finalized Video/Audio latent prefix directly inside the next target and samples only the new region.
Masked AV Continuation — new Standard backend
Less Attention overhead — FL2VA test reduced packed rows by 8.01%
Faster continuation Sampling — median Terminal Group sampling improved by 4.06%
Bit-exact Video + Audio prefix preservation
T2VA / I2VA / FL2VA supported
Reference Image + Reference Audio supported
Long Terminal Merge fully supported
Run Storage / Resume / Regenerate From supported
Compatibility mode keeps the previous V3.5 Reference Context backend
Chunk duration expanded to 4–30 seconds — 5–15 seconds remains recommended
V3.6.1 Hotfix
Improved mixed Timeline /
---prompt syntax handling and warningsPrevents unintended repeated chunk prompts
Non-Balanced Audio Continuity settings safely fall back to the compatible Reference Context path instead of stopping generation
Same sampling quality, less redundant continuation work, and full compatibility with existing V3.5 workflows.
H3 Continuum V3.5.2 — Stabilization & Optimization Update
V3.5.2 focuses on stability, efficiency, and cleanup rather than adding major new features. Repeated runs with the same Prompt/CLIP conditions now avoid unnecessary re-encoding, and Video Guide preprocessing uses substantially less temporary RAM on longer inputs.
Existing V3.5.x workflows and generation contracts remain unchanged. Sampling, Terminal Merge, Run Storage, Reference handling, and output behavior were preserved while the updated paths were validated with CPU and GPU A/B testing.
V3.5.2 is an internal stabilization and optimization update. Existing V3.5.1 workflows can be used as-is, with no changes required to nodes, connections, or workflow structure.
> v3.5.1 Update
This update improves the Second Pass workflow, reference handling, and seed behavior. It adds the new Conditioning Bridge V3.5 for external sampling workflows, optional Reference Audio conditioning.
Video reference inputs are also clarified as Video Guide Frames / Video Guide Size, while existing V3.4/V3.5 workflows and backend connections remain compatible.
Updating to V3.5 does not automatically replace V3.4 nodes in saved workflows. When updating an older workflow, replace not only the Sampler but also the assembler with H3 Continuum Assemble + Seam V3.5. Add H3 Continuum Hi-Res Fix V3.5 when Hi-Res Fix is required.
GitHub:
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
[ V3.5 ]
V3.5 introduces two major additions:
- Continuum-aware Second Pass / Hi-Res Fix
Refine externally processed H3 latents while preserving Continuum physical groups, prompts, seeds, ordering, and first-pass audio. An integrated one-node 2x Hi-Res Fix path is also included as an experimental feature.
- Low-memory Assemble + Seam V3.5
Adds Auto, RAM, and Disk-backed video-buffer modes. Disk-backed assembly significantly reduces system RAM/private-memory usage for long or high-resolution outputs while preserving Exact Duration, Seam, Terminal Merge, and audio behavior.
All V3.4 nodes remain available for saved-workflow compatibility. Existing V3.4 workflows continue to work unchanged.
The V3.5 release passed 430 automated tests and representative GPU acceptance tests.
Note: The integrated Hi-Res Fix remains experimental. Long 2x workflows can require substantial GPU VRAM.
#5 3 chunks x 15 seconds
#6 6 chunks x 15 seconds
W576xH576
[ v3.4 ]
Long-form MiniMax H3 video and audio generation for ComfyUI with chunked generation, persistent references, restartable runs, and user-controlled audio.
### What's new in v3.4
- Driving Audio: preserves the supplied audio as the final audio while guiding generation across chunks.
- Video Reference: provides persistent visual reference for identity, motion, framing, and scene appearance.
- Restartable chunks: reuse completed chunks with Run Storage and regenerate only the required part.
- Improved Core compatibility: unknown upstream or custom nodes are not rejected merely because they are not recognized by Continuum.
- Simpler stable interface: obsolete compatibility controls and experimental Timeline inputs are hidden from the V3.4 public workflow.
- Spectrum interoperability: Spectrum remains optional and can use the official H3 Continuum Interop API.
### Direction change from v3.3
V3.4 focuses on predictable reference workflows rather than experimental Timeline Video and timeline-audio generation.
Driving Audio preserves the original user-supplied audio. Video Reference provides persistent visual guidance without requiring exact frame-by-frame copying. Existing V3.3 workflows remain available through legacy compatibility paths.
### Updating
For an existing Git installation:
git pull --ff-only origin mainV3.4 input connection patterns
V3.4 separates the visual reference input from the driving-audio input. Choose the connection pattern that matches your source material.
1. Audio only
Connect Load Audio to driving_audio. Use this when an existing song, dialogue track, or sound effect should remain the final audio. A Video Reference is not required.
2. Video with its own audio
Connect Load Video (Upload) IMAGE to Video Reference. If the uploaded video contains the audio you want to preserve, connect its AUDIO output to driving_audio as well.
3. Video and audio from separate sources
Connect Load Video (Upload) IMAGE to Video Reference, then connect a separate Load Audio node to driving_audio. Use this when the visual reference video and the final audio source are different files.
