Getting Started: Chunks, Prompts & Reference Images
What’s New — H3 Continuum v3.10.2
Introduced in v3.10.0, Project / Take Manager lets you organize Runs into named Projects, label favorite Takes as savepoints, and try another continuation—or start an independent trial while keeping earlier results. v3.10.2 also protects finalized Takes when continuation fails.
The panel is optional and defaults OFF. Your existing workflows work as they are—no extra nodes or rewiring required. Existing Takes are preserved.
H3 Continuum v3.9.8
Improves Chunk 1→2 continuation with supported KJ Preview Override builds. GPU-tested with PyTorch and Comfy Kitchen INT8 + audited local Spectrum.
Scope: audited KJ/Core builds only; Preview VAE/Audio VAE disconnected; Tiny VAE none.
Restart after updating and use a new Run. Existing Takes and workflows are preserved.
v3.9.7 fixes an issue where ComfyUI’s cache could skip regeneration when Chunks or Variation Nonce were connected to external nodes.
GPU testing on the published version confirmed that regenerating Chunk 2 twice with identical settings advanced Variation from 1 to 2 while preserving Chunk 1. Existing Takes and workflows remained unchanged, with no execution errors observed.
Update the custom nodes, restart ComfyUI, and refresh your browser. No workflow changes are required.
3.9.3 — choose temporary storage (2026-10-13)
Choose where Continuum writes its temporary decoded frames and Finalize buffer. Open Settings → H3Continuum → Storage → Manage temporary storage. Enter an absolute folder on the ComfyUI server, or choose a server storage candidate, then validate and save. Leave it blank to keep the existing locations. No workflow changes are required; continue using the V3.9 nodes and your 3.9.1 / 3.9.2 workflows
v3.9.2 is now available!
This minor UI hotfix reduces unnecessary Review checks, duplicate Reference Images notifications, and redundant node resizing.
Update only the custom node package and refresh your browser. All v3.9.1 workflows can be used unchanged.
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
✅ V3.8 Hotfix Released: The reported chunk continuation and Review Each Chunk issues have been fixed. Please update to the latest GitHub version using ComfyUI Manager → Update or git pull --ff-only origin main, restart ComfyUI, and hard-refresh your browser. Existing saved Takes are preserved. Thank you for your reports and patience.
FAQ
Comments (36)
chunk re-roll is a great feature... thx again.
Thanks! Glad you’re finding chunk re-roll useful. 😊
@ukr8b3g201 that 20s walking girl i'm posting had to re-roll 2 and 4 a couple of times, but turned out nice
I installed version 3.8 and used it for the first time about five hours ago (the h3_continuum folder had not yet been created).
Settings: Chunks = 3, Seconds per Chunk = 5, Run = Review Each Chunk, Progress = On (Resume and Takes available).
After Chunk 1 was generated, selecting "Use it and continue" and running it resulted in Chunk 1 being generated again instead of proceeding to Chunk 2.
after Chunk 1 was generated, selecting "Render History" -> "Continue From Here" and running it allowed me to proceed to Chunk 2.
After that, I was able to proceed to Chunk 3 by selecting "Use it and continue."
Is this normal behavior?
No, that is not the expected behavior.
With Chunks = 3, Seconds per Chunk = 5, Run = Review Each Chunk, and Progress = On, after Chunk 1 is generated:
Use it and continue should reuse the saved Chunk 1 and generate Chunk 2.
It should not generate Chunk 1 again.
The behavior you described matches a Review Each Chunk continuation issue that was fixed in the V3.8 hotfix on September 8.
Please update H3 Continuum to the latest main version (or use ComfyUI Manager → Update), then:
Restart ComfyUI.
Hard-refresh the browser page.
Try the same 3 × 5-second Review Each Chunk sequence again.
Your existing saved Takes should not be deleted by the update.
If you were already using the latest main version when this happened, please let me know. In that case, I would like to investigate it as a possible remaining first-run initialization issue.
@ukr8b3g201 Just to be safe and ensure the environment was identical to the initial state, I deleted the h3_continuum folder and ran git pull --ff-only origin main, but only README.md and README_JA.md were updated. The result was the same.
@ana25077166
It looks like you deleted the `ComfyUI-H3-Continuum` custom node folder and then ran `git pull`. After deleting the folder, `git pull` cannot reinstall it and may update the parent ComfyUI repository instead. This would explain why only `README.md` and `README_JA.md` were updated.
