MiniMax H3 Hi-Res Regenerate | DEAS | ComfyUI

作者：Darksidewalker

✨ Open-H3-Regenerate for DEAS & ComfyUI 🔑 Native H3 Keys • 🗜️ SVD 0.999 Re-Ranked The Open-H3-Regenerate-2K refiner LoRA from NVlabs, converted so ComfyUI actually loads it. It takes a 768P MiniMax H3 draft and regenerates it at 2560×1440 (2K) in just three sampling steps!

• Support for 1080P and 4K outputs as well. ⚠️ Why this conversion exists The original upstream file uses Diffusers key names. ComfyUI's LoRA loader does not recognize them and skips every single key silently without throwing an error — meaning the original file literally does nothing in ComfyUI. This release maps everything to native H3 keys (fused q/k/v, fc1 halves in ComfyUI order) and loads cleanly with zero unloaded keys. 🌟 HIGHLIGHTS & PROS
• 🎨 Real Detail, Not Blurred Sharpening: The draft is upscaled in latent space and genuinely re-sampled, generating crisp new textures for fur, snow, foliage, and fine needles. Standard latent upscaling with a low-denoise pass stays noticeably soft by comparison (see showcase crops).

• 🗜️ Compact File, Identical Quality: Compressed via SVD re-ranking at 99.9% energy — shrinking the file from 8.77 GiB down to 2.35 GiB. In same-seed testing, both versions measure 39.7 dB PSNR apart and look indistinguishable to the naked eye.

• ⚡ Blazingly Fast: Runs in 3 steps, CFG 1, and requires no negative prompt.

• 🔊 Audio Stays Untouched: Your draft's original audio track passes through completely unchanged.

• 🎬 Full Workflow Compatibility: Continuity and Director features continue working seamlessly when paired with the DaSiWa custom node.

📦 WHAT'S INSIDE This is a single merged LoRA matching the exact upstream NVlabs recipe:

• Open-H3 Refiner (Rank 384)

• LightX2V FL2VA Turbo v1.2

• LMS Sharp v1.0

The keys were converted to ComfyUI's native H3 format, and the combined weights were SVD re-ranked (0.999 energy, 3.15% median relative error). It is a faithful, key-converted derivative with no additional training or altered behavior. 🛠️ REQUIREMENTS

• ComfyUI with native MiniMax H3 support

• Supported out of the box in DEAS - Dark Engine AI Studio (darksidewalker/DEAS), no manual wiring needed there. The node pins the recipe above, handles 4K via an intermediate 2K upscale plus tiling, and patches only a model clone.

• DaSiWa Custom Nodes (v0.5.49 or newer) - specifically the "MiniMaxH3 • Regenerate 2K" node:

• https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes

• MiniMax H3 Model - trained and tested on the FL2VA / Ref2VA and custom fine-tunes

• H3 Latent Upscaler - place minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors inside models/latent_upscale_models/ (from LBH-123-AI; the 2D variant also works)

• This LoRA — place in models/loras/

🔌 COMFYUI & DEAS SETUP 🚀 Option A One-Click Pipeline in DEAS (Dark Engine AI Studio): Full native support is built directly into DEAS:

• https://civitai.red/models/3028621/deas-dark-engine-ai-studio or

• ttps://github.com/darksidewalker/DEAS

DEAS automatically handles the backend, model downloads, latent upscaler routing, and the Regenerate 2K pass out of the box - no manual node wiring required. 🔧 Option B Manual ComfyUI Wiring:

1. Generate your draft as usual (1344×768 is the reference resolution; using Turbo for the draft pass is completely fine).
2. Add "MiniMaxH3 • Regenerate 2K" directly after your draft sampler:
3. model: Output from your plain H3 model loader (without Turbo or style LoRAs).

• conditioning: The same positive conditioning used for the draft.

• latent: The draft's packed video+audio latent.

• lora_name: Select this file. • target: Choose 1080P, 2K, or 4K.

• Optional inputs: Connect director_guide + vae (start/end frames are automatically re-encoded at the target resolution) and continuity_context (for Continuity continuations).

1. Decode the output latent using your standard VAE decode setup. ⚙️ Under the Hood The node automatically enforces the published recipe (Euler, CFG 1, sigmas 0.909375, 0.857143, 0.666667, 0, LoRA strength 1.0). It also handles 4K via an intermediate 2K upscale, spatial tiling, and VRAM management. Because the LoRA is applied to a cloned model instance, your upstream generation chain remains unpatched.
2. Resolution Grid: 1080P computes at 1920×1088 and 4K at 3840×2176 (matching the reference implementation — crop after decoding if you need exact 1080 or 2160 vertical pixels). Drafts near 16:9 snap to the exact 16:9 grid; other aspect ratios preserve their ratio scaled to the target's short side.

🧪 BENCHMARKS & TEST NOTES

• Hardware & Model: RTX 5090 (32 GB), FL2VA pruned bf16 base, 124 frames, 1344×768 → 2560×1440

• Spatial Tiling ON: ~26–27 min (draft + regenerate)

• Spatial Tiling OFF: ~19 min (slightly sharper output, no OOM on 32 GB VRAM)

• Bottleneck & Memory: The model exceeds 32 GB VRAM and streams from system RAM (which dominates total runtime); peak system RAM usage was 21–25 GB.

• Continuity Test: Tested a Continuity continuation ("new segment only") with Regenerate 2K — the join stays coherent, audio remains untouched, and no extra flicker is introduced.

⚠️ LIMITATIONS & KNOWN BEHAVIOR

• H3 Specific: Trained specifically on MiniMax H3 drafts; drafts from other video architectures are untested.

• Spatial Tiling Trade-off: Tiling can slightly reduce global coherence compared to a full-canvas pass. Disable tiling if your VRAM headroom permits.

• Seed Sensitivity: Same-seed outputs are not deterministic across different settings — sampling in 3 steps from high noise naturally amplifies minor numerical differences (expected behavior noted by NVlabs).

📜 CREDITS & LICENSES

• Open-H3-Regenerate-2K Refiner & Merged LoRA: NVlabs / Efficient-Large-Model https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K

• Code: https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K

• MiniMax H3 Base Model: MiniMax — Governed by the MiniMax H3 Community License (including territory, usage, and redistribution restrictions)