v1.2 adds an optional turbo-LoRA fast path: 4 steps for silent b-roll, 6 steps for anything with audio (dialogue, foley, ambience). Needs the pruned GGUF DiT. Base 20-25 steps stays the quality lane. Full notes in the README.MacMax runs MiniMax H3 locally on a Mac. Video and stereo audio come out of the same pass, so foley and dialogue are generated with the picture, not layered on after.MacMax runs text to video, image to video and first/last frame in one graph, each with native audio.Measured on a 48 GB M5 Pro, ComfyUI 0.30.0. Full docs, every measurement: https://github.com/Bambushu/minimax-h3-mac<h3 id="render-times">Render times</h3>0.5 MP vertical, Spectrum on, chunked VAE, ClipProj encoder. Base int8 path; the turbo GGUF path drops to 4 steps silent / 6 with audio.<pre><code>base, 20 steps:
3s image to video ~14 min
5s text to video ~24 min
5s image to video, chained link ~39 min

turbo, 4 silent / 6 with audio:
3s silent, 4 steps ~7 min
4s spoken, 6 steps ~11 min</code></pre>Cost tracks megapixels x seconds. A first-frame image adds little; duration and resolution are the levers.Sizing: about 22k tokens is comfortable on 48 GB. That is 0.6 MP at 5s, or 1.03 MP at 3s.<h3 id="chaining">Chaining</h3>Clips continue each other: motion carries across the cut, the scene holds, and so does the audio bed. Un-bypass Save Latent on a clip you may want to continue and it writes a small latent; to continue it, un-bypass three more nodes, set two clip indexes, queue.The continuation comes back exactly 22 frames shorter, because those frames are the pinned context and they get trimmed so the files concatenate cleanly. If yours is not 22 frames shorter, chaining did not engage. That is the check worth doing.Nine links ran as one sequence, 39s of continuous scene. The 25s video in the gallery is the front of it, hard concatenated, no crossfades and no level matching, so every join is visible as rendered.Two rules that cost me renders:<ul><li>Write each beat to fill the whole clip. If the action finishes early, the model can fill the rest by cutting to an animated version of one of your reference images. Reseeding does not fix it.</li><li>Clip 1's framing is inherited by everything after it. Lock it before you start.</li></ul><h3 id="smaller-text-encoder">Smaller text encoder</h3>An optional swap replaces the 14.6 GB GGUF encoder with Qwen3-VL-8B fp8 plus a small projection matrix, roughly 4 GB lighter. That is what makes 5s clips fit on 48 GB. It buys length, not speed: same-seed reruns held identity, wardrobe and a spoken line at the same wall clock. Sampling dominates either way.The projection is an approximation and proper nouns are where it shows. Same seed, same prompt, only the encoder changed: ClipProj gave "Apple, Andrian or your dakes top at home", GGUF gave "Apple, Android, or your desktop at home". Keep the shipped GGUF loader for anything with brand names or proper nouns, and swap to ClipProj when you need the length. GGUF is also lighter in practice than its size suggests, since ComfyUI frees it after encoding; ClipProj ships <code>mode: resident</code>, which pins its weights for the whole render.<h3 id="settings">Settings</h3>Spectrum ships on, worth about a quarter of the clock, and faces hold. EasyCache is faster still but smears mouths, so it ships off; fine for faceless b-roll, never alongside Spectrum. 20 steps, sampler <code>euler</code> rather than the stock templates' <code>res_multistep</code>. On unified memory leave <code>history_storage</code> on <code>system_ram</code>.<h3 id="setup">Setup</h3>Four model files, about 41 GB, all linked in the README. Use the GGUF text encoder, the stock one is CUDA only.ComfyUI 0.30.0 in its own checkout, launched with:<pre><code>ASFP8_INT8_EXT=1 python main.py --port 8288 --reserve-vram 10 --cache-none --disable-smart-memory</code></pre>Three node packs are required and two more are optional, for chaining and the smaller encoder. <code>./install_node_packs.sh all</code> clones the lot. Everything the optional packs add ships bypassed, so neither is needed to render.48 GB is what this was measured on. 32 GB works too, reported by users rather than tested here; expect to stay at the shorter durations.<h3 id="limits">Limits</h3>One machine, one config. Chaining is validated at 5s links; longer links are reported elsewhere to fall apart around 15s. Audio carried across ambience beds, not tested on a musical build. No metric here judges audio beyond whether Whisper recovered the words.

MacMax Turbo V1.2 (4-Step Turbo LoRA)

v1.1 adds clip chaining and an optional smaller text encoder, and corrects one claim from v1.0.Chaining. Both workflows now carry a Motion Context block. Un-bypass Save Latent on a clip you may want to continue and it writes a small latent; to continue it, un-bypass three more nodes, set two clip indexes and queue. The continuation comes back exactly 22 frames shorter, which is the free proof it engaged. Nine links ran as one sequence, 39s of continuous scene. The 25s video in the gallery is the front of it, hard concatenated, no crossfades and no level matching.Two rules that cost me renders. Write each beat to fill the whole clip: if the action finishes early, the model can fill the rest by cutting to an animated version of one of your reference images, and reseeding does not fix it. And clip 1's framing is inherited by everything after it, so lock it before you start.Smaller text encoder. An optional swap replaces the 14.6 GB GGUF encoder with Qwen3-VL-8B fp8 plus a small projection matrix, roughly 4 GB lighter. That is what makes 5s clips fit on 48 GB. It buys length, not speed: same-seed reruns held identity, wardrobe and a spoken line at the same wall clock.The correction. v1.0 said latent-space chaining was impossible. The measurement behind that stands, a causal VAE cannot have one latent sliced into a keyframe slot, but the conclusion was wrong: H3 pins a run of frames at their own time coordinates and re-injects them while sampling, which is a different mechanism.Both workflows ship sampler <code>euler</code> rather than the stock templates' <code>res_multistep</code>, and both optional packs ship fully bypassed, so neither is needed to render. Both were rendered end to end from the shipped files before posting.48 GB tested, ComfyUI 0.30.0, Spectrum pinned v0.2.3.

Files with SHA256: f33c83de0b154350564701a01578f1b10f8d0b74e01712b834ae21473d3b48f4