Generate Video with Sound in ComfyUI (Workflow Included) – MiniMax H3 Text-to-Video
Most video models generate silent clips. You add music in an editor afterward and hope it lines up. MiniMax H3 doesn’t work that way — it generates the picture and the sound together, in the same pass, so footsteps land on the frame where the foot lands. This is the text-to-video workflow for it in…
Annons
Annons
Most video models generate silent clips. You add music in an editor afterward and hope it lines up. MiniMax H3 doesn’t work that way — it generates the picture and the sound together, in the same pass, so footsteps land on the frame where the foot lands. This is the text-to-video workflow for it in ComfyUI. You type a description, and you get back a 5-second clip at 24fps with stereo audio already mixed in. Click to watch: four shots generated in a single 5-second pass, with the score and sound effects generated alongside them. In this tutorial, you will learn how to: Set up the MiniMax H3 text-to-video workflow in ComfyUI. Write a timecoded storyboard prompt that produces multiple shots in one clip. Control resolution and duration without hitting the model’s grid limits. Trade quality for speed with the optional Lightning LoRA. Software needed We will use ComfyUI, a free AI image and video generator. You can use it on Windows, Mac, or Google Colab. Think Diffusion provides an online ComfyUI service if you don’t have a capable local GPU. What is MiniMax H3? MiniMax H3 is MiniMax’s general-purpose, omni-modal model. Instead of treating video and audio as separate problems, it denoises visual tokens and audio latents together, so dialogue, sound effects and music stay aligned with what is happening on screen. Key highlights: Native stereo audio: sound is generated in the same forward pass as the picture, not layered on afterward. Multiple shots per clip: a single generation can cut between several shots if your prompt describes them. 24fps, up to about 15 seconds: the native cadence matches standard editing timelines. Reference inputs: the same model family accepts up to 9 reference images, 3 video clips and 3 audio tracks for identity and motion transfer. One practical note before you start: the unquantized weights are enormous. The bf16 checkpoint is about 37 GB and the full bf16 text encoder about 48 GB. The workflow here uses the int8-pruned builds instead, roughly 19.5 GB for the checkpoint and 14.6 GB for the text encoder, which is what makes local generation realistic on a consumer card. Quality is close, and the difference matters far less than your prompt. MiniMax H3 text-to-video workflow Step 0: Update ComfyUI MiniMax H3 support landed in ComfyUI 0.30.0, and the bundled template and prompt embeddings need a newer build than that. Update before you start. Click the Manager button on the top toolbar. Select Update ComfyUI. Restart ComfyUI. Good news on custom nodes: this workflow needs none. Every node in it — the resolution selector, the math expression, the switches — ships with ComfyUI itself. If you see red-bordered nodes after loading, your ComfyUI is simply out of date; update it rather than hunting for a node pack. Step 1: Download the models Download these four files and put them in their folders: Download minimax_h3_fl2va_pruned_int8_convrot.safetensors and put it in ComfyUI > models > diffusion_models. Download qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors and put it in ComfyUI > models > text_encoders. Download minimax_h3_video_vae_int8_convrot.safetensors and put it in ComfyUI > models > vae. Download minimax_h3_audio_vae_fp32.safetensors and put it in ComfyUI > models > vae. Note the two VAEs. H3 has a separate decoder for picture and for sound, and the workflow loads both. Miss the audio one and the clip comes out silent. (Optional) For the speed trick in Tips: Download minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors and put it in ComfyUI > models > loras. (Optional) For style shortcuts: Download any of the 10 style embeddings (minimaxh3_*.safetensors) and put them in ComfyUI > models > embeddings. Step 2: Load the workflow Download the workflow JSON file below. Drop it into ComfyUI to load. Download The demo prompt is already loaded in it, so you can run it as-is to reproduce the clip above, then swap in your own scene. The workflow is short because most of it is folded into a single subgraph node. Double-click it to see the sampler stack inside: the two VAE loaders, the text encoder, the diffusion model, the sampler, and the two decoders that split the result back into picture and sound. Step 3: Set the resolution and duration H3 is strict about canvas sizes, so the workflow computes them for you instead of letting you type any number. Open the Resolution Selector node. Set the aspect ratio to 16:9 (Widescreen) and the megapixels to 0.4. That gives you 864×480. Leave the multiple at 32. H3’s latent grid is built on multiples of 32, and rounding to anything else produces artifacts. Keep the megapixels at or below 1.0 for the generation pass. Past that the model degrades instead of improving. To go bigger, generate at 768p and upscale afterward (see Tips). Duration is the Float (duration) input on the subgraph node, in seconds. Set it to 5 for the demo. Step 4: Write the prompt This is where H3 differs most from an image model. It reads your prompt as a storyboard, not a caption, and it expects timecodes. Here are four tips for writing a successful prompt for the Minimax H4 model. Scene header. The first paragraph sets film stock, lens and grading once, so you don’t repeat it in every shot. Timecoded shots. [0s-1.5s] and so on. Overlapping ranges are fine and give you the rapid-cut feel. Named camera moves. “Extreme wide”, “low angle from the valley floor”. Concrete cinematography. Vague style words do much less. Audio described as action. Because sound is generated alongside the picture, describe what makes the noise (“cracking ice”, “the deep boom of each wingbeat”) rather than naming a sound file. Then add a separate line for the score, which the model treats as non-diegetic. Step 5: Generate the video Click the Run button. The first run loads about 37 GB of weights from disk, so give it time before you assume something is wrong. The result lands in the Save Video node. Click it to play the clip. The four shots the prompt asked for, in order. The clip arrives with stereo audio already attached — no separate audio file, no lining anything up. Settings SettingValue in the workflowEffectAspect ratio16:9Set in the Resolution Selector.Megapixels0.4864×480. Up to 0.98 (1344×768) is safe.Multiple32Leave it. Must match the latent grid.Duration5 secondsSnapped up to 124 frames.Steps208 when the Lightning LoRA is on.Sampler / schedulerres_multistep / simpleDenoise1.0Full generation.FPS24Native cadence of the model.Seed757358688076805 (fixed)Change it for a new take. Tips Iterate at low resolution first. Drop megapixels to 0.2 and generate several takes to test your storyboard and camera moves. Once the timing works, re-run at 0.4 or higher and keep the seed. Prompt changes are the slow part, not the resolution. Use the turbo LoRA to go faster. Set turbo toggle to true to use the LoRA. This will change the number of sampling steps from 20 to 8, reducing the generation time to half. Go bigger in a second pass, not the first. Generate at 768p, then upscale. Pushing megapixels past 1.0 in the generation pass costs you structure, not gains detail. Add a style embedding. With an embedding in ComfyUI > models > embeddings, reference it in the prompt by filename without the extension, e.g. embedding:minimaxh3_art_is_explosion. It’s a one-token shortcut for a whole look. Optional — the workflow runs fine without one. Change the seed for a new take. Everything else can stay identical; a new seed gives you a different performance of the same scene. Keep the negative instructions. That last paragraph of the prompt (no text, no logos, no cartoon look, etc) is acting as the negative prompt. Leave it in and adjust it to your scene rather than deleting it. What’s next Wan 2.2 text-to-video LTX Video Hunyuan Video MiniMax Video-01 Text-to-video: the basics AI upscalers — for the second pass at 768p and beyond. Video-to-video The post Generate Video with Sound in ComfyUI (Workflow Included) – MiniMax H3 Text-to-Video first appeared on Stable Diffusion Art.