Both inputs are optional. Connect Video Reference when visual guidance is needed, and connect driving_audio when the supplied audio should be preserved in the final output.
Video Reference frame rate
Use a 24 fps source for Video Reference. Load Video (Upload) may accept files recorded at 25 fps or another frame rate, but acceptance alone does not guarantee correct temporal alignment with H3. For a non-24 fps source, set force_rate to 24 in Load Video (Upload), or convert the file to 24 fps before loading it. If the source is already 24 fps, leave force_rate at its default and do not resample it.
Current validation status
[ v3.3 ]
V3.3 adds Timeline Video conditioning for long-form MiniMax H3 generation. A reference video can now be processed in chunk-local time slices, allowing motion and scene continuity to be carried across multiple 5-second chunks while keeping the reference resolution independent from the output resolution. The Efficient 0.4 MP mode helps reduce memory usage and processing time.
Video assembly has also been improved. Auto seam handling analyzes chunk boundaries and applies guarded corrections for transient flicker, micro-flash, exposure, and color differences. This helps produce more natural transitions between generated chunks without changing the original sampling process.
Existing V3.2.4 workflows remain available as Legacy nodes for compatibility.
[ v3.24 ]
Generate longer native MiniMax H3 video and audio sequences in ComfyUI.
H3 Continuum is a ComfyUI custom node that generates a longer sequence as connected chunks and assembles them into one continuous video.
```text
3 × 5-second chunks → 15-second video
6 × 5-second chunks → 30-second video
The previous video and audio latent context is passed into each continuation chunk. This is not a simple video concatenation workflow.
Main purpose: longer MiniMax H3 generation, not faster generation.
Easy Installation
H3 Continuum can be installed directly from ComfyUI Manager.
Open ComfyUI Manager
Search for H3 Continuum or Continuum
Select Install
Restart ComfyUI
Load one of the included sample workflows
Manual installation and the latest documentation are available on GitHub:
GitHub:
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
What It Does
H3 Continuum divides a longer generation into manageable chunks.
MiniMax H3 Model
↓
H3 Continuum Sampler
↓
ComfyUI Core Video / Audio VAE Decode
↓
H3 Continuum Assemble
↓
Final videoEach continuation chunk receives latent context from the preceding chunk. Overlapping context is removed during assembly, and the final frame and audio counts are aligned to the requested duration.
Main Features
Connected long-form MiniMax H3 generation
Native video and audio latent continuation
Fixed, List, and Timeline prompt formats
Automatic prompt-format detection
T2VA, I2VA, FL2VA, Last Frame and Reference workflows
Up to three Reference Images
Reference Audio conditioning
First Frame and Last Frame conditioning
Configurable continuity context
Run Storage and automatic resume
Partial regeneration from a selected chunk
Optional Spectrum interoperability
Standard and Turbo sample workflows
ComfyUI Core VAE Decode compatibility
Included Sample Workflows
Two example workflows are provided.
Standard Workflow
Recommended when output quality and temporal consistency are the priority.
Standard MiniMax H3 sampling
Spectrum can be enabled
Suitable for quality-focused generation
Reference Image and Reference Audio supported
RTX upscaling can be enabled when required
Turbo Workflow
Recommended for faster tests and iteration.
LightX2V MiniMax H3 Turbo LoRA
8-step example configuration
Spectrum is bypassed by default
Faster than the standard workflow in tested configurations
Some loss of facial detail or additional artifacts may occur
Turbo LoRA models:
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
MiniMax H3 models and documentation:
https://huggingface.co/MiniMaxAI/MiniMax-H3
Models and LoRAs are not included with this custom node.
Reference + Continuation
Reference Images remain available across all generated chunks.
A typical setup is:
Picture 1 → face and identity
Picture 2 → full-body appearance and clothing
Picture 3 → environment or an additional visual reference
Audio 1 → vocal, music or audio-performance referenceRef2VA is the reference-specialized checkpoint and is generally the first choice for stronger reference fidelity.
FL2VA with Reference conditioning is also allowed. H3 Continuum does not automatically replace or switch the connected model.
Spectrum Integration
Spectrum is optional. H3 Continuum also works without it.
With a compatible Spectrum release, H3 Continuum sends a continuation signal only when generating later chunks.
Chunk 1 → normal Spectrum sampling
Chunk 2+ → Continuum Actual Prefix 2This allows Spectrum to coordinate its spectral forecasting with the continuation context instead of treating every chunk as an unrelated generation.
Benefits include:
Automatic identification of continuation chunks
Actual Prefix applied only where required
No manual prefix switching between chunks
Reduced risk of duplicated prefix processing
Compatibility with standard ComfyUI workflow execution
Spectrum remains an approximate accelerator. Motion, anatomy, audio and detail can differ from a non-Spectrum result, so quality comparisons should use the same prompt and seed.
Spectrum:
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
Run Storage and Resume
Enable Save + Auto Resume to preserve completed raw video and audio chunks.
If a generation is interrupted, H3 Continuum can reuse compatible saved chunks and continue from the first missing chunk.
It can also regenerate from a selected chunk while preserving the compatible prefix.