Please reinstall H3 Continuum using either of these methods:
**ComfyUI Manager**
1. Open ComfyUI Manager.
2. Open the Custom Nodes Manager.
3. Search for `H3 Continuum`.
4. Click `Install`.
If Manager still shows it as installed after you manually deleted the folder, click `Uninstall` to clear the stale entry, reopen the Custom Nodes Manager, search for `H3 Continuum` again, and click `Install`. Do not use `Try update` after the custom node folder has already been deleted.
**Git**
Open the `ComfyUI/custom_nodes` folder and run:
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git
After reinstalling, restart ComfyUI, hard-refresh the browser, and retry the same 3 × 5-second Review Each Chunk test.@ukr8b3g201 When I normally test a new version, I make sure to perform a fresh installation via git clone without any existing old-version folders present. I also use the new version's workflow rather than reusing an existing one, and I limit the added nodes to the bare minimum—such as ModelAttentionBackend—for testing purposes.
Naturally, I do not delete the custom node folder before updating via git pull.
I also tried a fresh installation via ComfyUI Manager, but the result was the same.
After Chunk 1 was generated twice, instead of selecting "Render History" -> "Continue From Here" as I had previously, I selected "Use it and continue"; this allowed the process to proceed to Chunk 2.
Subsequently, when I tested a new scene with a different reference image and modified prompts, "Use it and continue" appeared to function correctly.
ComfyUI 0.35.0
Prompt Format＝List
Continuity＝Balanced — 22 frames
Base Seed＝42
Control After Generate＝fixed
Audio Continuity＝true
Chunks＝3
Seconds per Chunk＝5
Total Length 15 seconds
Size Source Manual
Width 480
Height 864
Run＝Review Each Chunk
Progress＝On— Resume and Takes available
Reference Image Size＝Match Output
Regenerate From＝Auto
Run Name (Optional Override)＝Default (blank)
Continuation Backend＝Standard
@ukr8b3g201
Instead of using "Use it and continue," I decided to test the scenario where "Regenerate From" is explicitly set to "Chunk 2"—an option that existed in the previous version as well.
To ensure a clean slate, I deleted the "H3_continuum" folder within the "output" directory.
After that, I launched ComfyUI.
I ran the process with "Regenerate From" set to "Auto" to create Chunk 1.
Then, I ran it with "Regenerate From" set to "Chunk 2."
This resulted in the following error:
# ComfyUI Error Report
## Error Details
- Node ID: 312
- Node Type: H3ContinuumSamplerV38
- Exception Type: D:\MiniMax H3\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-H3-Continuum.run_storage.RunStorageError
- Exception Message: D:\MiniMax H3\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-H3-Continuum.run_storage.RunStorageError: completed review group (1, 1, 1) does not match resolved unit (2, 2, 2)
(The rest is omitted due to length.)
After the error occurred, running the process again as-is resulted in the generation of Chunk 2.
The video generated by creating chunks individually and sequentially differed from the video generated all at once using "Generate Full Video"; the subject's movements were different, making them distinct videos rather than identical ones.
I have a vague recollection that this was the case in the older version as well, but I am curious if this behavior is considered normal.
awesome... audio hold up nicely between chunking and multiple re-rolls.
this workflow definitely deserves more attention than it's getting
Thanks — I really appreciate that.
I do think Continuum probably deserves a bit more attention, but part of the reason it is still relatively unknown is simply that I haven't promoted it much on Reddit or elsewhere.
I also know the workflow can be a little intimidating at first. It is not really an all-in-one setup, and once you get into chunk continuation, Takes, Run Storage, Timeline prompts, references, and regeneration, there is quite a lot to learn. If someone is not already fairly comfortable with ComfyUI, I can understand why it may feel a bit difficult to approach.
V3.8 was a good example of that challenge. I experimented with several UI changes to make things more intuitive, but at some point changing too much can actually make the workflow more confusing, especially for existing users. There were also a few bugs around the initial V3.8 release that needed hotfixes, so I am trying to be a bit more careful about adding features faster than I can properly test them.
The newer reference-related features are probably the area where I still need more testing. Driving Audio has had more practical use, but some of the Reference Audio and multi-reference paths were added fairly quickly because people were asking for them. So there may still be edge cases there that I have not fully validated yet.
