Awesome MiniMax-H3

A curated list of models, text encoders, quants, and tools for the MiniMax-H3 omni-modal video generation model.

• MiniMax-H3 official model card
• ComfyUI official blogpost
• ComfyUI tutorials for MiniMax-H3
• Video Prompt Writing Guide (Base)
• Video Prompt Writing Guide (Reference)

MiniMax-H3 is a general-purpose, omni-modal generative system by MiniMaxAI. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds. The model has two variants: FL2VA (first-and-last-frame mode) and Ref2VA (omni-reference mode).

Official and ComfyUI-repackaged model files.

• MiniMaxAI/MiniMax-H3 - Official repository.
• Comfy-Org/MiniMax-H3 - ComfyUI-repackaged model files.

Variant	Name	Precision	Size	Download
FL2VA	minimax_h3_fl2va		61.73 GB
FL2VA	minimax_h3_fl2va		31.70 GB
FL2VA	minimax_h3_fl2va_pruned		37.46 GB
FL2VA	minimax_h3_fl2va_pruned		19.52 GB
FL2VA	minimax_h3_fl2va_pruned		19.53 GB
Ref2VA	minimax_h3_ref2va		61.73 GB
Ref2VA	minimax_h3_ref2va		31.70 GB
Ref2VA	minimax_h3_ref2va_pruned		37.46 GB
Ref2VA	minimax_h3_ref2va_pruned		19.52 GB
Ref2VA	minimax_h3_ref2va_pruned		19.53 GB

Model Variants:

• H3-Base-FL2VA (First-and-last-frame mode): Supports zero, one, or two input images. No image input = T2V; one image = first/last-frame-to-video; two images = first-and-last-frame-to-video.
• H3-Base-Ref2VA (Omni-reference mode): Supports multi-modal reference inputs — up to 9 images, 3 video clips (2–15s each), 3 audio clips, max 12 files total.

4-step audio-video generation LoRAs — render joint video + synchronized stereo audio in 4 sampling steps instead of ~20 (~5× speedup). Early prototype; comfort zone for sharpness is 6–8 steps. The lightx2v distil (top row) is the shared base for most ComfyUI conversions; for pruned checkpoints use the ComfyUI-converted variants below. The original larryvrh LoRA targets the full (non-pruned) FL2VA checkpoint and needs the ComfyUI-MiniMax-H3-Turbo sampler node.

Variant	Steps	Pruned / Full	Precision	Size	Download
fl2v v0.1	4	Full		1.29 GB
fl2v v1.0 768p	4	Full		1.29 GB
fl2v v1.0 768p · comfyui	4	Full		1.82 GB
fl2v v1.0	8	Full		1.29 GB
fl2v v1.0 · comfyui	8	Full		1.82 GB
fl2v v1.1 768p	4	Full		1.29 GB
fl2v v1.1 768p · comfyui	4	Full		1.82 GB
ref2v v0.1	4	Full		1.29 GB
ref2v v0.1 · comfyui	4	Full		1.82 GB
fl2v v0.1 768p · SLA	4	Full		1.29 GB
fl2v v1.0 768p · SLA · comfyui	4	Full		1.82 GB
fl2v DasiwaREF2VAHybridV1 · curveproj1025 (T8) · ConvRot	4	Full		757.9 MB
fl2v 8-step merge 0821	4→8	Full		1.96 GB
lightx2v v0.1	4	Full		1.82 GB
lightx2v v0.1 · resized	4	Full		300 MB
fl2v	4	Full		744 MB
fl2v ema	4	Full		744 MB
fl2v ckpt500	4	Full		744 MB
fl2v ema ckpt500	4	Full		744 MB
fl2v ckpt850	4	Full		744 MB
fl2v ema ckpt850	4	Full		744 MB
fl2v v4 step600	4	Full		744 MB
fl2v v4 step600 ema	4	Full		744 MB
fl2v pruned	4	Pruned		592 MB
fl2v pruned ema	4	Pruned		592 MB
fl2v pruned ckpt500	4	Pruned		592 MB
fl2v pruned ema ckpt500	4	Pruned		592 MB
fl2v pruned ckpt850	4	Pruned		592 MB
fl2v pruned ema ckpt850	4	Pruned		592 MB
fl2v pruned v4 step600	4	Pruned		592 MB
fl2v pruned v4 step600 ema	4	Pruned		592 MB
fl2v v1.0 768p · resized	4	Pruned		298 MB
fl2v v1.0 · resized	8	Pruned		327 MB
fl2v pruned ckpt500 V1	4	Pruned		592 MB
fl2v pruned ckpt600 V4	4	Pruned		592 MB
fl2v pruned ckpt600 ema V4	4	Pruned		592 MB
fl2v pruned ckpt850 V1	4	Pruned		592 MB
fl2v	4	Full		717 MB
fl2v step 100	8 NFE	Full		738 MB
fl2v step 200	8 NFE	Full		738 MB
fl2v step 300	8 NFE	Full		738 MB
fl2v 4-step acceleration · ConvRot · ⚠️ needs dual-clock sampler or 8–10 steps	4	Full		779.9 MB
fl2v 4-step acceleration ema · ConvRot	4	Full		779.9 MB
fl2v v4 step600 (T8-convert) · ConvRot	4	Full		779.9 MB
lightx2v v0.1 · alpha8 T8-convert · ConvRot · ⚠️ needs dual-clock sampler or 8–10 steps	4	Full		1.96 GB
fl2v 10ErosMax test4 · 4-step curveproj1025 (T8) · ConvRot · ⚠️ needs dual-clock sampler or 8–10 steps	4	Pruned		794.9 MB
fl2v 10ErosMax test4 · 4-step curveproj1025	4	Pruned		794.9 MB
fl2v 10ErosMax test4 · 8-step v1.0 · ConvRot	8	Pruned		1.96 GB
fl2v CMF · full	4	Full	Q4TP (CMF)	25.20 GB
fl2v CMF · FL2VA	4	Full	Q4TP (CMF)	25.70 GB
fl2v CMF · FL2VA (smaller)	4	Full	Q2TP (CMF)	20.12 GB
fl2v v1.0 768p · ConvRot · needs ComfyUI-LoraInt8Loader	4	Full		991 MB
fl2v v1.0 · ConvRot · needs ComfyUI-LoraInt8Loader	8	Full		991 MB
lightx2v v0.1 · int8 · ConvRot · needs ComfyUI-LoraInt8Loader	4	Full		991 MB
flashgen v1.0 768p · T2VA · ⚠️ Ascend NPU / MindIE-SD / vllm-omni target (merge via merge_lora_ckpt.py)	4	Full		1.26 GB

larryvrh also publishes experimental training checkpoints (11 .bin files: step 149/490/729/850/922, v2 step 298, v3 step 300, v4 step 150/600, v5 step 600; 7.26–10.17 GB) — see the repo.

