Open-H3-Regenerate-2K: MiniMax H3 - 20261029 - For Patreon Supporters (Patreon)
Videos
-
773650503.mp4
Downloads
Content
Hey everyone!
If you watched the video, you saw the 480p-to-2K tests and which method I preferred. (https://www.patreon.com/aifuturetech/posts/173747267)
This post has both workflows as I ran them, the compare version with method 1 and method 2 side by side, plus every setting that mattered and the ones that ship correct.
The exclusive file here is the 2-method compare workflow. The free two-stage file from the public post is attached too, so everything lives in one place.
----------------------------------------
WHAT TO DOWNLOAD (MODEL FILES)
----------------------------------------
Base and encoders (what the files ship with):
- Minimax-h3_Singularity_ref2va_v2_int8.safetensors -> models/diffusion_models/minimax-h3 (any H3 base or fine-tune works; the workflow note also lists the smhfacct hybrid fl2va/ref2va int8 and the Comfy-Org fl2va pruned int8)
- qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors -> models/text_encoders/minimax-h3 (CLIPLoader type minimax)
- minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors -> models/vae/minimax-h3
Stage 1 LoRAs -> models/loras/minimax-h3:
- minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (the 4-step turbo, strength 1.0)
- h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors (the realism LoRA, strength 1.0, trigger prefix r34l1sm)
Stage 2, the regenerate chain -> models/loras/Open-H3-Regenerate-2K. Pick one refiner:
- minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors (about 5GB, Kijai's extracted build, the one I recommend)
- open_h3_regenerate_2k_lora.safetensors (12.4GB FP32 original from the NVIDIA repo, only for reproducing method 2 exactly)
- minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors (about 155MB, Kijai, the low-VRAM experiment build)
Plus, both paths:
- sharp_curve_branches.safetensors (1.2GB from the NVIDIA repo, the edge and color sharpener)
- minimax_h3_lms_v1.0_r64.safetensors (LMS Sharp from the Alissonerdx repo, strength 0.8)
Upscaler and fast decode:
- minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors (LBH-123-AI repo, also mirrored in the NVIDIA repo's latent_upsampler folder; keep it where your MMH3 loader points)
- fast_h3_vae_noncausal_24l.safetensors (NVIDIA repo's vae_fast folder, for the OpenH3RegenFastVAEDecode node)
Custom nodes: ComfyUI-Open-H3-Regenerate (github.com/rdh9616), Comfyui-MMH3-UltimateUpscale (github.com/bbaudio-2025), KJNodes, VideoHelperSuite, ComfyUI-Easy-Use, an RMBG-2.0 pack, and a current ComfyUI with the native MiniMax H3 nodes. Drag the JSON in and let the manager install what it flags.
----------------------------------------
FILES ATTACHED TO THIS POST
----------------------------------------
- MinimaxH3_2 method Compare-NVRegen2k-20261028.json (121 nodes, method 1 and method 2 side by side, exclusive here)
- MinimaxH3_2Stage-NVRegen2k-20261028 (1).json (88 nodes, the free two-stage file, same as the public post)
----------------------------------------
QUICK START
----------------------------------------
1. Install the node packs above, drop the model files where your loaders expect them, then drag in the compare JSON. Start there, it has both methods wired.
2. Check the loaders before queueing anything: diffusion model, CLIP loader type (minimax), both VAEs, and the two stage-2 LoRA chains. Almost every broken first run is a loader path.
3. Load your character image. The chain runs RMBG-2.0 on it and resizes to 1230x1024 center crop before the reference conditioning, so a clean full or half-body shot works best.
4. Write the prompt in the six-part H3 format below, and keep the r34l1sm prefix line if you keep the realism LoRA.
5. Leave stage 1 at 0.4 megapixels for the draft and the stage 2 target at multiplier 3 for 2K. Duration 8 seconds, FPS 24, seeds fixed as shipped so your runs actually compare.
6. Queue. Stage 1 gives you the small draft, stage 2 regenerates it at 2K, and the fast VAE decode writes the clip. That's the whole loop.
----------------------------------------
WHAT TO INPUT (PROMPT FORMAT)
----------------------------------------
The prompt box takes MiniMax H3's structured six-part format, not free text. The two blocks that matter most here:
subject_definitions:
<Subject 1> is the young woman whose appearance comes from <Picture 1>, with [hair, clothing and accessories in one run-on sentence].
summary:
[reference generation] [one-sentence scene: duration, framing, what the character does].
Then the detailed description carries the rest of the shot. The shipped coffee-shop example: a ten-second all-close-up coffee moment of <Subject 1> at a cafe table, steam rising over her wrapped hands, one slow sip held on her lips, the cup lowering into a quiet satisfied breath, cut tight on face, hands and cup.
Rules that held up in my runs:
- Describe the character from the reference image in subject_definitions, not in the summary. The <Picture 1> tag is how the model ties the look to your loaded image.
- The StringConcatenate node prepends "r34l1sm , ultra realistic video. " to everything you write. That is the realism LoRA trigger. Keep prompt and LoRA together, drop both together.
- Time events in words ("midway", "near the end"). The duration float drives a math node that snaps frames to H3's 17-frame latent windows, so write seconds, never frame counts.