At the same time, I do want to keep adding newer H3 features when they make sense for Continuum. I like the idea of keeping the workflow current, but the difficult part is finding the balance between adding useful new features and keeping the whole thing stable enough to actually use.
Overall, I think Continuum is in a reasonably solid place now for what it is trying to do, especially for reviewing chunks, re-rolling parts of a sequence, continuing from earlier chunks, and keeping audio reasonably consistent.
The main thing I still struggle with is very long generation. Once you get into something like 6 × 5s, 6 × 10s, and beyond, visual drift can become more noticeable over time. I am still looking at ways to reduce that, but I do not have a complete solution yet.
It is possible that higher-precision models such as BF16 could help reduce some of that degradation, but testing those properly is a bit difficult on my current hardware because of the VRAM and memory requirements.
So for now I am trying to improve it carefully rather than turning it into a huge all-in-one workflow and introducing even more bugs.
And seriously, thanks for testing it. Feedback from people who are actually using multiple chunks, doing re-rolls, and trying the newer reference features is especially helpful.
Audio voice reference would be a nice addition, as well as starting video if I already have a short clip.
it already supports audio voice ref. i only managed to get it working on 1 char only tho.
Audio Voice Reference is already supported in V3.8.
Continuum has an H3 Continuum Reference Audios node which supports up to 3 ordered reference audios, corresponding to <Audio 1>, <Audio 2>, and <Audio 3>.
For multiple characters, according to the official MiniMax H3 prompt guide, it is better to explicitly define the relationship between each subject and each voice reference, rather than only writing separate dialogue lines.
For example:
subject_definitions:
<Subject 1> is the woman (S1).
<Subject 2> is the man (S2).
<Audio 1> is the voice-timbre reference for <Subject 1> (S1).
<Audio 2> is the voice-timbre reference for <Subject 2> (S2).
[Shot 1]
<Subject 1> (S1) says in the voice referenced from <Audio 1>:
<d>[English] Where are you going?</d>
<Subject 2> (S2) replies in the voice referenced from <Audio 2>:
<d>[English] I'll be back soon.</d>
So if multiple voices are getting mixed up, I would try explicitly pairing <Subject N>, (S1/S2), and <Audio N> like this.
This follows the MiniMax H3 prompt-guide style, but I haven't personally validated multi-character voice separation with this exact setup yet, so I don't want to claim that it will always keep the voices perfectly separated.
For video, Continuum also already has H3 Continuum Load Video → Video Guide Frames. That lets an existing video act as persistent visual guidance.
However, this is not currently the same as a true Starting Video feature where an existing short clip becomes "Chunk 0", Continuum starts generating directly from the end of that clip, and then automatically assembles the original clip plus the generated continuation.
I agree that this would be a useful addition.
Hey guys i have a couple of questions, one how do I add more references since there only 3, two any prompts that will help in NOT making my character keep facing forward every chunk especially for RevCow and Dog style scenes, i get the twisted torso and ass cheek combo everytime, and finally can any explain the setting sin h3 continuum cause i tried asking chat gpt for help and still don't get stuff like run storage, match out vs max identity, and the differences im fixed, timeline, list and auto. So far timeline has been working, but of course I walways get the character facing forward everytime no matter how much i prompt not too. Thanks for the workflow!!
Thanks! A few answers:
For reference images, Continuum currently supports up to 3 Reference Images. They are used as persistent identity/appearance references throughout the generation. There isn't currently a way to add a 4th, 5th, etc. reference through the standard workflow. I may look at expanding this in the future.
For the character constantly turning toward the camera, I would first look at the prompt itself.
With H3, I generally recommend describing the pose and orientation you want, rather than repeatedly describing what you don't want.
For example, instead of:
do not face the camera, don't turn forward, don't show her back/butt
try something more like:
She remains in a left-facing three-quarter profile, looking toward the other character. Her shoulders, torso, and hips stay naturally aligned in the same direction. The camera remains at her side and slowly follows the action.
Or:
She turns her whole body toward the other character and continues the action in side profile. Her pelvis and shoulders rotate together naturally. The camera observes her from a rear three-quarter angle while following the movement.
This is especially important because words such as back, butt, hips, rear, etc. can still become strong concepts in the prompt even when they appear inside a negative instruction. If you keep mentioning the unwanted pose, H3 may still latch onto those concepts.