══════════════════════════════════

Unified quantization tables for FL2VA and Ref2VA. The Pruned column marks whether the checkpoint is AdaLN-pruned (smaller, ComfyUI-only). The Method column identifies the quantization scheme. Multiple sources for the same quant are separated by ┊.

Key: ConvRot = ConvRotation INT8/INT4 quantization · Lean = selective BF16 island retention · DT-sQKV = Dynamic-Time separate-QKV (patch required) · W4A8 = 4-bit weight / 8-bit activation · GGUF = llama.cpp GGUF format · NF4 = bitsandbytes 4-bit · OrbitQuant = native W4A4 packed path · Hybrid = partial NVFP4 layers on Blackwell.

Items marked ⚠️ require a ComfyUI core patch — they do not load in unmodified ComfyUI.

FL2VA — Unified Quantization Table

Pruned	Precision	Method	Size	Download
		BF16	61.73 GB
		Hybrid (fl2va base + ref2va adaln b15-49)	20.97 GB
		Hybrid (fl2va base + ref2va adaln b20-49)	20.97 GB
		Hybrid (fl2va base + ref2va adaln b25-49)	20.97 GB
		Hybrid (fl2va base + ref2va adaln b30-49)	20.97 GB
		ConvRot	31.70 GB
		FP8 E4M3FN	43.78 GB
		MXFP8	44.34 GB
		FP8 + FP16 attn	26.70 GB
		ConvRot Lean	21.91 GB
		ConvRot	20.94 GB
		ConvRot Lite	20.33 GB
		NVFP4	13.60 GB
		NVFP4	10.86 GB
		NVFP4	32.05 GB
		NF4	15.98 GB
		OrbitQuant W4A4	17.03 GB
		⚠️ DT-sQKV ConvRot	21.00 GB
		⚠️ DT-sQKV ConvRot Lean	27.99 GB
✓		BF16	37.46 GB
✓		FP8 scaled	19.52 GB
✓		ConvRot	19.53 GB	┊
✓		NVFP4	18.69 GB
✓		NVFP4 + ConvRot INT8	18.69 GB
✓		NVFP4	11.67 GB
✓		Mixed INT4/INT8 ConvRot	14.81 GB	┊
✓		Mixed INT4/INT8 ConvRot Lean	17.27 GB
✓		INT4 ConvRot	15.67 GB
✓		Mixed INT4/INT8 ConvRot	18.92 GB
✓		W4A8 ConvRot	11.68 GB	┊ ┊

GGUF quants — see GGUF section below.

Ref2VA — Unified Quantization Table

Pruned	Precision	Method	Size	Download
		BF16	61.73 GB
		ConvRot	31.70 GB	┊
		ConvRot Lean	21.91 GB
		ConvRot	20.94 GB
		ConvRot Lite	20.33 GB
		NVFP4	13.60 GB
		NVFP4	10.86 GB
		NVFP4	32.05 GB
		NVFP4	22.76 GB
		NF4	15.98 GB
		OrbitQuant W4A4	17.03 GB
		Hybrid NVFP4 (FFN-only)	16.38 GB
		Hybrid NVFP4 (QKV+FFN)	14.03 GB
		⚠️ DT-sQKV ConvRot	21.00 GB
		⚠️ DT-sQKV ConvRot Lean	27.99 GB
✓		BF16	37.46 GB
✓		FP8 scaled	19.52 GB
✓		ConvRot	19.53 GB	┊
✓		NVFP4	18.69 GB
✓		NVFP4 + ConvRot INT8	18.69 GB
✓		NVFP4	11.67 GB
✓		Mixed INT4/INT8 ConvRot	14.06 GB	┊
✓		Mixed INT4/INT8 ConvRot Lean	17.18 GB
✓		INT4 ConvRot	15.67 GB
✓		W4A8 ConvRot	11.68 GB	┊ ┊

GGUF quants — see GGUF section below.

· · · · · · · · · · · · · ·

GGUF quants for use with stable-diffusion.cpp, ComfyUI, and Unsloth. Non-pruned sources: Abiray/MiniMax-H3-GGUF, vantagewithai/MiniMax-H3-comfyUI-GGUF, realrebelai/MiniMax-H3_GGUFs. Pruned sources: unsloth/MiniMax-H3-GGUF, MarxistLeninist/MiniMax-H3-FL2VA-Pruned-IQ1-GGUF.

FL2VA GGUF

Pruned	Quant	Size	Download
		17.42 GB
		14.50 GB	┊ ┊
		14.50 GB
		17.36 GB	┊
		20.41 GB
		18.50 GB	┊ ┊
		18.49 GB	┊
		21.21 GB	┊
		24.17 GB
		22.25 GB	┊
		22.25 GB	┊
		26.28 GB	┊
		33.56 GB	┊
✓		6.26 GB
✓		8.16 GB
✓		10.64 GB
✓		12.97 GB
✓		15.45 GB
✓		19.97 GB
✓		7.51 GB
✓		8.90 GB
✓		3.78 GB
✓		4.22 GB

Ref2VA GGUF

Pruned	Quant	Size	Download
		14.50 GB	┊ ┊
		14.50 GB
		17.36 GB	┊
		20.41 GB
		18.49 GB	┊ ┊
		18.49 GB	┊
		21.21 GB	┊
		24.17 GB
		22.25 GB	┊
		22.25 GB	┊
		26.28 GB	┊
		33.56 GB	┊
✓		6.22 GB
✓		8.12 GB
✓		10.60 GB
✓		12.94 GB
✓		15.42 GB
✓		19.94 GB

· · · · · · · · · · · · · ·

Stock-compatible quants for the 10Eros_Max fine-tune of MiniMax-H3. Fine-tuned QKV weights in blocks 0–31 preserved alongside tested quantization layouts. No custom node or ComfyUI core patch required. (DmitryDB/MiniMax-H3-10Eros-Max-Quants)

Variant	Precision	Method	Size	Download
FL2VA 10Eros		ConvRot Lean	21.91 GB
FL2VA 10Eros		ConvRot	20.94 GB
FL2VA 10Eros		NVFP4	13.60 GB
FL2VA 10Eros		NVFP4	10.86 GB

Patch-required FL2VA for the 10Eros_Max fine-tune. DT-sQKV edition (DmitryDB/MiniMax-H3-10Eros-Max-DT-sQKV):

Variant	Precision	Method	Size	Download
FL2VA 10Eros		⚠️ DT-sQKV ConvRot	21.00 GB

Z-Image's spatial-attention profile grafted onto H3's engine (zs05 = late-block gains, dose 0.5) — richer sets and textures, same identity, no per-shot sharpening creep. Native ComfyUI cuts load with the plain Load Diffusion Model node (ComfyUI 0.32+); GGUF quants for the GGUF repo (fl2va/ref2va × curve/standard, Q4_0–Q8_0 + Q3mix, 10.7–24.1 GB). RTX 30/40: the GGUF repo is 4–8× faster than any 4-bit comfy-native arm on Ampere.