Chunk 1–3 completed
↓
Generation interrupted
↓
Queue the workflow again
↓
Chunks 1–3 reused
↓
Generation continues from Chunk 4Run Storage verifies the sampling contract, model route, references, resolution and saved chunk files before reuse.
Prompt Formats
Fixed
One prompt is used for every chunk.
List
Separate prompts are divided with:
---Timeline
[0-5s]
First scene description
[5-10s]
Second scene description
[10-15s]
Third scene descriptionPrompt Format = Auto detects the appropriate format automatically.
Incomplete timeline coverage produces diagnostics and safe fallback behavior rather than unnecessarily stopping every generation. Structurally unusable input is still reported as an error.
Tested Configuration
The current Windows implementation has been tested with:
GPU NVIDIA RTX 5060 Ti 16GB
ComfyUI MiniMax H3-compatible Core build
Chunk Duration 5 seconds
Typical Length 3 or 6 chunks
Continuity Balanced 22 frames
Standard Sampling RES Multistep
Spectrum Interop Actual Prefix 2The node is not limited to RTX 50-series GPUs. Actual compatibility, generation speed and usable resolution depend on the MiniMax H3 model, GPU memory, ComfyUI configuration and installed acceleration nodes.
RTX 4060 and other configurations have not been formally validated by this project.
Frequently Asked Questions
Is this only a workflow?
No. H3 Continuum is a ComfyUI custom node package. The included workflows are ready-to-use examples.
Does it generate one native 30-second sample?
No. It generates connected chunks and assembles them into one longer output while carrying video and audio latent context forward.
Does it make MiniMax H3 faster?
Speed is not the primary purpose. H3 Continuum is designed for longer generation. Spectrum and Turbo LoRAs can reduce generation time in some configurations.
Is Spectrum required?
No. It is an optional acceleration and interoperability path.
Can I use the Turbo LoRA?
Yes. A Turbo sample workflow is provided. Spectrum is bypassed by default in that workflow because combining both can change quality or introduce artifacts.
Which model should I use for Reference Images?
Ref2VA is the reference-specialized option. FL2VA with Reference conditioning is also allowed, but reference fidelity may differ.
Are the models included?
No. MiniMax H3 checkpoints, text encoders, VAEs, Turbo LoRAs and optional acceleration nodes must be installed separately.
Can an interrupted generation be resumed?
Yes. Enable Run Storage before generation. Compatible completed chunks can then be reused.
Can I regenerate only the later part?
Yes. Run Storage supports regeneration from a selected chunk while retaining a compatible earlier prefix.
Are chunk boundaries always invisible?
No generative continuation system can guarantee a completely invisible boundary. H3 Continuum preserves latent context and removes duplicated overlap, but difficult motion, lighting changes and large prompt transitions can still produce flicker or visual changes.
Does Reference Audio guarantee exact lip synchronization?
Reference Audio conditions MiniMax H3’s native joint video/audio generation. It can guide vocals, rhythm, expression and mouth movement, but it does not guarantee sample-identical audio reproduction or frame-perfect lip synchronization in every generation.
Does it support audio continuity?
Yes. Video and audio latent context are carried together. The assembler also provides an optional Audio Seam mode for boundary-local audio correction.
Is RTX 5090 required?
No. Development and runtime validation were performed on an RTX 5060 Ti 16GB. Lower-memory configurations may require reduced resolution, offloading or other ComfyUI memory optimizations.
What license is used?
H3 Continuum is released under the MIT License.
Links
Custom Nodes
The included Standard and Turbo workflows use the following custom nodes.
- H3 Continuum
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
- rgthree-comfy
https://github.com/rgthree/rgthree-comfy
- ComfyUI-Easy-Use
https://github.com/yolain/ComfyUI-Easy-Use
- ComfyUI-KJNodes
https://github.com/kijai/ComfyUI-KJNodes
- ComfyUI-Spectrum-MiniMax-H3
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
Spectrum are optional generation paths, but installing all listed custom nodes allows the included workflows to load without missing-node warnings.
Models
- MiniMax H3
https://huggingface.co/MiniMaxAI/MiniMax-H3
- LightX2V MiniMax H3 Turbo LoRA
https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
Models and LoRAs are not included in the workflow ZIP.
Main Links
- GitHub and documentation
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum
- Install from ComfyUI Manager
Search for H3 Continuum
Description
FAQ
Comments (45)
Has anyone tried it with 8GB of VRAM?
I'm on 8GB VRAM. 32 RAM. Works well for me.
What's your production speed for a 15-second video?@hatt2
Hi, If I need to modify only the second part keeping the first part of the generation is it possible with this workflow?
Yes, it is possible, but Run Storage must be enabled during the original generation.
For example, if your video is configured as 2 chunks × 5 seconds, Chunk 1 is the first 5 seconds and Chunk 2 is the second 5 seconds.
Before the first generation:
Set Run Storage to Save + Auto Resume.
Give the run a fixed Run Name.
Keep a fixed Base Seed.
Generate the complete 10-second video normally.
Continuum will save the generated latent data for both chunks.