So I would describe the desired body orientation directly:
left-facing side profile
right-facing three-quarter profile
facing the other character
shoulders and hips aligned
whole body turns together
camera remains beside the character
rear three-quarter camera angle
For Continuum Timeline prompts, you can also carry that orientation into each chunk rather than allowing each chunk to re-establish the pose.
Example:
The same character and environment continue throughout the sequence. Natural continuous movement and consistent camera direction.
[Chunk 1]
integrated_multimodal_description: [Shot 1] She stands in a left-facing three-quarter profile, looking toward the other character. Her shoulders and hips are naturally aligned toward them. The camera stays beside her and slowly tracks the action.
[Chunk 2]
integrated_multimodal_description: [Shot 1] She continues from the same body orientation and movement. She turns further toward the other character with her whole body, keeping her shoulders, torso, and hips aligned. The camera continues tracking from the same side.
[Chunk 3]
integrated_multimodal_description: [Shot 1] She continues the interaction from the same three-quarter orientation. Her movement flows naturally from the previous chunk while the camera gradually moves around the pair without re-establishing a frontal pose.
That is generally closer to how I recommend writing H3/Continuum prompts: describe the current visible state, the next action, and the camera, rather than relying heavily on negatives.
As for the settings:
Fixed — the same Sequence Prompt is reused for every chunk.
List — each --- separated section is assigned to the next chunk.
Timeline — lets you explicitly write [Chunk 1], [Chunk 2], etc., or aligned time sections such as [0-5s], [5-10s]. For scenes where the action or pose should evolve across chunks, Timeline is usually the easiest to control.
Auto — Continuum detects whether your prompt is Fixed, List, or Timeline automatically.
If Timeline is already working well for you, I would stay with Timeline for this type of scene.
Run Storage saves completed chunks so Continuum can reuse compatible earlier chunks instead of generating everything again. For example, you can keep Chunk 1 and regenerate starting from Chunk 2.
For Reference Image Size:
Match Output is the practical/default choice.
Max Identity keeps more detail from the reference image and can improve identity/reference preservation, but it can also use more memory and be heavier to process.
Hope that clears it up a bit. The prompt orientation issue is probably the first thing I would experiment with in your case.
Something that worked well for me in this regard is writing the prompt in such a way that it actively tells h3 to discard and/or change the gaze of the person from the references.
For example in the summary you could add:
Standing studio poses and studio gazes from <Picture 1> are discarded.
And then in the detailed description do sth like this:
Change <Subject 1>’s gaze from the flat studio stare in <Picture 1> and from the other woman’s look in <Picture 2> to look down at the camera with a gleeful tease.
As ukr8b3g201 already said, h3 doesnt really understand negatives, so you want to format instructions in an active way (do this, that is happening...) rather than passive ones (do not, thay are not)
@Xarfai Yes, I think this is a good way to describe it.
Rather than simply telling H3 what not to copy from the reference, explicitly changing the reference state into the desired state seems much more useful — especially for gaze, pose and body orientation.
I would probably take it one step further and describe both sides of the transition: what should be discarded from the reference, and exactly what replaces it.
For example, instead of only “do not keep the studio gaze,” something like:
The flat studio gaze from <Picture 1> is replaced by a downward gaze toward the camera, with the head and body naturally following that orientation.
Combined with the other useful suggestion here — explicitly starting later chunks with something like Continuation of Chunk 1 — I think this could become a useful general prompting rule for Continuum.
Thanks for adding this. I’ll keep this approach in mind for the prompt examples/guidance.
Hello, and thank you for the workflow.
I’m not sure if this has been discussed before or if anyone else has encountered this problem.
I’m experiencing a loss of detail in the face and hands from the second segment onwards, caused by Spectrum, which I didn’t have with the old 3.7 workflow.
In any case, it’s great to be able to rerun the segments.
All the best!
Thanks for reporting this, and thank you for testing the workflow.
What you are describing is interesting, especially because the loss of facial and hand detail starts from the second segment onward.
Spectrum is not required by Continuum itself; it is an external optional accelerator, although the supplied V3.8 workflow currently has Spectrum included in the graph. So the first thing I would recommend is a simple A/B test with exactly the same seed and settings:
Spectrum ON
Spectrum OFF
If the detail loss only appears with Spectrum enabled, then I would treat it as a Spectrum/continuation interaction rather than a problem with the basic Continuum continuation path.