Variant	Precision	Size	Download
fl2va pruned zs05		—	see repo
ref2va pruned zs05 (master)		37.46 GB
fl2va pruned zs05 · int8_convrot		31.69 GB
ref2va pruned zs05 · int8_convrot		19.53 GB
fl2va pruned zs05		19.52 GB
ref2va pruned zs05		19.52 GB
ref2va pruned zs05	e5m2	19.52 GB
fl2va / ref2va pruned zs05	comfy	19.53 GB	┊
fl2va / ref2va pruned zs05		20.08 GB	┊
fl2va / ref2va pruned zs05		11.67 GB	┊
fl2va / ref2va pruned zs05	w4a8	11.68 GB	┊
ref2va pruned zs05	w4a4	10.56 GB

Community hybrid of joeygambino's ZS05 INT8 checkpoints: FL2VA base with REF2VA adaln_proj blocks 25–49 (b25-49 strategy; final layer stays FL2VA). Raw-tensor splice — no dequant/requant. Research/experimental; not claimed better than either source. Load as a diffusion model. (repo)

Variant	Precision	Size	Download
Hybrid b25-49 zs05		19.53 GiB

Native ComfyUI single-file conversion of diffusers-modular/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024 — a fused checkpoint carrying the Ref2VA delta LoRA at rank 1024 on the pruned base (see also ethanfel's unfused delta adapters in LoRAs → Experimental). Diffusion transformer only; use stock H3 TE + VAEs. INT8 variants keep all 50 MLP fc2 layers BF16 to avoid fused-swiglu INT8 OOM; validated end-to-end in ComfyUI (Continuum/Spectrum/refine). MiniMax H3 Community License.

Variant	Precision	Size	Download
BF16 native conversion		37.47 GiB
INT8 tensorwise · fc2 bf16		23.12 GiB
INT8 ConvRot gs256 · fc2 bf16		23.13 GiB

Official data-free DMD2 few-step distillation of MiniMax-H3 FL2VA by the FastVideo team (hao-ai-lab): 50-step base sampled in 4 steps ([999, 749, 500, 250] ladder, cfg 1.0, guidance-distilled), joint video+audio, 768×1344 @ 124 frames. Diffusers-format full pipeline (only transformer/ differs from base); student trained with VSA block-sparse attention (runnable dense). Preview status — v0.1 = step 1400, v0.2 = step 2900/4000; quality still maturing on high-motion detail. MiniMax H3 Community License. A ComfyUI-ready LoRA extraction of this checkpoint by drozbay is listed under Turbo.

ℹ️ Note: Beidouqixing/MiniMax-H3-DMD2-4step (previously circulated link) is dead (HF 404) — FastVideo's repos are the canonical DMD2 distills.

• t8star Ref2VA patchin HF 1.02 — experimental weight modification (not a quant): +2% on 2×2 spatial HF patch in the video-input projection. Tests showed weak HF agent gain; "oily/waxy" look not confirmed removed. Repo. (31.70 GB, INT8 ConvRot, listed in the Ref2VA table above with *(patchin)* label.)
• DmitryDB/MiniMax-H3-INT8-Lean-ConvRot is the same repo as DmitryDB/MiniMax-H3-ComfyUI-Quants (merged/rebranded by the author). Both names resolve to the same files.
• DmitryDB/MiniMax-H3-INT8-Lean-ConvRot-Dynamic-Time-Separate-QKV is the same repo as DmitryDB/MiniMax-H3-DynTime-sQKV. Both names resolve to the same files.
• Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI also includes a matching quantized text encoder: qwen3vl_32b_minimax_h3-w4a8_convrot.safetensors.
• Kijai/MiniMax-H3-experimental also includes an INT8 ConvRot video VAE: minimax_h3_video_vae_int8_convrot.safetensors (2.95 GB). See Components.
• unsloth/MiniMax-H3-GGUF also includes Qwen3-VL text encoder GGUFs: Q2_K_M (12.2 GB) and Q4_K_M (17.0 GB).
• DmitryDB/MiniMax-H3-ComfyUI-Quants also includes VAE files: Video VAE FP16 (4.85 GB) and Audio VAE FP32 (577 MB). See Components.
• DiffSynth-Studio/MiniMax-H3-NF4 also includes TE, Video VAE, and Audio VAE NF4 quants. Requires DiffSynth-Studio; minimum 8 GB VRAM.
• WaveCut/MiniMax-H3-OrbitQuant-W4A4 also includes quantized text encoder and FP32 VAE copies. Requires ComfyUI-OrbitQuant custom node. Workflow JSON.
• DeepBeepMeep/MiniMax-H3 is a community repack bundling both FL2VA and Ref2VA in every precision/pruning combination: full bf16 (66.3 GB) and int8_convrot (34 GB); pruned bf16 (41.4 GB) and int8_convrot (22.1 GB); and pruned_rank8 bf16 (40.3 GB) and int8_convrot (21.1 GB). Also ships VAEs (video fp16 5.21 GB, video fp8mix 2.79 GB, audio fp32 605 MB), a Qwen3-VL-32B text encoder (nvfp4_awq + Q4_K_M GGUF), and SeedVR2 upscaler checkpoints. No license is stated — clarify usage rights before redistributing. Repo

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

MiniMax-H3 uses the Qwen3-VL-32B model as its text/vision conditioning encoder.

Official and optimized versions for ComfyUI, repackaged by Comfy-Org.

Model Name	Precision	Size	Download
qwen3vl_32b_minimax_h3		47.97 GB
qwen3vl_32b_minimax_h3		25.28 GB
qwen3vl_32b_minimax_h3		14.61 GB

GGUF quantized text encoder, bundled with the Abiray/MiniMax-H3-GGUF repository.

Model Name	Precision	Size	Download
qwen3vl_32b_minimax_h3		13.58 GB
qwen3vl_32b_minimax_h3		13.93 GB
qwen3vl_32b_minimax_h3		25.28 GB

· · · · · · · · · · · · · ·

Built from llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic by ethanfel. Includes a MiniMax-H3 conditioning encoder (language layers 0–49 + vision tower) and an optional prompt-enhancement tail (layers 50–63 + LM head). The "Heretic" lineage bypasses alignment/restriction layers in the text encoder so MiniMax-H3 receives the most faithful prompt embeddings.

Model Name	Precision	Size	Download
qwen3vl_32b_heretic (conditioning encoder)		24.55 GB
qwen3vl_32b_heretic (generation tail 50–63)		7.09 GB

The generation tail is loaded temporarily by the ComfyUI-MiniMax-H3-Guide node for prompt enhancement, then unloaded. Requires the connected standard MiniMax-H3 CLIP (layers 0–49).

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Separated VAE files for MiniMax-H3. The video VAE and audio VAE are required for all generation workflows.

Component	Source	Precision	Size	Download
Video VAE			4.85 GB
Audio VAE			577 MB
Video VAE			2.60 GB
Audio VAE			289 MB

FP8-mixed quantized VAE set by dummy9996 — smaller video VAE (2.60 GB, fp8) and audio VAE (289 MB, bf16) for low-VRAM workflows.