To create a different second part:
Keep the same Run Name, Base Seed, model, LoRA, resolution, sampler, steps, references and Continuation settings.
Set Regenerate From to Chunk 2.
Queue the workflow again.
Continuum will load the saved Chunk 1 instead of sampling it again. It will use the end of Chunk 1 as the continuation context and generate a new Chunk 2. The status report should show:
1 reused, 1 generatedThis means the first 5-second chunk was preserved and only the second 5-second chunk was sampled again. The complete 10-second video is then decoded and assembled again.
If you want different instructions for each part, prepare separate prompts before the first generation:
Prompt for the first 5 seconds
---
Prompt for the second 5 secondsChanging the model, LoRA, resolution, sampling settings, references or prompt plan after the original run may create a new Run Storage revision. In that case, Continuum may need to regenerate the complete video.
If the original video was generated with Run Storage = Off, the first chunk was not saved and cannot be reused afterward.
Important exception: an FL2VA workflow using Long Terminal Merge samples the final two chunks as one unit, so those two chunks must be regenerated together.
@ukr8b3g201 Thanks!!
Back with more questions. 🙂
Is it possible for us to change the Shift?
- Video shift
- Audio shift
Yes. Current ComfyUI Core provides a ModelSamplingMiniMaxH3 node with separate controls for:
Video shift — default 12.0
Audio shift — default 3.0
Place it after the MiniMax H3 diffusion-model loader and before the rest of the model chain connected to Continuum. Continuum already preserves this official upstream model patch, so no special Continuum implementation is required.
Changing either value alters the denoising schedule, so it should be treated as an advanced/experimental setting and regenerated from scratch rather than reusing an older Run Storage revision. I have not established recommended alternative values yet.
@ukr8b3g201 Ok thank you for confirming!
thanks for the workflow. 2 things could any share some of the prompts cause i not think i fully understand how to prompt for this and does any one know why im getting random dialog that is not in the prompt
thanks
IF I understood it correctly which seems to work in my prompting...separate chunks (video parts) as [Chunk 1] etc, then in the main thing you set how many chunks to generate, and the chunk length aka seconds/duration of each chunk
for example
[Chunk 1]
bla bla bla character picture 1 bla bla
she walks in and waves
[Chunk 2]
bla bla bla character picture 1 bla bla
she walks out and closes the door
random dialogue (if using reference audio) might be due to sampler/scheduler, too few or too many steps, random words outside of desired places, like putting trigger words in the beginning of the prompt (happened to me several times), or you don't shut up the character by saying something like 'she finishes speaking her sentence'
after the dialogue
OR you have some other trigger words in the prompt which can have a starting effect, like "says" "said" "voice" etc
ya my prompt look the same and ive put the she finishes speaking her sentence at the end of dialog and tried removing my audio ref still doing. could you tell me what sampler/scheduler your using and any speed lora and how many steps you do please. thank you
@themarkedone830 right now for this workflow Euler - beta57, 4 steps first run and 4 steps 0.4 denoise second run with x1.0 upscale factor... seems aight but I'm still experimenting
using minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600_comfy_fro_no_adaln_proj.safetensors lora on 1 strength, perhaps important factor.. I'm using a hybrid checkpoint called minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors
There's a bunch of different ways to prompt. You can see another method from the skateboarding video.
I tend to put complete prompts for every segment, so far gives me the best results, since I mainly do longer segments (15s chunks). Technically shouldnt make a difference, but its what I had the best experience with.
The checkpoint I use is the minimaxH3Ref2vaXUELUO_v10.safetensors one, with int8 Text encoder and VAE (links on the civitai site of the checkpoint)
Then is use an 8 Step lora minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
Sampler is er_sde and simple scheduler.
At 1mp with 2x RTX upscale that usually gives me enough quality that I dont need the 2nd pass, but if I want, ill add the 10 steps with only 1x scale, just to refine the video.
Also I find that seed hunting is very much helpful, I find myself often just doing a 5s segment of the video to find a good starting scene and then screenshotting it and using it as a reference image to cut down on the variance.
For now this format is working for me. If I deviate from this I get weird results. Like bad camera work or voices that shouldn't be talking etc
Line 1 [Chunk 1]
Line 2 integrated_multimodal_description: [Shot 1] detail scene
Line 3 overall_soundscape: detail sound stuff
Line 4 non_diegetic_music: N/A
Line 5
Line 6 [Chunk 2]
Line 7 integrated_multimodal_description: [Shot 1] detail scene
Line 8 overall_soundscape: detail sound stuff
Line 9 non_diegetic_music: N/A
Line 10
Line 11 [Chunk 2]
Line 12 integrated_multimodal_description: [Shot 1] detail scene
Line 13 overall_soundscape: detail sound stuff
Line 14 non_diegetic_music: N/A
Line 15
thanks everyone go for the help
This is my favorite workflow, and tonight v37 dropped, however i have a few questions (am novice)
1. for v37 it says i need the continuumv37 sampler, and after git clone of the repo, and reboot i still dont have this node.