----------------------------------------
KEY SETTINGS (AND WHY)
----------------------------------------
- Stage 1: 0.4MP (864x480), euler, simple, 8 steps, denoise 1.0, sigma shift 12/3. The draft is supposed to be small and soft; the detail comes from stage 2. Ships correct, leave it.
- SoL Attention: the BlockSparseAttention node sits on the "sol-attn" preset with the sage attention and chunk feedforward patches. The NVIDIA repo says the engine needs this sparse attention for the forward pass. Ships correct, leave it.
- Stage 2, method 1: sa_solver, simple, 3 steps, denoise 0.35. Three steps is the framework's specialty; the refiner was trained to start from a partly re-noised upscaled draft instead of pure noise. The free 2Stage file ships 4 steps here; 3 also works and is what the compare file runs.
- Stage 2 LoRA strengths: LMS r64 at 0.8, refiner at 1.0, sharp_curve_branches at 1.0. LMS loads first in the upscale group in both files. At 1.0 the LMS over-sharpens in my tests; 0.8 keeps skin readable.
- Stage 2 target: ResolutionSelector multiplier 3 from the 480p draft lands at 2K. The template table in the workflow note has the full mapping, 2.0 is about 1080p.
- Method 2 only: OpenH3RegenUpscale preset "2k (2560x1440)" bf16, OpenH3RegenSigmas 3 steps shift 0 (that node builds the framework's own sigma schedule, 0.909 down to 0), RefineLatent on "silence (reference)", refine noise fixed. Decode through OpenH3RegenFastVAEDecode with the fast 24-layer VAE; the stock VAE decode at 2K is the slow part of this whole pipeline.
- Low VRAM: keep the MMH3 spatial split (864x480 tiles, 128 overlap) and temporal split (136 frames, 17 window), switch the final decode to VAEDecodeTiled, and leave the cleanGpuUsed and clearCacheAll nodes between stages. Plenty of VRAM and want speed? Delete the two split nodes, that is what the workflow note says.
----------------------------------------
TIPS
----------------------------------------
1. Skip the 12GB FP32 LoRA unless you are reproducing method 2 for science. Consumer GPUs cannot use FP32 anyway, and the rank 384 BF16 gave me the same pipeline principle at a fifth of the download. I only pulled the big file for the experiment.
2. Rank 16 is the low-VRAM door at about 155MB. The video only tested rank 384, so treat rank 16 as an experiment and expect softer detail.
3. sharp_curve_branches is a taste dial. In: sharper edges, punchier coloration, the coffee-cup steam reads cleaner. Out: softer, more natural look. Some of you will prefer it out, and that is fine.
4. Far shots are the money test. At 480p the character's face in a wide street shot is unidentifiable; after the 2K pass it holds up even while she walks fast. Judge your runs on the far shots, not the close-ups.
5. The 2K pass is not the finish line if you want crystal. Topaz Starlight Mini on top is the paid clean-up; a second round through any free open-source upscaler is the free one.
6. Timing: with SoL Attention on and the 3-step regenerate, a 2K clip came out in about a minute on my rig, faster than my old latent-upscale-only second stage. Budget your tests around that.
7. Any H3 base or fine-tune works in stage 1, and the Turbo and Realism LoRAs stack fine. I ran Singularity V2 int8; the note nodes list the hybrid and pruned alternatives if that is what you already have.
----------------------------------------
COMMON ISSUES
----------------------------------------
Method 2 output looks blurrier than method 1: not the LoRA's fault. The custom-node example's curve and attention table settings did not load into the diffusion model the way their example shows, likely a pack update thing. Run method 1; same pipeline principle, better output in my tests.
Out of VRAM at stage 2: tiled VAE decode, keep both split nodes, keep the cache-clean nodes. Still dying? Drop to the rank 16 refiner.
Regenerate stage errors or does nothing: check the sol-attn preset on BlockSparseAttention, and that the latent upscaler file is the 3d_conv_v1 bf16 build the node names.
Output over-sharp: lower LMS below 0.8 or drop sharp_curve_branches first. Touch the refiner strength last.
Character does not match your reference: keep the RMBG cutout and the 1230x1024 center crop in the chain; the workflow was tuned with the cutout character on a flat grey, and subject_definitions has to point at <Picture 1>.
Duration or frames look wrong: change the Duration float in seconds. The math node snaps frames to the 17-frame latent windows; hand-editing frame counts fights it.
----------------------------------------
Both workflow files are attached. Start with method 1, get one clean 2K clip out of it, then open method 2 to see why I prefer the simple path.
Questions? Drop a comment with your GPU, VRAM, and the exact error.
Thanks for supporting the channel. You guys make these deep dives possible.
- Benji
Resources:
- Project page: https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/
- NVIDIA GitHub (Sana, sol-engine): https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K
- Official model repo: https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K
- Kijai extracted refiner LoRA rank 384 (~5GB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors
- Kijai extracted refiner LoRA rank 16 (~155MB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors
- ComfyUI-Open-H3-Regenerate custom nodes: https://github.com/rdh9616/ComfyUI-Open-H3-Regenerate
- MMH3 Ultimate Upscale pack (method 1): https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale
- MiniMax H3 latent upscaler model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler
- LMS Sharp LoRA: https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/blob/main/loras/minimax_h3_lms_v1.0_r64.safetensors
- MiniMax H3 VAE and text encoders: https://huggingface.co/Comfy-Org/MiniMax-H3
- H3 hybrid diffusion model (normal): https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models