An i2v / t2v MiniMax H3 workflow tuned for maximum speed on a single 24 GB card, with native stereo audio in the same pass. On our RTX 4090 a 10 s clip that took ~27 min on the dense hi-res pipeline renders in ~6 min with this stack (same seed; the hi-res pass is the optional extra below). The Turbo LoRA alone already measured x1.9 end-to-end.

What is inside

• 10Eros-Max base (INT8) - NSFW-focused finetune of MiniMax H3 by cicalooo, int8 convrot skip_edges variant (first/last blocks kept in BF16, the quality-safe option). Same loader as the official base: swap one filename to go back SFW.
• 4-step Turbo distill LoRA (larryvrh, v4_step600_ema, the strongest checkpoint of that line) + the dedicated dual-clock sampler: video and audio each keep their own shift (12/3), so 4-6 steps do not wreck the audio.
• Sparse attention (Zironic's H3-Optimizations, Advanced node): author-measured 1.57-1.69x on 243-frame clips, Kitchen INT8 backend, zero extra dependencies.
• SageAttention patch (KJNodes): another ~20% off sampling in our A/B.

Downloads - models

Put each file in the listed ComfyUI folder.

• models/diffusion_models (pick one base):
  10Eros-Max INT8 skip_edges (21.7 GB, gated repo: log in to HuggingFace and accept the terms; file 10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot_skip_edges.safetensors)
  or the official INT8 base (21 GB, SFW drop-in)
• models/loras: minimax_h3_turbo_v4_step600_ema.safetensors (780 MB)
• models/text_encoders: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB)
• models/vae: video VAE fp16 (5.2 GB) + audio VAE fp32 (0.6 GB)

Downloads - custom nodes

• ComfyUI-MiniMax-H3-Turbo (larryvrh) - the Turbo LoRA loader + sampler. Keep THIS loader: on the quantized base the LoRA must run in bypass mode and re-inject the adaln keys the pruned base needs. A generic LoRA loader either silently does nothing or crawls at minutes per step.
• H3-Optimizations (Zironic) - H3SparseAttentionAdvanced.
• ComfyUI-KJNodes (kijai) - SageAttention patch node. Needs the sageattention pip package; cannot install it? Set the node to disabled, you only lose ~20%.

Requirements

• ComfyUI 0.30+ (native INT8 convrot support).
• Tested on a 24 GB card (RTX 4090, 23.3/24 GB used at 0.3 MP). The NVFP4 text encoder is emulated on Ada: works, just slower at encode time.
• Frame count must satisfy count mod 17 == 5: 56, 73, 90, ... 243. 56 frames = 2.3 s at 24 fps.
• Stay at or under 0.5 MP; go big with the latent hi-res pass below instead.

Settings that matter

• Steps 6 shipped (4 = fastest and the author's minimum; 6-8 looks noticeably better; above 8 it over-sharpens).
• Scheduler simple with the Turbo LoRA (switch to beta + ~20 steps if you remove it; node 17 is left in the graph to rewire).
• Sparse level shipped at the author defaults (budget 0.3, edges [email protected]). Honest note: sparse attention is not free - on complex prompts adherence can drop, because with 4-6 steps the sparsity also touches the composition steps. Fix: raise to budget 0.5 / edges 0.8, or unplug node 810 (guider back to node 600) and keep the rest of the stack.
• Prompt recipe H3 responds to: style, then <Picture 1> + subject, the action, ONE camera move, then an Audio: block (voices, sfx, music - the audio is generated from your prompt in the same pass).
• t2v: the first_frame input of the MiniMax node is optional - drop the image and prompt only.

Going further (optional)

• Comfyui-MMH3-UltimateUpscale (bbaudio-2025): upscales the nested video+audio latent BEFORE decode - no tile seams, audio untouched. This is the stage our 27-to-6-min number runs sparse.
• Heretic NVFP4 text encoder (sakamakismile): uncensored drop-in replacement for the official encoder (same folder, swap clip_name in node 13) if it refuses your prompts.

Credits

MiniMax (H3 open weights), Comfy-Org (ComfyUI repackage), larryvrh (Turbo LoRA + nodes), Zironic (H3-Optimizations), cicalooo (10Eros-Max), kijai (KJNodes), bbaudio-2025 (MMH3 UltimateUpscale), sakamakismile (Heretic encoder). This workflow just wires their work together - go like their repos.