I have not tested Spectrum deeply enough across all multi-chunk/reference combinations yet, so I do not want to claim that its behavior is fully validated in V3.8. This is one of the areas I still need to investigate more carefully.
For the moment, if you prefer the image quality you were getting with V3.7, I would recommend running V3.8 with Spectrum disabled. You can still use the V3.8 Review Each Chunk / reroll features without Spectrum.
If possible, please let me know whether the same seed and settings keep the facial and hand detail when Spectrum is OFF. That comparison would be very useful for narrowing this down.
And thanks again — reports like this are especially helpful because quality degradation that only appears after Chunk 1 can be difficult to catch with short tests.
@ukr8b3g201 Actually, I’ve sorted this problem out by turning Spectrum off. I mainly create ‘Ref to video’ clips, each segment lasting 5 seconds. I hope I’ve answered your questions, as I’m not very confident with English or technical terms.
That said, I do have a question about LORAs: if I want to activate a specific LORA on a segment (the third one, for example), it breaks the sequence and I have to start again from the first segment. Is there a way to get round this problem?
video ref question: i have a 10sec/240fr ref video. if i create 1 chunk, 10sec clip, result video tracks pretty close to ref. if i create 2 chunk, 5sec clips, 1st chunk matches ref video, 2nd chunk continuity and transition is good but does not match ref video. even if i stop and re-roll chunk 2, it never gets close to ref video. i'm using fl2va model. tried ref2va, but same behavior. is there a setting to indicate to sampler to use next 'chunk' of ref frames in successive chunks?
Thanks — your observation is correct.
I checked the current Continuum implementation again. At the moment, there is no separate Timeline Video input in the public V3.8 workflow. The available video input is Video Guide Frames.
Currently, Video Guide Frames does not automatically advance through a longer reference video according to the Continuum chunk position.
So in your example:
10-second / 240-frame reference
2 × 5-second Continuum chunks
Chunk 1 can follow the first part of the reference closely, but Chunk 2 is not automatically given the second 5-second section of that reference.
There is currently no sampler setting that changes this behavior. Switching between FL2VA and Ref2VA will not solve it either, because the limitation is in how Continuum currently supplies the video guide across chunks.
Your test is therefore useful — per-chunk timeline slicing of Video Guide Frames is something that needs to be added to Continuum.
Thanks for catching this.
@ukr8b3g201 is there a 'practical' limit for how long 1st chunk can be/ how many 'video guide frames' can be used in that 1st chunk?
Nice work i like this workflow but could you add 2 more slots for reference images please :)
Got any prompt examples or llm instructions you use for prompts? no matter what i try pretty much every chunk tries to do the whole video. so the end is just a video repeating itself 3 times. can't even get anything basic done.
[0-5s] the girl walks forward on a street
[5s-10s] the scene switches to the girl walking in a parking garage, she is looking around,she sees a red motorcycle and starts walking towards it
[10s-15s] the scene changes to the girl riding the motorcycle through a street, she stops at a red light.
what am i doing wrong with that prompt? chunk 1 tries to do all the cuts every time.
EDIT: well I'm a moron i figured out the mistake. I'll leave the comment in case someone also has issues with this.
i was doing the prompt right except for the fact that you're supposed to put the prompt below the parenthesis not next to it.
[0-5s]
the girl walks forward on a street
[5s-10s]
the scene switches to the girl walking in a parking garage, she is looking around,she sees a red motorcycle and starts walking towards it
[10s-15s]
the scene changes to the girl riding the motorcycle through a street, she stops at a red light.
This is what actually works lol. I hate reading
The Tag with the seconds has to be in its own line.
My System Prompt is:
You are an expert prompt engineer for MiniMax H3 Ref2VA (the open-source 33B reference-to-video-audio model, 768p, 24 fps, 3-15 s, 32 kHz stereo audio, ≤9 images / ≤3 video clips / ≤3 audio clips, ≤12 reference files total).
Your job: turn a user's scenario + reference material into a structurally compatible H3 prompt that the model can condition on. The prompt drives BOTH video and audio — the audio sections are as important as the visuals.
---
## INPUT HANDLING & DECISION MAKING
1. Asset Mapping: Map each reference asset to a unique label: images → <Picture N>, videos → <Video N>, audio → <Audio N>. Number each category independently.
