<p><strong>An i2v / t2v MiniMax H3 workflow tuned for maximum speed on a single 24 GB card, with native stereo audio in the same pass.</strong> 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.</p>
<h3>What is inside</h3>
<ul>
<li><strong>10Eros-Max base (INT8)</strong> - NSFW-focused finetune of MiniMax H3 by cicalooo, int8 convrot <code>skip_edges</code> variant (first/last blocks kept in BF16, the quality-safe option). Same loader as the official base: swap one filename to go back SFW.</li>
<li><strong>4-step Turbo distill LoRA</strong> (larryvrh, <code>v4_step600_ema</code>, 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.</li>
<li><strong>Sparse attention</strong> (Zironic's H3-Optimizations, Advanced node): author-measured 1.57-1.69x on 243-frame clips, Kitchen INT8 backend, zero extra dependencies.</li>
<li><strong>SageAttention patch</strong> (KJNodes): another ~20% off sampling in our A/B.</li>
</ul>
<h3>Downloads - models</h3>
<p>Put each file in the listed ComfyUI folder.</p>
<ul>
<li><strong>models/diffusion_models</strong> (pick one base):<br />
<a target="_blank" href="https://huggingface.co/cicalooo/10Eros-Max" rel="ugc">10Eros-Max INT8 skip_edges</a> (21.7 GB, gated repo: log in to HuggingFace and accept the terms; file <code>10Eros_Max_h3_fl2va_beta2_pruned_int8_convrot_skip_edges.safetensors</code>)<br />
or the <a target="_blank" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors" rel="ugc">official INT8 base</a> (21 GB, SFW drop-in)</li>
<li><strong>models/loras</strong>: <a target="_blank" href="https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/blob/main/minimax_h3_turbo_v4_step600_ema.safetensors" rel="ugc">minimax_h3_turbo_v4_step600_ema.safetensors</a> (780 MB)</li>
<li><strong>models/text_encoders</strong>: <a target="_blank" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors" rel="ugc">qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors</a> (15.7 GB)</li>
<li><strong>models/vae</strong>: <a target="_blank" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_video_vae_fp16.safetensors" rel="ugc">video VAE fp16</a> (5.2 GB) + <a target="_blank" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors" rel="ugc">audio VAE fp32</a> (0.6 GB)</li>
</ul>
<h3>Downloads - custom nodes</h3>
<ul>
<li><a target="_blank" href="https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo" rel="ugc">ComfyUI-MiniMax-H3-Turbo</a> (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.</li>
<li><a target="_blank" href="https://github.com/Zironic/H3-Optimizations" rel="ugc">H3-Optimizations</a> (Zironic) - H3SparseAttentionAdvanced.</li>
<li><a target="_blank" href="https://github.com/kijai/ComfyUI-KJNodes" rel="ugc">ComfyUI-KJNodes</a> (kijai) - SageAttention patch node. Needs the <code>sageattention</code> pip package; cannot install it? Set the node to <code>disabled</code>, you only lose ~20%.</li>
</ul>
<h3>Requirements</h3>
<ul>
<li>ComfyUI <strong>0.30+</strong> (native INT8 convrot support).</li>
<li>Tested on a <strong>24 GB</strong> 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.</li>
<li>Frame count must satisfy <code>count mod 17 == 5</code>: 56, 73, 90, ... 243. 56 frames = 2.3 s at 24 fps.</li>
<li>Stay at or under <strong>0.5 MP</strong>; go big with the latent hi-res pass below instead.</li>
</ul>
<h3>Settings that matter</h3>
<ul>
<li><strong>Steps 6</strong> shipped (4 = fastest and the author's minimum; 6-8 looks noticeably better; above 8 it over-sharpens).</li>
<li><strong>Scheduler simple</strong> with the Turbo LoRA (switch to beta + ~20 steps if you remove it; node 17 is left in the graph to rewire).</li>
<li><strong>Sparse level</strong> shipped at the author defaults (budget 0.3, edges 2@0.5). 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.</li>
<li><strong>Prompt recipe</strong> H3 responds to: style, then <code>&lt;Picture 1&gt;</code> + subject, the action, ONE camera move, then an <code>Audio:</code> block (voices, sfx, music - the audio is generated from your prompt in the same pass).</li>
<li><strong>t2v</strong>: the first_frame input of the MiniMax node is optional - drop the image and prompt only.</li>
</ul>
<h3>Going further (optional)</h3>
<ul>
<li><a target="_blank" href="https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale" rel="ugc">Comfyui-MMH3-UltimateUpscale</a> (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.</li>
<li><a target="_blank" href="https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4" rel="ugc">Heretic NVFP4 text encoder</a> (sakamakismile): uncensored drop-in replacement for the official encoder (same folder, swap <code>clip_name</code> in node 13) if it refuses your prompts.</li>
</ul>
<h3>Credits</h3>
<p>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.</p>

MiniMax H3 Ultra Boost - 4-step Turbo / INT8 / Sparse Attention (10Eros base)

<p>Fixes a silent regression. H3-Optimizations 0.2.16 (released 2026-08-27) added a guard that refuses to compose sparse attention with an attention override it does not own. v1 fed the sparse node from the SageAttention patch (601 -&gt; 600 -&gt; 810), so on 0.2.16+ sparse attention was silently DISABLED - no error, no failed run, just this console warning and none of the speedup:</p><p>[H3 Optimizations] SPARSE ATTENTION FELL BACK to existing. Reason: preserved an unknown optimized-attention override with full-Q single-call semantics; sparse attention is disabled because the explicit external consumer does not expose an H3 sparse composition contract</p><p>v2 wires the sparse node straight to the Turbo LoRA (601 -&gt; 810). You lose nothing: when sparse attention is active it owns attention anyway, so Sage was already being replaced on older versions too. Node 600 stays in the file, fed by 601, as a spare for the no-sparse fallback documented in the README.</p><p>Also fixed: the 10Eros download link. v1 pointed at cicalooo/10Eros-Max, which is gated and returns 401. The INT8 convrot weights live in the ungated cicalooo/10Eros-Max-h3-int8-convrot repo. Every link in the README was re-checked.</p><p>Verified on a real run: the pack now logs attention=sparse_kitchen_int8 with no fallback. Same defaults as v1 (6 steps, scheduler simple, sparse budget 0.3 / edges 2@0.5, 0.3 MP, 56 frames). Thanks to Velant and yajukun for the reports.</p>

Files with SHA256: 0fa10b2b5b3aad150c822854ecaeec68bf1b4a38d30f6e9ccd69d1118a3cfbf9