keys-2k MiniMax-H3 — Parallel processing for Two DGX Sparks
One-command pull · pinned quality stack · master-K0 dual-Spark parallel · ESRGAN “2K” · Spectrum audio fix · Contex-Loop / MultiRef · optional realism LoRA
| Docker (arm64, published) | ghcr.io/drowzeys/keys-2k-minimax-h3-parallel-two-dgx-sparks:0.31.1-pp20260811 (~16GB, Comfy+nodes; mount models) |
| HF package | this repo |
| Power Pack | keys-DGX-Sparkticus-Ultimate-Power-Pack-Unleashed |
| Dual-serve foundation | tonyd2wild/ds4-h3-video-gen-factory (Tony) |
Dual-serve DS4 + dual H3 on two DGX Sparks was made possible by Tony (tonyd2wild).
This package is the keyspark quality specialization (ablit co-tenant profile + sealed bf16 finals with stock int8-convrot TE — heretic retired — + farm-mode dual parallel + 2K workflows).
Why this exists
Agents and users often git clone old Spectrum / Motion Context / Comfy builds and lose:
- Spectrum v0.2.1 audio fix (
offline_smoothing_replay=true) - Contex-Loop / MultiRef seam fixes
- Legal native sizes for upscale-to-2K
- Dual-node master-K0 parallel spans
This package pins the full integrated stack in a Docker image + manifest so the default path is the upgraded one.
One-command pull
curl -fsSL https://huggingface.co/drowzeys/keys-2k-MiniMax-H3-Parallel-Two-DGX-Sparks/raw/main/scripts/one_command_pull.sh | bash
Or:
git clone https://huggingface.co/drowzeys/keys-2k-MiniMax-H3-Parallel-Two-DGX-Sparks
cd keys-2k-MiniMax-H3-Parallel-Two-DGX-Sparks
bash scripts/one_command_pull.sh
Pull Docker only:
docker pull ghcr.io/drowzeys/keys-2k-minimax-h3-parallel-two-dgx-sparks:0.31.1-pp20260811
# also tagged :latest
What’s in the image (pinned)
| Component | Version / pin |
|---|---|
| ComfyUI | 0.31.1 |
| Spectrum MiniMax H3 | 0.2.1 (audio fix defaults) |
| Contex-Loop | 0.3.8 |
| Motion Context + MultiRef | pinned SHAs in manifest/H3_STACK_MANIFEST.json |
| Sol-attn Blackwell + Triton SolAttn | pinned |
| KJNodes (Sage) | 1.4.9 · sageattention==1.0.6 |
| NKD Preview Tools | 3.3.0 |
| Launch flags | --disable-pinned-memory --reserve-vram 48 --vram-headroom 10 |
Not for quality: Turbo / few-step LoRA graphs (present optional; leave off).
See manifest/H3_STACK_MANIFEST.json and scripts/verify_h3_stack.sh.
2K path (open weights)
Open MiniMax-H3 does not native 2K in one pass. This package’s production path:
Native denoise 704 × 1280 (multiple of 32 — 720×1280 fails)
↓ RealESRGAN ×2 (inline on spans)
Delivery 2K ~1408 × 2560
Parallelism: master-K0 keyframes + parallel FL2VA spans on two Sparks
(h3-spans.py --kf-mode master-parallel --upscale).
Workflow package: workflows/anime_2k_bench/
Lab sample run 0811_012837: ~18.8 min wall for ~4.8 s @ 704×1280 dual-node (co-tenant DS4).
Dual DGX Spark — parallel H3
export HEAD=10.100.10.1 WORKER=10.100.10.5 # your fabric IPs
export MODELS_DIR=$HOME/keys-2k-minimax-h3-parallel/models
# weights once
bash scripts/fetch_weights.sh
# or: rsync -a /path/to/lab/h3-cotenancy/ComfyUI/models/ $MODELS_DIR/
# H3 on both nodes
bash scripts/run_dual_h3.sh
# quality 2K parallel bench
bash scripts/run_anime_2k_bench.sh
# optional realism LoRA (file in models/loras/):
REALISM=1 bash scripts/run_anime_2k_bench.sh
Single node smoke:
docker run --rm --gpus all --network host \
-v $MODELS_DIR:/opt/powerpack-h3/ComfyUI/models \
-v $HOME/Videos/h3-out:/opt/powerpack-h3/ComfyUI/output \
ghcr.io/drowzeys/keys-2k-minimax-h3-parallel-two-dgx-sparks:0.31.1-pp20260811
# UI/API: http://<host>:8188
Compose: compose/docker-compose.yml
Co-tenancy with DeepSeek-V4-Flash (Tony’s rule)
If running DS4 + H3 on the same two Sparks:
- DS4 first until
/v1/modelshealthy - H3 second (this image)
- Power Pack util 0.76 / 888k (or ≤ 0.85 hard cap)
- Teardown: H3 → DS4
- One heavy video job per Spark (
concurrency=2= two boxes)
Full dual-boot recipe: Power Pack deploy/keyspark/bringup.sh.
A/V glitch mitigations (baked into graphs)
| Fix | Setting |
|---|---|
| Spectrum speech stutter | offline_smoothing_replay=true, audio_blend_weight=0 |
| Multishot seams | Motion Context / Contex-Loop / MultiRef |
| Long-gen hallucination | master-K0 short parallel spans (not one long gen) |
Weights (not fully redistributed here)
Place under models/ (bind-mount into the container):
| File | Role |
|---|---|
diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors |
FL2VA DiT — FINALS (sealed) |
diffusion_models/minimax_h3_ref2va_pruned_bf16.safetensors |
ref2va — FINALS (sealed) |
diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors |
FL2VA DiT — draft tier only |
diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors |
ref2va — draft tier only |
text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
STOCK int8-convrot TE (sealed; heretic retired) |
vae/minimax_h3_video_vae_fp16.safetensors |
video VAE |
vae/minimax_h3_audio_vae_fp32.safetensors |
audio VAE |
upscale_models/RealESRGAN_x2plus.pth |
2K upscale |
loras/h3-realism-people-t2v-i2v-r2v.safetensors |
optional realism |
Helpers: scripts/fetch_weights.sh
Upstream: Comfy-Org/MiniMax-H3, Kijai/MiniMax-H3_comfy.
Verify (agents must pass)
bash scripts/verify_h3_stack.sh 10.100.10.1
bash scripts/verify_h3_stack.sh 10.100.10.5
Fails if Spectrum audio-fix default is wrong or core nodes are missing.
Credits
| Dual-serve co-tenancy | Tony / tonyd2wild — ds4-h3-video-gen-factory |
| Spectrum v0.2.1 audio fix | xmarre / ComfyUI-Spectrum-MiniMax-H3 |
| Motion Context | NikoDemon80 |
| MultiRef | seitanism |
| Contex-Loop | ethanfel |
| NKD Preview | Nekodificador |
| Sol / Sage / heretic | NVIDIA Sol-engine community · KJNodes · sakamakismile et al. |
| Packaging | keyspark / @drowzeys Power Pack |
License
- Package scripts/docs: see Power Pack repo
- Model weights: their respective licenses (not all redistributed here)
- ComfyUI and custom nodes: upstream licenses (GPL where applicable)