ComfyUI VELA H3

ComfyUI VELA H3

Version 1.0.0 · developed and maintained by Speach1sdef178

VELA is a trajectory-aware production execution optimization layer for VDN-H3 on MiniMax H3 in ComfyUI. It is the separate continuation of the optimization work started in ComfyUI-VDN-H3-24GB. VELA keeps the VDN checkpoint format and model behavior while selectively changing how expensive attention paths are executed.

VELA is not a distilled MiniMax model and does not ship model weights. The VDN checkpoint is required separately.

Why VELA exists

The project began with a simple question: which parts of the VDN + MiniMax H3 runtime can be made faster on a 24 GB Ampere GPU without treating every transformer block, attention branch or diffusion step as equally tolerant to approximation?

The resulting production policy is deliberately heterogeneous. Faster approximate backends are used where validation showed they were acceptable, while trajectory-sensitive regions are returned to exact execution. The final release is therefore the result of profiling, numerical A/B testing and block/timestep sensitivity mapping rather than a global backend replacement.

VELA 1.0.0 production policy

• QKV projection: native comfy-kitchen INT8 / ConvRot path.
• GROUP attention: Stock Sage path (C40).
• GLOBAL attention: Sage normally.
• Critical GLOBAL island: gate label 2, transformer blocks 40–49 use exact direct-cuDNN SDPA.
• ANCHOR attention: exact.
• Anchor-column optimization: extra anchor columns are removed only in blocks 0 and 36.
• Research/deep profilers are disabled in production.
• The validated VELA configuration is internalized; no external VDN_H3_* BAT flags are required.
• Experimental Triton QKV is not enabled in VELA 1.0.0.

The internal research checkpoint that became this public release was VELA V3.28. The public version starts at 1.0.0 and follows normal semantic versioning.

Benchmarks

Matched end-to-end measurements on the project test system:

| Test | VDN 1.1.0 | VELA 1.0.0 | Change | |---|---:|---:|---:| | 0.4 MP / 5 sec | 1:14 | 1:01 | 17.6% faster | | 0.8 MP / 5 sec | 2:53 | 2:13 | 23.1% faster | | 0.8 MP / 7 sec | 3:10 | 3:02 | 4.2% faster | | 0.8 MP / 8 sec | 3:31 | 3:23 | 3.8% faster | | 0.8 MP / 9 sec | 4:07 | 3:57 | 4.0% faster | | 0.8 MP / 10 sec | 4:12 | 4:17 | 2.0% slower |

Additional VELA scaling validation:

• 1.0 MP / 5 sec / 8 steps = 2:46 total sampler time, 20.80 s/it. There is no matched VDN 1.1.0 1.0 MP measurement, so no speedup claim is made for this point.
• 1344×768 / 16:9 / 1.032 MP / 8 sec / 8 steps = 4:40.
• 768×1344 / 9:16 / 1.032 MP / 5 sec / 8 steps = 3:03 on the first portrait run and 2:52 on a subsequent warm run.

These results are intentionally reported as a full table rather than a single “up to” number. VELA's benefit is workload-dependent; the relative cost of the optimized attention paths changes with spatial and temporal sequence geometry.

Visual quality comparison — Raw 20 steps vs VDN + VELA 8 steps

The comparisons below are intended for direct visual inspection of output quality, not as a pure VELA-only speed benchmark. The Raw MiniMax H3 examples use 20 steps, while the VDN + VELA examples use 8 steps, so the full timing difference must not be attributed to VELA alone.

The purpose is to show how closely the practical 8-step VDN + VELA output can approach the visual result of a much slower 20-step raw generation under the same output geometry.

Anime — 768×1344 · 9:16 · 1.032 MP · 5 sec

Raw MiniMax H3: 20 steps · 8:07
VDN + VELA 1.0.0: 8 steps · 3:03

https://github.com/user-attachments/assets/5ec53d7e-bb50-4d54-8717-e827bd85d3a2

Realistic — 768×1344 · 9:16 · 1.032 MP · 5 sec

Raw MiniMax H3: 20 steps · 7:50
VDN + VELA 1.0.0: 8 steps · 2:52 (warm run)

https://github.com/user-attachments/assets/89e198f4-63e0-48de-b7ab-36da21e8ca48

These examples are presented as qualitative A/B comparisons. No claim of pixel-identical or universally identical quality is made; the videos are provided so the visual difference can be judged directly.

See BENCHMARKS.md for methodology and RESEARCH.md for the complete research history.

Tested stack

• NVIDIA RTX 3090 Ti 24 GB
• Windows
• Python 3.13.15
• PyTorch 2.13.0 + CUDA 13.0
• ComfyUI 0.33.0 (v0.33.0-36-g76135e55)
• comfy-kitchen 0.2.31
• comfy-aimdo 0.4.13
• MiniMax H3 INT8 ConvRot + VDN

This is the validated stack, not a guarantee that every future ComfyUI/PyTorch/comfy-kitchen combination is numerically or performance-equivalent.

Installation

ComfyUI Manager / Registry

After the package is published to the Comfy Registry, install ComfyUI VELA H3 through ComfyUI Manager and restart ComfyUI.

Manual

Clone or extract the repository to:

<ComfyUI>/custom_nodes/ComfyUI-VELA-H3/

The directory should directly contain __init__.py, vdn_h3_24gb/, pyproject.toml and this README.

VDN checkpoint

VELA uses the same separately distributed VDN stage format as the preceding VDN-H3 project. Place the complete stage at:

<ComfyUI>/models/vdn/stage-dmd-step-250-int8_convrot_comfyui/
├─ model_spec.json
├─ linear_branch/
│  └─ model_int8_convrot_comfyui.safetensors
└─ adapters/
   ├─ default/
   │  ├─ adapter_config.json
   │  └─ adapter_model.safetensors
   └─ turbo/
      ├─ adapter_config.json
      └─ adapter_model.safetensors

Checkpoint source: speach1sdef178/VDN-H3-INT8-ConvRot-ComfyUI. Model weights are not included in this repository.

Core-patch-free integration

VELA v1.0.0 does not modify comfy/ldm/minimax/model.py or any other ComfyUI core source file. The older VDN research line included an experimental block-loop hook for LongCache; that hook is deliberately not part of the VELA 1.0.0 production path. LongCache remains research-only and disabled in this release. This avoids a fragile on-disk dependency on MiniMax-H3 core implementation details and improves compatibility with upstream ComfyUI features that use the standard model patch system.

Expected release markers

At startup:

[VELA v1.0.0] production config active; external VDN_H3_* BAT flags are not required

During generation:

[VELA v1.0.0] ACTIVE production path | GLOBAL gate2 blocks=40-49 exact direct-cuDNN | other GLOBAL Sage | ANCHOR exact | GROUP Stock Sage | anchor-column removal blocks=0,36

These markers are useful when validating that the release package, rather than an older experimental VELA folder, is active.

Project lineage

VELA is published as a separate project because its purpose is broader than the original 24 GB VDN integration, but it is explicitly a continuation of that work:

OpenVDN / VideoDeltaNet → ComfyUI VDN-H3 port → Speach1sdef178/ComfyUI-VDN-H3-24GB → ComfyUI VELA H3

See NOTICE and LICENSE for attribution.

Research documentation

• RESEARCH.md — full technical research article and experimental history.
• BENCHMARKS.md — publication benchmark table and methodology.
• CHANGELOG.md — public release history.

License

Apache License 2.0. Existing upstream notices are preserved. VELA modifications are Copyright 2026 Speach1sdef178. MiniMax-H3 model weights and separately distributed checkpoints may be governed by their own licenses; review those terms before redistribution or commercial use.