2. i still dont understand fully the chunks, i can do a 30 second video by selecting 15seconds @ 2 chunks, but my output is the same video twice in 30 seconds, like a loop, not a continuation or a 30 second long single scene.
3. how do i change the resolution of my image? i see the 4mp and 6mp options, but unsure of how to upscale this to get a fhd 2k or 4k output.
thank you for your work on this workflow, its one of my favorites, but im sure im barely using it as intended.
Thank you — I’m glad you enjoy the workflow! Here are answers to each question:
1. V3.7 Sampler node is missing
The display name is *H3 Continuum Sampler V3.7**.
Please confirm that:
- The repository is directly under ComfyUI/custom_nodes/ComfyUI-H3-Continuum
- ComfyUI is version 0.32.0 or newer
- You completely restarted the ComfyUI backend, then refreshed the browser
- You do not have both a Manager installation and a second manual clone
For an existing Git installation:
```bash
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
```
During startup, you should see something similar to:
```text
H3 Continuum 3.7.0 loaded
```
If that line does not appear, please send the startup-console error around ComfyUI-H3-Continuum.
The V3.6 sampler remains compatible, so you can continue using it while resolving the installation. V3.7 does not change the existing Production defaults.
2. The 30-second result appears to repeat
Continuum does not generate one native 30-second H3 sample. With 2 chunks × 15 seconds, it generates two physical sections and passes the end context of the first section into the second.
If both chunks receive the same Fixed prompt, H3 may repeat the same described action. For more deliberate progression, use Chunk/List prompts:
```text
[Chunk 1]
The woman enters the hallway. The camera follows her from behind.
[Chunk 2]
Without a cut, she continues forward, turns right, and enters the next room.
The camera keeps tracking her in the same direction.
```
Avoid restarting the action in Chunk 2—for example, do not say “she enters the hallway” again.
For initial testing, I recommend:
- Standard continuation
- Balanced — 22 frames
- Audio Continuity ON
- 3 × 10 seconds or 6 × 5 seconds
- A different prompt section for each chunk
If the two halves are literally identical rather than merely similar, temporarily set Run Storage = Off, disconnect Video Guide Frames if used, and check the Detailed Report for reused chunks.
3. Resolution and 2K/4K output
The Efficient 0.4 MP and Balanced 0.6 MP controls are for Reference/Video Guide preprocessing. They do not set the final output resolution.
The First Pass output resolution comes from the workflow’s Resolution Selector, which feeds the Sampler width and height. H3 is best used near its native canvas—up to approximately a 768-pixel short edge and 768×1344, aligned to multiples of 32.
For FHD/2K/4K, the practical approach is:
```text
Generate near native H3 resolution
→ Assemble the complete video
→ Apply a video upscaler afterward
```
The workflow also includes the Experimental H3 Continuum Hi-Res Fix V3.5. A 2× setting can produce a higher-resolution generative Second Pass, but it is much slower and uses substantially more VRAM. A single 576→1152 five-second test passed, while long 2× sequences can exceed 16 GB VRAM.
Therefore, for a 30-second video I recommend generating at native H3 resolution first and using a post-process video upscaler for 1080p, 2K, or 4K.
1. The V3.7 Sampler node is missing
The exact node name is H3 Continuum Sampler V3.7.
Cloning the repository does not update an older installation that may already exist. If you previously installed Continuum with Git, please run:
cd ComfyUI/custom_nodes/ComfyUI-H3-Continuum
git pull --ff-only origin main
Then completely restart the ComfyUI backend and refresh the browser.
If you installed it through ComfyUI Manager, use Manager’s Update button instead. Please do not keep both a Manager installation and a second manual Git clone, because duplicate copies can cause ComfyUI to load the wrong version.
During startup, look for:
H3 Continuum 3.7.0 loaded
Then search for the exact node name:
H3 Continuum Sampler V3.7
If it is still missing, please share the startup-console lines mentioning ComfyUI-H3-Continuum. The V3.6 sampler remains compatible while the installation issue is being resolved.
@ukr8b3g201 What a great and detailed reply, Thank you
1. i was able to resolve the "missing node" i think you were right about cached browser, i chose rightclick reboot from the comfyui window, but nodes did not refresh, so after closing and restarting the red missing node disappeared.
2. i will use [Chunk 1] , [Chunk 2] etc.. to space out my story, follow up on this, if i describe the details of the character at the top will it follow all the chunks? or do i have to re-add them? some chunks i forsee the character may go "off screen", how do i keep continuity?
also what are the maximum chunks, can i create a whole story like 100 chunks?
3. ok generate, then upscale afterwards, do you recommend/create any?
at this point im hoarding workflows and going cross eyed, but there are a handful (like this one) which are just great!, keep up the awesome work!
@gigomikol 最後の追加質問への返答案です。そのままCivitaiへ投稿できます。
You’re welcome—and yes, those are good follow-up questions. 😊
1. Shared character description and off-screen continuity
Text placed before the first [Chunk 1] header is treated as a global preamble. Continuum automatically includes it in every chunk prompt.
For example:
GLOBAL CONTINUITY:
The same young woman appears throughout the story. She has long black hair, green eyes, a red jacket, black jeans, and silver earrings. Her face, clothing, age, and body proportions remain unchanged in every chunk.