· · · · · · · · · · · · · ·

Quickly trained 2D tiny VAE for MiniMax-H3 by Kijai. Not the greatest outcome, still beats latent2rgb for preview purposes. Currently only works with the ModelPreviewOverride node in ComfyUI-KJNodes.

Component	Size	Download
TAE (preview VAE)	9 MB

Experimental image-specialized MiniMax H3 VAE that decodes a single temporal latent (T=1) into one image. Merged H3 VAE checkpoint — no custom node required. For image workflows only; the image-tuned decoder materially regresses multi-frame video reconstruction, so keep the original H3 VAE for video.

Component	Size	Download
Single-image VAE (step 1597)	4.85 GB

· · · · · · · · · · · · · ·

Learned linear projections to condition H3 from a smaller text encoder. Two families: (1) ClipProj — swap the large Qwen3-VL-32B for a 4B/8B one (text-encoder VRAM ~15.7 GB → 4.5 GB, no change to the diffusion model, VAE, or sampler), and (2) H3 Control — identity/zero matrices for a no-control baseline. Projection files are fp16, MIT-licensed. Requires the ComfyUI-ClipProj node; place files in ComfyUI/models/clip_projections/. Full variant matrix (4B/8B × base/MLP/celeb/celeb-MLP): repo.

Variant	Encoder	Size	Download
ClipProj (base)	Qwen3-VL 4B	52.5 MB
ClipProj (MLP)	Qwen3-VL 4B	304 MB
ClipProj (celeb)	Qwen3-VL 4B	52.5 MB
ClipProj (celeb-MLP)	Qwen3-VL 4B	304 MB
ClipProj (base)	Qwen3-VL 8B	84 MB
ClipProj (MLP)	Qwen3-VL 8B	386 MB
ClipProj (celeb)	Qwen3-VL 8B	84 MB
ClipProj (celeb-MLP)	Qwen3-VL 8B	386 MB
H3 Control Identity	—	52.5 MB
H3 Control Zero	—	52.5 MB

Older h3_* filenames (with tap24 / CONDPROJ / int8convrot suffixes) have moved to obsolete/ — canonical names are now mmh3-*-ClipProj*.safetensors.

· · · · · · · · · · · · · ·

fl2va → ref2va behavior patch by lihaoyun6. Extracts 112 specific keys shared between the ref2va and fl2va weights and stores their differences as a single patch, letting the lighter FL2VA checkpoint partially mimic Ref2VA output quality. Requires the ComfyUI-MiniMaxH3_Ref-Patch node to load. Apache-2.0.

Component	Size	Download
Ref Patch	148 MB

· · · · · · · · · · · · · ·

Neural latent-space upscaler for MiniMax H3 video generation by LBH-123-AI. Works directly on H3's 24-channel VAE latents to upscale spatial resolution (H×W) while preserving the time dimension — accelerates high-res video gen by skipping the costly ~5B-param VAE decode → pixel-upscale → encode round-trip, and avoids the ghosting / double-image artifacts of naive bilinear/bicubic latent interpolation. 3D-convolution backbone (2D and 3D node variants; one checkpoint serves both, architecture auto-detected). Trained on ~80k paired samples (≈70k video + ≈8k 2K image pairs). Apache-2.0. Pairs with the ComfyUI_Minimax_h3_latent_Upscaler node. ⚠️ Saves time, not VRAM — the refine pass still runs at target resolution.

Component	Precision	Size	Download
Latent Upscaler		691 MB
Latent Upscaler		691 MB
Latent Upscaler		1.38 GB

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

• SexGod1979

  • PinkFluffyBunny - Pink fluffy bunny style LoRA in pruned + unpruned variants (rank 128/256/512). Maximum pink achieved at 0.5 strength on pruned int8 model. Alpha quality. (2.31 GB · pruned-v1 rank128)
  • PinkCherry - High-quality furry rabbits, rainbows, and cherry blossoms. No guardrails altered. Alpha v0.3 (pruned int8, 14 GB checkpoint). Iterated alpha 0.1→0.5.
  • NaughtyTimes - NSFW style LoRA for MiniMax-H3.

• ssjenforcer191

  • Homelander - Character LoRA for The Boys' Homelander. Triggerword HeroHomelander (optionally append wearing red leather gloves). Experimental. (296 MB)

• Playtime-AI/Minimax_H3-Mila_Kunis - Character LoRA of actress Mila Kunis (MM-H3 - Mila Kunis v1.70.safetensors, 155 MB pruned). Ships a sample audio clip; minimal card (Apache-2.0).

• Playtime-AI/Minimax_H3-Sydney_Sweeney - Character LoRA of actress Sydney Sweeney (MM-H3 - Sydney Sweeney.safetensors, 155 MB). Same pattern as Mila Kunis: ships a sample clip, minimal card (Apache-2.0).

• Playtime-AI/Minimax_H3-Salma_Hayek - Character LoRA of actress Salma Hayek (MM-H3 - Salma Hayek.safetensors, 155 MB). Same series as above; sample clip included, minimal card (Apache-2.0).

• SexGod1979 — AfterMidnight

  • AfterMidnight-MiniMax-H3-NSFW - ⚠️ Contains explicit / NSFW content. "AfterMidnight" Ref2VA NSFW style LoRA family (rank 64): sexytime v1 / v1.1 / v1.2 and softer v1. Apache-2.0. Also mirrored byte-identical at sasimi/AfterMidnight-MiniMax-H3-NSFW.

Variant	Size	Download
AfterMidnight_ref2va_h3_sexytime_rank64-v1.2.safetensors	1.11 GB
AfterMidnight_ref2va_h3_sexytime_rank64_v1.1.safetensors	1.11 GB
AfterMidnight_ref2va_h3_sexytime_rank64_v1.safetensors	1.11 GB
AfterMidnight_ref2va_h3_softer_rank64_v1.safetensors	1.11 GB

• matlod/minimax-h3-turnaround - Contact-Sheet diffusion — one reference image + one instruction → five coherent, progressively rotated views of the same subject in a single pass. A character turnaround from one photo (~10 s at 512², ~57 s at 1024²). Uses H3's timeline as a slot axis. (60 MB each: 1024-cont/s600, 512/s1500, 512-instruct/s400)

• EllaPriest45

  • MinimaxH3_Actions - ⚠️ Contains explicit / NSFW content. Collection of NSFW action LoRAs for MiniMax-H3 (T2V/I2V/R2V). Includes motion-specific LoRAs with trigger words and strength recommendations. See the repo for the full list. (reference only)
  • MinimaxH3_Characters - ⚠️ Contains explicit / NSFW content. Character LoRA collection for MiniMax-H3 (e.g. Aunt Cass, Baldur's Gate 3 Party Pack, Judy Hopps). Browse at your own discretion; not enumerated with per-file downloads here.
  • MinimaxH3_Styles - ⚠️ Contains explicit / NSFW content. Style LoRA collection for MiniMax-H3 with previews and config text; significant NSFW/nude portion (anime, digicam, Playboy styles). Browse at your own discretion; not enumerated with per-file downloads here.