2. Duration Planning: If the user did not state the target duration, default to ~10 seconds. Ensure all cut timestamps fit strictly inside the total duration.
3. Mode Selection:
- Frame-anchor mode: Use this if a reference image is explicitly meant to be the video's first frame, last frame, or a hard structural keyframe.
- Full-reference mode: Use this if references only serve to guide character identity, scene appearance, style, or generalized actions.
4. Language Enforcement & Translation: The user may provide instructions, descriptions, or scenarios in languages other than English (e.g., Traditional Chinese, Simplified Chinese). You must automatically translate all user concepts, visual scenes, and soundscape layouts into English. The final structured output must be 100% in English across all sections.
5. Clarification: If the request is genuinely underspecified (no scenario, no references), ask one concise clarifying question before generating. Otherwise, rely on creative choices.
---
## OUTPUT STRUCTURE MODES
Select one of the two modes based on input criteria. Output ONLY the raw formatted prompt matching the section headers below. Do not include preambles, explanations, or introductory text.
### Mode A: Full-Reference Mode (6 sections, strict order)
#### subject_definitions:
Define each reusable item (person, object, scene, style, action).
Rule: Only isolate <Picture N> on its own line if it serves as a concrete frame anchor; otherwise, cite it nested inside a <Subject N> definition. All definitions must be written in English.
Format:
<Subject 1> is ... [Detailed description in English]
<Picture N> is ... [Only if used as a strict composition anchor]
#### summary:
Must start with task type prefixes enclosed in brackets (e.g., [reference generation + audio reference]). Write one short paragraph in English summarizing the target video narrative and its relationship with the reference assets.
Task Types: keyframe completion · reference generation · video editing · video continuation · audio reuse · audio reference. Combine with + if multiple apply.
#### retention_analysis:
List all relevant labels line-by-line, stating which shots they appear in and defining their exact retention/preservation strength. Use the fixed English markers, followed by a brief clarification description written entirely in English.
Visible Markers: fully_preserved · partially_preserved · attribute_transfer · weak_reference
Audio Markers: fully_copy · partially_copy · reference · weak_reference
Format:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - [Brief descriptive detail in English]
#### detailed_description:
The very first line must be one or two English sentences establishing the overarching visual style, lighting, and color palette. This is immediately followed by a thorough shot-by-shot breakdown.
Word Count Constraint: For generation tasks, this section must contain 350–500 English words in total distributed across shots to maximize temporal density.
#### overall_soundscape:
Write 1–4 sentences in English detailing the ambient environment and physical sounds (wind, footsteps, mechanical clicks, fluid movement) playing across the entire video. Do not repeat dialogue text. Use N/A only for absolute silence.
#### non_diegetic_music:
Write 1–3 sentences in English specifying audience-only background music, instrumentation, tempo, and dynamics. Avoid abstract emotional adjectives. Use N/A if there is no score.
---
### Mode B: Frame-Anchor Mode (Simplified Keyframe Entry)
When exactly one image is explicitly designated as the video's absolute first frame (0.00 seconds), output these exact simplified blocks in English instead:
```text
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description:
[Shot 1] (Anchor the shot on the image's composition, then develop forward sequentially in English: anchor state → action onset → continuous development → result)
overall_soundscape:
[English description of ambient and physical sound layers]
non_diegetic_music:
[English specification of score layers or N/A]
```
---
## SHOT, CAMERA, & DIALOGUE CORE RULES
1. Shot Timeline Syntax:
- [Shot 1] must never feature a timestamp.
- Subsequent cuts must be explicitly marked with strictly increasing timestamps matching this format: [Shot 2] At 00:03.500, the camera cuts to ....
2. Standardized Camera Motion: Camera descriptions must combine motion type + amplitude + speed in natural English (e.g., "The camera pushes in with small amplitude at slow speed").
Authorized Vocabulary: Zoom In/Out, Push In/Pull Out, Pan, Truck, Tilt, Pedestal, Arc, Tracking, Static, Shake Slightly/Strongly, POV, Roll.
3. Speaker Tracking & Dialogue Tagging:
- Assign permanent tracking IDs to characters: (S1), (S2), or (S1, S2) for unison.
- Dialogue must be enclosed inside <d> tags featuring a language attribute.
- Language Rule: If the user provides dialogue in a native language (e.g., Chinese) and explicitly wants the character to speak it, preserve that exact native script inside the <d> tag. Otherwise, translate the dialogue to match the specified tag language.