The location remains the same modern hotel, with warm evening lighting and dark wooden walls.
[Chunk 1]
The woman enters the hotel hallway. The camera follows her from behind.
[Chunk 2]
Without a cut, the same woman continues down the same hallway. She turns right and exits the frame while the camera remains in the corridor.
[Chunk 3]
The same woman re-enters from the right side, still wearing the same red jacket, black jeans, and silver earrings. She opens the next door.You therefore do not have to copy the complete character description into every chunk.
However, for important characters and long sequences, I still recommend briefly repeating the most important details when the character returns:
The same woman re-enters from the right, wearing the unchanged red jacket and silver earrings.This gives H3 both the global identity description and a local reminder.
If the character remains off-screen for several chunks, identity drift becomes more likely because there is no visible character information in the recent video context. To reduce this:
Connect a clear character image through Reference Image
Keep the shared identity description above [Chunk 1]
Explicitly say that the character is temporarily off-screen rather than replaced
Restate her identity, clothing, and entry direction when she returns
Avoid keeping an important character absent for too many consecutive chunks
Reference conditioning and Continuum context improve consistency, but they cannot guarantee perfect identity over an unlimited story.
2. Maximum number of chunks
The current sampler supports 1–16 logical chunks, so 100 chunks cannot be generated as one Continuum run.
Although 16 chunks are available, I would not recommend treating one run as a complete movie. Small differences in identity, lighting, camera position, contrast, and motion can accumulate as the sequence becomes longer.
A more reliable structure is:
Scene 1: 3–6 chunks
Scene 2: 3–6 chunks
Scene 3: 3–6 chunksGenerate each scene separately, then use an accepted final frame from the previous scene as the next scene’s First Frame. Keep the same Reference Image and global character description, and assemble the completed scenes afterward in a video editor.
That gives you natural scene boundaries and makes it much easier to reroll or repair only one part of the story.
3. Post-upscaling recommendations
For a beginner, I recommend keeping upscaling separate from the Continuum generation workflow:
Generate near native H3 resolution
→ assemble the complete video with audio
→ upscale the finished videoThe easiest paid option is Topaz Video. It accepts a completed video directly and is simpler than processing frames manually.
Inside ComfyUI, the lighter option is a conventional ESRGAN-style image upscaler using the built-in UpscaleModelLoader, followed by video reassembly. This is relatively easy and works on modest hardware, but independent frame upscaling may introduce some temporal shimmer.
A more advanced video-aware option is SeedVR2 Video Upscaler. It can preserve temporal consistency better, but it is significantly heavier. Its documentation warns that even the 3B model may need around 18 GB of VRAM, so it is not my first recommendation for a novice or a 16 GB GPU.
I may provide a separate optional post-upscale workflow rather than adding more complexity to the main Continuum workflow. For now, generate and approve the complete video first—then upscale only the final accepted result.
And thank you again. “Hoarding workflows and going cross-eyed” is a very accurate description of learning ComfyUI. 😄
@ukr8b3g201 Awesome, great information for my toolkit :) one last post (for today) you mentioned
"Generate each scene separately, then use an accepted final frame from the previous scene as the next scene’s First Frame"
if i create a low resolution video for now before upscaling, but want to go forward with creating the next "chunk batch", how do i use your workflow so that when the video is finalized the last frame is generated as a high resolution still (i hope this makes sense)
thereforce i can use that as the first frame like you mentioned, without having to wait for the upscale and try to screenshot it from the up-scaled video?
I promise i wont use this as tech support, i appreciate all of your feedback
@gigomikol Yes, that makes sense—and you do not need to upscale the entire video or take a screenshot.
Connect the images output from H3 Continuum Assemble + Seam V3.5 to the ComfyUI Core node:
Get Image from Batch
batch_index = -1
length = 1-1 selects the final frame of the assembled video. Connect its output to Save Image, then load that saved image as the First Frame of the next scene batch.
I recommend using this original native-resolution final frame for continuation. It matches the preceding video exactly and H3 will resize it internally to the next generation canvas. A separately upscaled image does not normally give H3 additional native detail and may introduce sharpening or changed details at the scene boundary.
If you also want a high-resolution still for archiving, create a second branch:
Assemble images
→ Get Image from Batch (-1, length 1)
→ Image Upscale With Model
→ Save ImageUse the original extracted frame for the next batch, and keep the upscaled copy as an optional high-resolution still.
Make sure you extract the frame after H3 Continuum Assemble + Seam V3.5, rather than directly from an individual decoded chunk. That gives you the actual final frame after assembly and exact-duration processing.
I may add this small “Save Final Frame” branch to a future workflow template because it is useful for building longer stories scene by scene.
@ukr8b3g201 absolute legend, thank you for taking the time to help me!
@ukr8b3g201 wording "Without a cut" dont work, there clearly visible cut in movement, colors and surrounding. And in sound.
Hello,
I have more questions. 😅
I was reading the official minimax H3 documentation today for REF2VA.