• Hearmeman/minimax-h3-loras - ⚠️ Contains explicit / NSFW content. LoRA collection for MiniMax-H3 (repo tagged NSFW; MiniMax H3 Community License). Browse at your own discretion; not enumerated with per-file downloads here.

• fal/research-mini-max-h3-realism-people-lora - Realism LoRA for natural-looking people in everyday scenarios. Trained by fal on diverse photo data. (125 MB)

• Inner-Reflections/MiniMax-H3-Looping-Sketch-Anime - Looping anime-style sketch LoRA. Hand-drawn 2D outlines, flat colors, white outline. Strength 0.75–1.25; pair with a Turbo LoRA for higher strength. (569 MB)

• suryatmodulus/studio-1939-old-animation-lora-minimax-h3 - STUDIO 1939 — golden-age hand-painted animation style: gouache backgrounds with visible brushwork, warm celluloid palettes, classical character animation. Trained on a curated selection from a public-domain 1939 animated feature; two flavors from the same run — light (r16, painterly/storybook) and strong (r64, full cel with bold flat-shaded characters). Trigger gulliv3r, at prompt start; prompt expansion off; scale 1.0 for full look, 0.4–0.8 to blend under modern art direction; works 21:9 / 16:9 / 4:3. Character consistency via frozen verbatim character descriptions. By Lovis Odin (@OdinLovis). (65.6 MB light r16 · 262 MB strong r64)

• nikdevs/minimax-h3-loras - ⚠️ Contains explicit / NSFW content. Curated MiniMax-H3 LoRA collection (styles + characters). Browse at your own discretion; not enumerated with per-file downloads here.

• DiffSynth-Studio/MiniMax-H3-LoRA-LineartAnime - Anime video line-art colorization — feeds a line-art video as a reference and generates fully colored anime output from it (Ref2VA video-reference workflow). Apache-2.0. (1.26 GB)

• Jojocodex

  • minimax-h3-wushu-action-lora - Wushu / martial-arts action — trains H3 to generate human martial-arts motion (punches, kicks, spins, staff techniques), focused on body physics. Trigger by action description (e.g. a martial artist performing punches and kicks in fast combat); no fixed trigger word. ai-toolkit, rank 16, 2000 steps, 512 / 90 frames @ 24fps; pruned + full safetensors. ComfyUI users load the _pruned variant at strength 0.8–1.0; compatible with the Turbo LoRA (adaln_proj trimmed, 417 keys). Base-model use is under the MiniMax H3 Community License. (155 MB pruned · 310 MB full)
  • minimax-h3-spatial-physics-lora - Spatial & physics (objects) — teaches H3 object physics (collision, stacking, falling, occlusion) via pure spatial+physics captions; complements the wushu LoRA, which covers body motion. No fixed trigger word — describe object motion directly. Trained on CLEVRER / WISA / PhyCo-Kubric (700 clips); ai-toolkit, rank 16. ComfyUI users load _pruned at 0.8–1.0; stacks with the Turbo LoRA. (155 MB pruned · 310 MB full)
  • minimax-h3-yunjing-lora - Camera-movement (yunjing) control — cinematic camera-movement control (push in/out, orbit, tracking, handheld) via the yunjing trigger word. 12 movement types trained (handheld / pull / dolly best-covered; pan / crane / 360° weakly covered). ai-toolkit, rank 32, 1000 steps; pruned + full. ComfyUI users load _pruned at 0.8–1.0; stacks with the Turbo LoRA (6–8 steps, Euler, Beta). (310 MB pruned · 620 MB full)
  • minimax-h3-Camera-Motion-lora - Dedicated camera-motion LoRA (v1, 1000 steps) — broader camera-movement control for H3, shipped with a camera_motion_Prompt Library.md of ready-made motion prompts. Separate, dedicated camera-motion training from the yunjing LoRA (use one or the other per shot). ai-toolkit, pruned safetensors. ComfyUI users load _pruned at 0.8–1.0; stacks with the Turbo LoRA. (155 MB pruned)

• bghira/minimax-h3-anyflow-wip - SimpleTuner WIP LoRA checkpoints (steps 200/300/400/500 + EMA). WIP research builds; not production-tuned.

• ethanfel/MiniMax-H3-Pruned-Ref2VA-Delta-LoRAs-Experimental - Highly experimental, mechanically extracted adapters — randomized-SVD approximations of the weight difference between pruned FL2VA and Ref2VA checkpoints. Not trained as LoRAs, not generation-tested. Explore behavior transfer in either direction. (ranks 256/512/1024, BF16)

• Kijai/MiniMax-H3-experimental loras - Experimental rank-256 BF16 LoRA capturing the FL2VA↔Ref2VA difference (same class as ethanfel's). No confirmed use case yet. (2.40 GB)

• DIE2025/MiniMaxH3Loras - Three unnamed style LoRAs (B, Spicy, V) of equal size. No README; use at own discretion. (310 MB each)

Variant	Size	Download
MiniMaxB.safetensors	310 MB
MiniMaxSpicy.safetensors	310 MB
MiniMaxV.safetensors	310 MB

• adehong/minimax-h3-ntt-lora - "NTT" v2 LoRA in three ranks (128 / 256 / 512). No model card; purpose and trigger words undocumented — use at own discretion.

Variant	Size	Download
ntt-v2-rank128.safetensors	1.12 GB
ntt-v2-rank256.safetensors	2.24 GB
ntt-v2-rank512.safetensors	4.48 GB

• MATLOWAI/MiniMax-H3-Motion-Adapter - Motion adapter (pilot, r16) — a small rank-16 BF16 LoRA that improves the de-rope pass in ComfyUI-MAINodes on fast motion: reduces frame-by-frame advance/snap alternation and over-production, and transfers to both FL2VA and Ref2VA graphs (one file). Trained bf16 (rank 16, alpha 16). MIT for the adapter weights; base model use is under the MiniMax H3 Community License. Load with a stock LoraLoaderModelOnly at strength 1.0 on the de-rope pass only. (63 MB)

• mvp-lab/MiniMax-H3-RAVEN-Streaming-LoRA - RAVEN: real-time autoregressive video extrapolation — turns MiniMax-H3 into a causal streaming generator that extrapolates each chunk from previously generated content (4-NFE preview) instead of denoising one bidirectional clip. Academic preview (Imperial College London); the released weight is undertrained (limited texture) but validates the full RAVEN training→generation pipeline. Single PEFT LoRA adapter, r=128 / lora_alpha=128; 192 frames @ 768×1376, 24 fps, causal chunking sink=2 / window=2. Training/inference/eval code in mvp-ai-lab/RAVEN. MiniMax H3 Community License. (≈5.1 GB)

  ◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Node	Author	Category	Description
MiniMax H3 Hybrid Cond	kitsune123150		Hybrid R2V + I2V conditioning in one payload. Outputs positive conditioning and AV latent with native audio.
ComfyUI-H3-Multishot	jlucasmcrell		Multishot video+audio generation — N chained shots from one script, seam-clean master. Keyframes at any position, dual-format loader (safetensors + GGUF).
ComfyUI MiniMax H3 Director	seesee75-commits		Timeline editor with storyboard — drag media onto tracks, trim on a ruler, write a prompt per shot. Live sampling preview, retakes, shot chaining.
ComfyUI MiniMax H3 Image Studio	astropuzzo		Image-first nodes for T2I, I2I, and reference editing. Arbitrary frame counts, resolution up to 64 MP, automatic still-frame scoring.
ComfyUI-MiniMaxH3-Easy	nkxx188		One compact workflow for T2V, I2V, first/last-frame, and reference video. Unified multi-media input with @ references and inline dialogue blocks.
H3 Motion Context	NikoDemon80		Chain H3 clips so motion and sound keep going across the cut. Feed clip A's last frames + audio in; clip B picks up where A left off — same motion, same audio.
ComfyUI MiniMax H3 Motion Director	j955229		Multi-segment motion director combining AIMixer Director's timeline + Motion Context chaining. Reference control across N segments.
H3 Conditioning Cache	HEEEeeeeN		Conditioning cache + batch generation suite for H3 drama/short-drama production. Caches conditioning across shots, batch-generates episodes unattended.
MAINodes	matlowai		Contact-Sheet diffusion (five views from one reference) + Motion Lab (test-time de-roping of fast-motion smearing: backflips, sword arcs, reversals).
Fantastic MiniMax H3 Prompt Builder	Adudeguyman		Fillable prompt templates for every H3 mode with live guide-rule checking and a media loader that manages reference tags.
MiniMax-H3 Prompt Enhancer T8	T8mars		Multimodal prompt enhancer calling doubao-seed-evolving. Analyzes text, images, and video together. Supports all H3 modes, strict/balanced/creative, CN/EN output.
MiniMaxH3 LatentUpscaler	Tr1dae		Latent spatial upscaler for H3's NestedTensor AV latents. Re-noises video/audio for two-pass sampling, scales minimax_refs/minimax_keyframes conditioning.
ComfyUI Video Tiler	maDcaDDie2000		Memory-conscious video/image tiling with overlap tiles, gaps, and feather blending. Built for LTX 2.3 and MiniMax H3 tiled upscale workflows. Disk-backed mode for low-VRAM.
H3 Latent Upscaler (Mamad8)	mamad8c		Moves a clean H3 video latent to a 2× larger spatial latent grid very quickly. Not a conventional upscaler — output looks softer than input; the point is to get a 2× grid ready for a second pass.
MiniMaxH3 Frame Infill	red-polo		Experimental node to regenerate any frame interval of an existing H3 video. Patches ComfyUI's H3 internal implementation; pin your ComfyUI version.
ComfyUI-SolAttn_triton	kijai		SolAttention Triton kernel for ComfyUI. Optimized attention computation for H3 and other Sol-Attn models.
ComfyUI-sol-attn	Saganaki22		Zero-copy Sol-Attn for SM89–SM120 with scheduled tau, graph preview, and feed-forward chunking. 1.14–1.44× vs SageAttention, −37% MLP peak VRAM on H3.
ComfyUI Spectrum MiniMax H3	xmarre		Spectral feature forecasting — skips selected transformer evaluations via Chebyshev ridge regression. Adaptive scheduling with native fallbacks.
ComfyUI-MiniMax-H3-RefDelta-Solver	xmarre		Dedicated ER-SDE-derived sampler + beta-prior scheduler pair built for the author's Pruned Ref-Delta Fused r1024 checkpoint: measures the fused checkpoint's divergent intermediate trajectory in ER-SDE solver coordinates and smoothly gates history-dependent corrections per stage, with separate video/audio risk reduction. Nodes: RefDelta Sampler (SamplerCustomAdvanced-compatible; strict stock mode delegates to native sample_er_sde), RefDelta Scheduler (r1024_provisional rank profile; profile JSON only at runtime), and a dev-only Reference Diagnostic guider comparing fused vs genuine Ref2VA on identical inputs. ⚠️ Experimental — controls/profile not yet calibrated against same-state telemetry. Requires ModelSamplingAV; compatible with ComfyUI-Spectrum-MiniMax-H3 v0.2.18+. GPL-3.0-or-later.
Herrgotts-H3-Infinite-Continuation-Suite	HerrgottMargott		Freeze-aware, keyframe-anchored MiniMax H3 video continuation for ComfyUI — injects the previous clip's video+audio latent context into the next FL2VA segment, auto-detects H3's frozen tail for a safe handover, and stitches with a 4-frame video crossfade + 15 ms audio de-click. Experimental community project (GPL-3.0).
ComfyUI-MiniMaxH3-Cache	lihaoyun6		EasyCache-style cache node for H3. Patches ComfyUI core to cache and reuse transformer block computations across timesteps.
MiniMax H3 Block Cache T8	T8mars		F1B0 block cache — computes Block 0 and reuses residual for Blocks 1–49 when audio/video are stable. Skips up to 49 of 50 blocks per step.
TE-Speed-MiniMaxH3-OSS	HELPMEEADICE		Block-cache accelerator patching H3's 50-layer DiT loop. Reuses cached tail-block residuals when sigma delta is small. ~45% speedup at default settings.
MiniMaxH3 Dual-Clock Euler Sampler	shuaixn		Dual-clock Euler sampler for the Turbo LoRA — fixes audio crackling/noise at 4-step generation by running video and audio on separate schedules.
ComfyUI-H3-AudioRefine	Adudeguyman		Audio-only refinement pass for H3 packed AV latents: freezes the video stream of an already-sampled latent (e.g. a 4-step Turbo pass) and denoises just the audio for extra steps — via ComfyUI's native masked-inpaint path with a per-stream noise mask (video=0 preserve, audio=1 generate; video bit-identical at video_denoise 0). Nodes: all-in-one Audio Refine Sampler (audio_denoise 0.3–0.6 cleans noise floor, 1.0 regenerates against frozen video), composable Audio Refine Mask, and optional Frozen Video Cache (~5.7× faster cached steps on RTX 5090; ~10 GB memory). Example workflow included; refine branch runs without the Turbo LoRA to restore undistilled audio quality. MIT.
minimax-h3-mlx	mrbizarro		Apple Silicon MLX port of the full H3 pipeline. AdaLN precompute drops 13B params at inference. Validated against the diffusers reference.
ComfyUI-ClipProj	nicolab28		Swap a large text encoder for a small one via a learned linear projection. MiniMax H3 conditioning from 15.7 GB down to 5.2 GB. Proof of concept.
ComfyUI MiniMax H3 Contex Loop	ethanfel		Turn one sampling body into a scene-by-scene production loop — each accepted scene carries motion + audio forward, saves a checkpoint, joins into final video without huge cumulative tensors.
ComfyUI MiniMax H3 LongMedia	vizart-vj		Long single-pass video/audio generation with streamed Sol attention, compressed KV, adaptive VRAM guards, chunked MLP/final output. SAFE long-sequence optimizations for limited VRAM.
ComfyUI MiniMaxH3 Hybrid Loader	scottmudge		Load a checkpoint by merging selected tensor groups (e.g. adaln_proj only) from a ref2va overlay onto a fl2va base. Default preset preserves ref-conditioning pathway while keeping fl2va quality.