- Example: the young woman (S1) says, <d>[English] Hands' up, pervert! </d>
4. Language & Style Constraints: Write exclusively in English for all sections, titles, and descriptions. Use the present tense. Be purely descriptive, objective, and non-instructional. Do NOT use meta-phrases like "show me," "create a video of," or "the camera displays." Describe the scene directly as if it is already unfolding.
5. Specialized Audio-Editing Triggers:
- Voiceovers must use this exact phrasing: says in an off-screen voiceover and explicitly note that the character's lips remain closed.
- Use <scenetrans> inside the dialogue block if speech smoothly crosses over a visual camera cut.
- Use <cutoff> at the very end of the line if dialogue is abruptly cut off by the video ending.
6. On-Screen Text: If physical text appears written on an object or sign within the environment, display it verbatim inside standard English double quotes.
---
## QUALITY & OUTPUT CONTROL
- Granular Density: Every individual shot block must concretely account for subject appearance, exact framing position, active physical progression, structural state changes, environmental shifts, lighting profiles, and the precise locations where referenced assets are visually rendered. Never compromise a shot breakdown into a brief plot summary.
- Section Headers: All section header titles (e.g., subject_definitions:, detailed_description:) must be written entirely in lowercase.
- Absolute Output Restriction: Do not include conversational remarks, meta-commentary, notes, markdown code block wrappers around the entire response, or conversational framing. Output begins immediately at the first required structural block header.
______________________________
Then I have some additional guidelines, the most important are:
Structure
Every video is four fully independent 15-second segments: [0-15s], [15-30s], [30-45s], [45-60s].
Each segment is a complete self-contained prompt. No relative or cross-segment language (continues, maintains, further, still, again as a callback).
Every segment fully restates subjects, setting, actions, fluids, expressions, camera, and sound as if it stands alone.
Section headers stay lowercase (subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, non_diegetic_music:).
Output starts at the first header. No conversational wrapper around the prompt itself.
Subjects and language
Always use explicit tags: <Subject 1>, <Subject 2>, <Picture N>, <Audio N>.
No pronouns (she, her, his, him, they) and no generic possessives (the cock, the pussy). Form: <Subject 1>’s cock, <Subject 1>’s mouth.
Gaze and facing are stated explicitly, and stated as different from the reference stares when the scene requires it.
On-screen text, if any, goes in English double quotes.
References
Images map independently: <Picture 1>, <Picture 2>, <Picture 3>.
Each picture has a job: body lock, face lock, pose/camera/setting lock, prop lock.
Unwanted parts are discarded in active language (X is discarded), not as “do not / don’t / do not copy.”
Pose-only or camera-only refs discard the original people, clothes, and rooms.
Later overrides win where specified (example: <Picture 2> overrides <Picture 1> for face and teeth).
Audio refs keep speaker identity, pitch, tone, and speech patterns. The words of the clip are discarded unless the user asks for those lines.
Camera and picture
Prefer exact camera of the scene reference when one exists.
Static POV stays static: no zoom, push-in, or orbit unless the user asks for movement.
When movement is requested, describe orbits, pans, and cranes that still show bodies, legs, and faces.
@Xarfai Thanks for posting the update — and yes, the Timeline tag has to be on its own line.
So this:
[0-5s]
the girl walks forward on a street
[5-10s]
the scene switches to the girl walking in a parking garage...
[10-15s]
the scene changes to the girl riding the motorcycle...
is the correct Continuum Timeline format.
One additional resource you may find useful: I’ve been gradually updating a small prompt-writing skill specifically for MiniMax H3 + H3 Continuum:
https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum/blob/main/H3-Continuum-Skill-v1.zip
It includes Continuum-specific Timeline/Chunk formatting rules, continuation wording, reference handling, audio guidance, and some example prompt structures. I’ve also just updated it to make the “Timeline header must be on its own line” rule more explicit, since this is clearly an easy mistake to make.
It’s still a work in progress and not intended as a complete or definitive H3 prompting guide yet, but it may be useful as a reference or as an instruction skill for an LLM when writing Continuum prompts.
For example, if you give an LLM a rough scenario and your actual Continuum settings such as 3 chunks × 5 seconds, the skill is intended to produce the prompt in the correct Continuum structure rather than treating the whole 15-second story as one prompt.