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
I was wondering if we should use the same format rules for Continuum? With their format, I'm not sure where to put [Chunk 1] [Chunk 2] etc
subject_definitionsDefines referenced content and its reference labelssummarySummarizes the task type, target video, and main reference relationships
retention_analysisDescribes how referenced content is preserved, transferred, or reused
detailed_descriptionDescribes visuals, actions, shots, sound, and dialogue in playback order
overall_soundscapeSummarizes ambience and physical sounds
non_diegetic_musicDescribes background music audible only to the audience
技術的ですが、結論は明確です。REF2VAを使う場合は公式の6セクション形式が推奨ですが、[Chunk N]はMiniMax形式ではなくContinuum側の振り分け記号です。
Civitaiには、次のように返すと分かりやすいです。
Yes—when you are actually using the Ref2VA model and Reference inputs, following MiniMax’s six-section Ref2VA format is a good idea. However, it is not required for every Continuum workflow. T2VA/I2VA/FL2VA can continue using the normal H3 prompt format.
The important distinction is:
subject_definitions, summary, retention_analysis, etc. are instructions for the H3 model.
[Chunk 1], [Chunk 2], or [0-10s] are routing headers used only by Continuum.
Continuum removes the Chunk header before sending the resolved prompt to H3.
For two 10-second chunks, I recommend Prompt Format = Timeline and this structure:
subject_definitions:
<Subject 1> is the woman referenced from <Picture 1>, preserving her face, hair, clothing, and overall identity.
[Chunk 1]
summary:
[reference generation] The first target segment introduces <Subject 1> and begins the action.
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - her identity, face, hairstyle, and clothing are retained.
detailed_description:
[Shot 1] A medium close-up introduces <Subject 1>...
[Shot 2] At 00:05.000, the camera moves...
overall_soundscape:
The environmental ambience and physical sounds for the first segment...
non_diegetic_music:
A restrained electronic score begins and continues without resolving.
[Chunk 2]
summary:
[reference generation] The second target segment continues the same subject, environment, action, and audiovisual direction.
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the same identity, clothing, and visual characteristics continue.
detailed_description:
[Shot 1] The action continues immediately from the inherited ending state, without resetting the pose, camera direction, lighting, or environment...
[Shot 2] At 00:05.000, the camera moves...
overall_soundscape:
The ambience and physical sounds continue naturally from the previous segment.
non_diegetic_music:
The same score continues across the segment boundary without restarting.Text before the first [Chunk 1] header is a global preamble. Continuum automatically places it at the beginning of every resolved chunk prompt, so shared subject_definitions are a good fit there.
Each chunk should still contain its own:
summary
retention_analysis
detailed_description
overall_soundscape
non_diegetic_music
Within each chunk, restart the model-facing shot notation at [Shot 1]. The [Shot N] timestamps are local to that chunk, while [Chunk N] belongs to Continuum.
A few important points:
Put [Chunk 1] and [Chunk 2] on separate lines.
Do not mix [Chunk N] sections with --- List separators.
Define every configured chunk; otherwise Continuum may reuse the previous prompt as a fallback.
If one identical full Ref2VA prompt should apply to every chunk, simply use Fixed mode—no Chunk headers are needed.
Use [video continuation] in the Ref2VA summary only when an actual reference video is being continued. Continuum’s internal latent continuation alone does not automatically make the external Ref2VA task type video continuation.
Also, if a reference image only defines a person’s identity, define it as a <Subject N> sourced from <Picture N>. Use a standalone <Picture N> entry mainly when that picture is a concrete First Frame, Last Frame, keyframe, or composition anchor.
The six-section rules come from MiniMax’s official Full-Reference Mode guide. Continuum does not rewrite those sections; it only selects the appropriate complete prompt for each chunk and passes it through to H3.
ok thank you for the informative reply! I have a lot of studying to still do. I will be back with some test results soon ! 🙂
First thanks for your hard work!
In the last version (3.7) seem is broken the spectrum, if enabled the workflow will stop the execution.
Thank you for the report, and for your kind words. Spectrum passed our V3.7 validation, so this may depend on the ComfyUI/Spectrum version or workflow configuration. Could you please share the full console error, your ComfyUI and Spectrum versions, and the workflow JSON? Also, please let us know whether it stops with Spectrum alone or only when combined with SageAttention or Sol-Attn. We’ll investigate it.
@ukr8b3g201 Solved. There was an update for the Spectrum node too yesterday. I've updated and now 3.7 work with Spectrum enabled.
Can you help me, please. I do 3 chunks, and all 3 chunks starts from beginnings, from first frame. Where my mistake?
Could you describe exactly how you are generating the video?
Which workflow and H3 Continuum version are you using?
Which node is set to 3 chunks?
What are your Chunk Seconds, Continuity, and Continuation Backend settings?
Where is the starting image connected: First Frame, Video Guide Frames, Still Image Guide, or somewhere else?
Are you using one Continuum Sampler, or three separate sampler nodes?
Are you seeing the restart in the individual chunk previews, or also in the final video after H3 Continuum Assemble?
At what exact times does the video return to the first frame?
A screenshot or video is optional. A written description of your connections and settings is enough for the first check.