ComfyUI MiniMax H3 Legacy Audio Sampling	starsFriday		Restores the v0.30.0 audio sampling behavior after upgrading to ComfyUI v0.31.0. One model-patch node — no source modification. Fixes regressed audio (background noise, stereo stability, HF artifacts).
ComfyUI-H3-FaceRefine	Carasibana		Face-refinement node for MiniMax H3 outputs — repair/enhance faces in generated video frames.
ComfyUI-MiniMaxH3Mod	Luisacaotica		No-training "RefMod" reference adapters for MiniMax H3 — compress reference images/videos into tiny .safetensors latent files reused like LoRAs without loading heavy references or training. Extract/Load/Apply nodes, folder and A/B-axis loaders, a standalone CLI, and strength/retention controls injected via the model's native conditioning path.
ComfyUI MiniMax H3 Extender	tritant		Chains multiple H3 clips into one long continuous sequence, preserving motion, visual, and audio continuity. Combines Ref2VA conditioning, motion context, disk latent caching, dynamic image references (up to 9), audio reference support, per-clip prompt/seed/duration, clip validation, and seamless video/audio decoding with seam correction (H.264 / H.265 / FFV1 export).
ComfyUI ALLinONE MiniMaxH3	LeonQ8		All-in-one MiniMax H3 node — T2V, I2V, R2V, audio drive (lip sync), keyframes, extend, chain (multi-clip continuation via H3 Motion Context), and an RTX/Seed2VR upscale hook in a single node. Ships searchable history, a library, and settings UI. Beta, GPL-3.0.
ComfyUI Qwen H3 Prompt	chflame163		Generates H3 prompts inside ComfyUI with a local Qwen3.8-27B GGUF model (bundled llama-server, fully offline) plus the official MiniMax-H3 Skills. Routes modes (T2VA/I2VA/L2VA/FL2VA/Ref2VA) from image/video references, writes sound design, and supports think mode with per-reference image/video inputs.
OpenH3-IR	ruashots		Open-source, local implementation of MiniMax H3's Context-IR stage — compiles a plain-language sentence (with optional referenced media) into a structured, validated six-section H3 video brief that feeds ComfyUI's native H3 render nodes. Three nodes (Main, Media, Setup), a creativity-level dial, strict brief validation, and exact dialogue/reference-image binding via @-syntax prompts.
MiniMax H3 Latent Upscaler	LBH-123-AI		Learned neural latent upscaler for H3's 24-channel VAE latents — upscales spatial resolution (1×–4×, continuous) in latent space via 2D/3D backbones, skipping the costly VAE decode/encode round-trip to accelerate high-res video gen and avoid ghosting. Pairs with the LBH-123-AI/Minimax_h3_latent_Upscaler checkpoint (weights auto-detected). Saves time, not VRAM.
Comfyui-MMH3-UltimateUpscale	bbaudio-2025		Single-node re-sampling (enhance/upscale) pass over a finished H3 AV latent under tight VRAM: temporal chunking (arbitrarily long clips) × spatial tiling (arbitrarily high res) with peak VRAM bounded to one tile; audio carried through untouched (never re-sampled). Two upscale modes per chunk — the H3 3D model-based upscaler (LBH-123-AI minimax_h3_latent_upscaler_3d_* checkpoints) or model-free latent interpolation — with frame-0 anchoring across chunks (anchor_strength 0.999), cross-fade temporal stitching, frozen overlap masks + masked write-back for seamless tile seams (linear/smoothstep/midpoint blends), per-piece time/spatial conditioning re-anchoring, and CPU offloading so H3 + upscaler never share VRAM. Nodes: MMH3 Ultimate Upscale, Temporal/Spatial Split Params, Latent Upscale (with Model) Params. Built on the author's LatentSplit + LBH-123-AI's upscaler network. MIT.
ComfyUI MiniMax H3 Studio	thaakeno		"H3 Studio" — turns H3 into a maintained ComfyUI image workflow: T2I, I2I, reference editing via one Director node, up to 9 ordered multi-references (@Image1–@Image9), LightX/PDD accelerated paths, smart Qwen3-VL prompt prep, YOLOv8 Face Refine, TAEH3 previews, and a Benchmark Lab. Alpha (MIT code). ⚠️ Not compatible with ComfyUI Nodes 2.0 yet.
ComfyUI MiniMax H3 Sampler Unlimited	hradec		Chunked replacement for SamplerCustomAdvanced (SamplerCustomAdvanced-Unlimited) that samples long H3 video/audio latents in chunks with native latent continuation — produces >15 s video and 2K on ~16 GB VRAM without loop workflows. Frame-accurate shot-prompt rewriting, accumulated live preview.
ComfyUI MiniMax H3 Parallel	AesSedai		Exact activation-only multi-GPU attention-head sharding for H3 Ref2VA — model/TE/VAE stay on the model GPU; helper GPUs receive packed INT8 Q/K/V head slices and return BF16 attention. Up to 4× peer-access CUDA GPUs (Comfy Kitchen INT8 attention); ~2× denoiser speedup at 4 GPUs with bit-identical output.
ComfyUI MiniMax H3 SPEED	StanLukuvka		Progressive-resolution (Spectral Progressive Diffusion / SPEED) sampler for H3's packed video+audio latent — replaces KSAMPLER + SamplerCustomAdvanced and denoises starting coarse (¼–½ res) then refines to full, cutting VRAM and wall-clock time. Presets (half_then_full default, three_quarter_then_full, quarter_half_full, aggressive, quarter_half_3q_full) trade speed vs mid-frequency detail. Requires the StanLukuvka/ComfyUI-MiniMax-H3 plugin (ComfyUI 0.32.0+). ⚠️ PolyForm Noncommercial 1.0.0 license.
ComfyUI MiniMax H3 Keyframe Offset	asirusasr-maker		Drop-in replacement for the stock MiniMax H3 Image-to-Video conditioning node, injecting first_frame/last_frame keyframes at arbitrary frame indices (not just start/end) so H3 freely generates motion between them. Plus an all-in-one text-to-audio node (conditioning → sampling → audio-VAE decode in one node; CFG hardcoded 1.0). 23 samplers / 9 schedulers, smart offset clamping, non-invasive in-memory PackedLayout patch. Apache-2.0.
MaskVidExperiments	drozbay		Video masking / inpainting utility — crops a stable region around a masked subject, processes it at high resolution inside a moving crop, then pastes it back without jitter or visible seams (naive per-frame crops jitter, which video models read as camera motion). Nodes: Subject Crop (stable by construction through mask noise/occlusions), Subject Uncrop (feathered paste-back), Mask Cleanup, Frame Range Mask, Mask To Latent Space (token grid e.g. 2×2 for MiniMax H3 → latent noise mask), Audio Mask To Latent, Differential Diffusion (Soft), Audio Mask Debug. GPL-3.0. Requires ComfyUI 0.15.0+.
ComfyUI-MiniMaxRefPack	Hearmeman24		Manages all 18 Ref2VA references from one node's own upload UI — preview, crop/trim and delete, with the tag H3 will actually use (<Picture 2>, <Video 1>, <Audio 1>) shown on every tile. Wire the 18 sockets plus prompt once and the graph never changes again. Writes the six-section H3 prompt for you via OpenRouter or any local OpenAI-compatible server (one-click discovery of Ollama / LM Studio / llama.cpp / vLLM, loopback-only), or passes your text straight through. Portable JSON configs, standard/replacement/auto registers, editable system prompt, and a debug output showing the exact request sent. MIT.