I get "RuntimeError: aimdo memory compile error" no matter what I try with 3.8. Only 3.7 works for me. Really not sure what to do here.
EDIT: I installed a fresh comfyui and was able to get it working.
On a 4070 super it takes me 10 or 15 times as long to generate a 5 second video on 3.8 as it does a 10 second video on 3.7. So I think I will return to 3.7
Thanks for testing this, and the fresh ComfyUI result is useful information.
The fact that the aimdo memory compile error disappeared with a fresh ComfyUI install suggests that the original error may have involved the environment or another custom node rather than Continuum itself, although I can't confirm the exact cause yet.
The 10–15× slowdown is definitely not what I would expect from V3.8 itself, though. Before I classify that as a V3.8 performance regression, I would want to compare the two workflows under the same conditions:
same H3 model
same resolution
same sampler / sigmas / steps
same LoRA setup
same Spectrum / Sage / Comfy Kitchen configuration
same generation mode
One important difference is that the supplied V3.8 workflow is set up around the Spectrum path, with the Turbo LoRA disabled by default. So if your V3.7 workflow was using a faster Turbo or different acceleration path, the timing difference may not be coming from Continuum itself.
If you still have both workflows, even just screenshots of the model, LoRA, sampler, steps, resolution and acceleration settings would be enough for me to compare them. If the settings are equivalent and V3.8 is still anywhere near 10–15× slower, then I would consider that a real performance issue worth investigating.
In the meantime, using 3.7 makes sense if it is currently much faster on your 4070 Super.
@ukr8b3g201 Not exactly related and I dont have exact times from 3.7 anymore, but the initialization time of 3.8 is definitely slower. Without spectrum and not on a fresh load, it currently taks 3-3.5 minutes before the first steps of the sampler are generating. For 5s videos on my setup, that is about half the time of the whole generation for me. As i said, I dont have exact numbers, but fairly certain that initialization process was not more than 60s before.
@ukr8b3g201 I have the same everything. And yes, as @Xarfai said the initializatioin time can take up to 5 minutes. Then I hit 5% progress. 20 minutes later I still haven't hit 10%. At this point I gave up. Since in 20 minutes I could have already generated a 15 second video on 3.7. Really makes me wish I had a 5090. But I'm not paying 6000 dollars for that lmao. Honestly I wouldn't be too worried about catering to my problems. Most people seem to do just fine with 3.8
@oldthrashbar Thanks for the additional information. This is useful because it confirms that the slowdown is not limited to the initialization stage.
What you are describing appears to be two separate issues:
initialization taking up to around 5 minutes before sampling really starts
sampling itself then progressing far more slowly than it did in V3.7
Since V3.7 was already usable on the same hardware, I don’t think this should simply be treated as a case of needing a faster GPU. A 5090 should not be required just to recover the performance you were already getting with V3.7.
I’m comparing the V3.7 and V3.8 code paths now, particularly the additional work around initialization, Run Storage / Review handling, reference-state validation, and anything that could affect the sampling path.
You don’t need to spend more time troubleshooting it for me. Between your report and Xarfai’s report, there is enough evidence for me to investigate this as a possible V3.8 regression or a slow path that only appears in certain environments.
For now, going back to V3.7 is completely reasonable if it gives you much better performance.
Thanks again for reporting the timings. They are useful even if most users are not seeing the same behavior.
3.8 works great thank you!
Is there a clean way to disable audio. No prompting. I want to disable the audio vae.
Thanks! Glad V3.8 is working well for you.
At the moment there isn’t a clean “Video Only / Disable Audio” switch in Continuum.
MiniMax H3 generates video and audio latents together, so turning Audio Continuity off only disables carrying the previous chunk’s audio context — it does not stop H3 from generating the audio latent.
The actual Audio VAE decode happens separately in the workflow through ComfyUI Core. So if you simply do not need audio in the final file, you can bypass/remove the VAEDecodeAudio side and use a video-only output path. However, the supplied Continuum Finalize workflow currently expects decoded audio as part of its normal AV assembly path, so this is not a one-click option yet.
I think a proper Video Only option that skips the Audio VAE decode/final audio path would be useful. I’ll consider adding that as a clean workflow option rather than requiring users to manually rewire the graph.
Also, if you are not using Driving Audio or Reference Audio, you do not need an Audio VAE for those conditioning inputs.