Using Continum v37 just downloaded.
H3 Continuum Sampler V3.7 node - second choise - "Chunks"=3
Chunk seconds 15, Continuity=22, Continuity backend=standart, Spectrum=off
I see the final video after assembling.
The video returns to the first frame each 15 seconds.
Tried the Eros, PinkCherry and Dasiwa checkpoints.
My prompt devided by "---" contains only a description of the actions of a specific ---fragment---.
Only the first paragraph contains a basic description of the character and the reference. Is this right?
And one more question, please. How can I use a specific LORA for a specific chunk?
@jafdeth2030105 Thanks, that information helps.
With one H3 Continuum Sampler V3.7, the second and third chunks should continue from the preceding chunk. Returning exactly to the original first frame at 15 and 30 seconds is not expected behavior.
Please try this clean test:
Connect the starting image only to First Frame.
Disconnect Video Guide Frames, Still Image Guide, Last Frame, and all Reference inputs.
Use one V3.7 sampler and one H3 Continuum Assemble path.
Set Continuity = Balanced 22, Continuation Backend = Standard, Audio Continuity = ON, and temporarily set Run Storage = Off.
First test 3 chunks × 5 seconds. If it continues correctly, repeat the test with 15-second chunks.
About the prompt: when using --- List mode, every section is an independent chunk prompt. The character description in the first section is not automatically added to the later sections. Repeat the persistent character, appearance, location, and current-state details in all three sections.
Alternatively, use Timeline format and place the shared description before the first time header:
Shared character and scene description.
[0-15s]
First action...
[15-30s]
Continuing from the previous action...
[30-45s]
Continuing from the previous action...Missing descriptions in later prompts can cause identity drift or abrupt cuts, but it should not normally cause an exact return to the original first frame.
If the clean test still restarts at exactly 15 and 30 seconds, please attach the workflow JSON and the sampler’s Detailed Report. That would allow me to determine whether this is a connection problem or a reproducible V3.7 issue.
Regarding LoRA: V3.7 currently receives one MODEL chain for the entire sampler run. A different LoRA for each chunk is therefore not supported inside one three-chunk sampler execution. Separate sampler runs followed by manual assembly are possible, but they do not preserve the same native Continuum sequence contract.
@ukr8b3g201 Thanks, I'll try and give you feedback.
@ukr8b3g201 I disconnected the "video guide frames", and only then did everything start working. Thanks for help!
@ukr8b3g201 always cuts regardless of description.
Have you any stable example which actually have no any cuts?
@velanteg Yes, I understand what you mean. At the moment I don’t want to claim that Continuum can reliably produce a completely invisible transition at every chunk boundary.
The previous “Without a cut” wording was only prompt guidance — it is not a real fix for chunk-boundary continuity.
In my tests, Continuum can preserve the sequence across chunks, but depending on the seed and scene there can still be noticeable changes in motion, color, environment, or audio at the boundary. I’m treating this as a Continuation Quality issue and investigating the boundary behavior itself rather than trying to solve it with prompt wording.
I’ll prepare a reproducible test/example using fixed settings so I can show both what currently works and where the limitation remains.
How do I avoid videos slowly getting "burned" with each extension?
I second that question, Ive noticed with recent releases that skin degrades a lot more often the longer the video goes.
@Xarfai Thanks to both of you for reporting this. I’m investigating the cumulative “burned,” oily, or degraded-skin appearance during long extensions.
My current tests suggest that accumulated visual self-conditioning may become too strong under some continuation settings. As a temporary workaround, please try:
Continuation Backend: Standard
Continuity: Balanced — 22 frames
Avoid Compatibility / Reference Context for long sequences
If possible, please share your workflow JSON, Continuum version, number and length of chunks, Continuation Backend, Continuity setting, and whether First Image, Last Image, or References were used. A comparison showing the first and last chunks would also be very helpful.
Since you mentioned that it happens more often in recent releases, I’ll also compare the same workflow across versions. I’ll report back once I can isolate whether this is a regression or a setting-specific accumulation issue.
@ukr8b3g201 I'm using 10Eros Beta4 as the checkpoint and it has a Turbo Lora merged into it. I did 10 steps (recommended is 6-8) and three 15s chunks. By the third chunk quality was severely degraded.
@OhNoMyChickens Thanks — that configuration is very useful information. The strongest suspect here is the Turbo-merged checkpoint being run at 10 steps, above its recommended 6–8 steps. Turbo models can become overprocessed or “burned” when oversampled, and Continuum may carry that degradation into later chunks.
Please first try:
Use the checkpoint’s recommended 6–8 steps
Continuation Backend: Standard
Continuity: Balanced — 22 frames
Test 5–10 second chunks instead of 15 seconds
If possible, also compare the same prompt with a clean non-Turbo checkpoint. Because the Turbo LoRA is merged, its strength cannot be reduced independently. A clean checkpoint comparison will help determine whether this is primarily Turbo oversampling or Continuum accumulation.
For diagnosis, please share the workflow JSON and first/third chunk comparison if you can. I’ll include this configuration in the cross-version test.