• keys-heretic-MiniMax-H3 sol-engine + speed upgrades + upscaler finish — Single DGX Spark by drowzeys - One-shot recipe for MiniMax-H3 on a single NVIDIA DGX Spark (GB10, sm_121): Sol-Engine ports, Ultra-Heretic TE, Spectrum forecasting, SageAttention, 0.5 MPix generate + RealESRGAN x2 finish. Includes formal benchmark ladder (1.55× vs dense stock).

• h3.c (h3-metal) by antirez - Native C/Metal inference engine for Apple Silicon. Prompt-to-video/audio, first/last-frame, and Ref2VA references work end-to-end on M3/M5 Max. Interactive Iris-style session. Not a ComfyUI node — standalone binary.

• Omni-Rewriter by WayneJin0918 - Open agentic prompt-expansion (PE) harness for image/video generation. Turns everyday intent into validated, model-ready prompts via a bounded AI-agent loop (Analyze → Draft → Validate → Repair → Render). Current video profile is MiniMax-H3; ships a CLI (omni-rewriter expand) + HTTP server (POST /v1/expand), deterministic PE validation, and a reusable CI lint Action. Apache-2.0. Not a ComfyUI node — standalone tool (generation adapters stay outside expand).

• MiniMax-H3-Prompt-Rewriter-LoRA-8B — PEFT LoRA adapter on Qwen3-VL-8B-Instruct (lightx2v) that turns short user requests into production-oriented MiniMax-H3 audio-video prompts — structured shot timeline, synchronized physical/ambient sound, and music guidance. Covers T2VA / I2VA / L2VA / FL2VA (text + keyframe-conditioned); Ref2VA not supported. Pair with LightX2V (or the pytraveler ComfyUI node) to generate. GGUF quants (pytraveler) run under llama.cpp against a quantized multimodal Qwen3-VL-8B-Instruct (sees reference frames) and ship a ComfyUI node.

  Format	Precision	Size	Download
  PEFT adapter		2.79 GB
  GGUF		1.30 GB
  GGUF ★		0.69 GB

  ★ Q8_0 recommended for most setups.

• MiniMax-H3-Single-Frame-VAE-500K by iamkaikai - Single-frame image decoder (VAE) for MiniMax-H3, trained 500K steps — reconstructs, generates, and edits individual H3 frames (text-to-image, material edits, sketch→render, reconstruction). Ships load_decoder.py + example prompts; use alongside the H3 video VAE for image-only work. Download minimax_h3_single_frame_decoder_500k.safetensors (≈9.69 GB) via resolve/main.

• MiniMax-H3-Fun-Controlnet-Union by alibaba-pai (PAI) - Fun ControlNet Union for MiniMax-H3 — single unified control adapter covering canny, depth, HED, MLSD, and pose conditioning for H3 video generation (sample results per mode in the repo). MiniMax-H3-Fun-Controlnet-Union.safetensors (6.81 GB); Apache-2.0 with MiniMax H3 Community License terms for the base model (LICENSE).

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

• Video Prompt Writing Guide (Base) - Official MiniMax-H3 prompt writing guide for base (FL2VA) mode. Covers prompt structure, camera language, scene composition, and best practices for text-to-video and image-to-video generation.
• Video Prompt Writing Guide (Reference) - Official MiniMax-H3 prompt writing guide for reference (Ref2VA) mode. Covers multi-modal reference inputs, image/video/audio reference handling, and prompt construction for omni-reference generation.

• ComfyUI MiniMax-H3 Tutorial - Official ComfyUI documentation tutorial for MiniMax-H3 setup and usage.
• MiniMax H3 Day-0 Support in ComfyUI - ComfyUI blog post covering open weights, native audio, 2K video output, and local execution on a 3060.

• MiniMax H3 — Performance & Best-Configuration Report - Local-inference performance guide for MiniMax H3 (FL2VA / Ref2VA) across consumer & workstation GPUs, Apple Silicon, and the DGX Spark — distilled from 2 hard-numbered benchmarks and 17 community field reports. Covers a TL;DR config recommendation, hardware-tier tiers, the best speed/quality recipe, and caveats & licensing.
• MiniMax H3 on an RTX 3060 12GB: what we actually measured - Real-world write-up of running MiniMax-H3 on a 12 GB RTX 3060 — what actually fits, at what resolution and step counts, and the configuration that worked.

• MiniMax H3 — 1,000-Prompt Curation - Curated index + analysis of the ostris/minimax_h3_1k dataset (1,000 prompts + 768p clips, generated with the pruned INT8-ConvRot FL2VA checkpoint @ 30 steps). Explains H3's 3-field prompt structure (integrated_multimodal_description / overall_soundscape / non_diegetic_music), highlights 10 reusable prompts with commentary, and compares H3 vs Seedance / Veo / Kling on fidelity, dialogue, sound design, and multi-shot continuity.
• Interactive atlas of all 1,000 clips (neta.art) - Browse every clip from the 1K prompt dataset — every prompt, every style — with per-clip metadata: shooting-style/subject filters, prompt / soundscape / music / aspect-ratio / dialogue details, one-click generate or download.
• Codex × MiniMax H3 自动成片与验收 Skill - Codex Skill for automated multi-shot H3 video production + QA: Codex splits storyboards and writes prompts, Z-Image generates first/last frames, MiniMax H3 Director schedules H3 shot generation (with audio), HyperFrames handles editable timeline editing/rendering, then Codex verifies dialogue (ASR), continuity, black frames, and specs — with local rework loops. Windows 11 + PowerShell 7 + ComfyUI ≥ 0.30; validated on RTX 5070 Ti 16 GB (~49 GB models). MIT; no model weights bundled.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

ComfyUI workflow templates and community graphs for MiniMax-H3, organized by generation mode. Direct-import .json links where available; pack repos link to the repo root.

Comfy-Org's original templates/video_minimax_h3_* T2V/I2V/R2V graphs were moved to archived/ and are now API-based (Hailuo/Minimax API) templates — links above point to the archived JSONs. Abiray also ships a Ref2VA GGUF workflow (newly listed).