MiniMax H3 ComfyUI Guide (2026): Every Workflow, File and Fix

MiniMax H3 is the open video model I have spent the most time with in ComfyUI. It gives you video and synchronized stereo audio from one sampling pass, it handles text-to-video, image-to-video, first/last frame and reference-to-video, and the community has built a whole toolbox around it: Turbo and Acc LoRAs, GGUF files, frame interpolation, upscalers, an extender for longer films and a section editor. The catch is that the pieces are scattered across a dozen repos and node packs. This page is the map. It tells you which H3 workflow fits your goal, which files you need, how to make it run on a 16 GB GPU, and links to my detailed guide for every step.

What Is MiniMax H3?

MiniMax H3 is an open-weights video generation model that outputs a video track and a matching stereo audio track from the same sampling run. Footsteps, rain, room tone, music beds and short spoken lines come out in sync with the picture, so you do not need a separate audio model or a lip-sync pass for most clips.

In ComfyUI it ships in a few flavours, and the names confuse people:

• FL2VA (first/last frame to video + audio): the weights you use for image-to-video, first/last frame and plain text-to-video. For text-to-video you simply leave the frame inputs empty. ComfyUI’s own template does the same.
• Ref2VA (reference to video + audio): give it one or more reference images of a character or product and it keeps them consistent across the clip.
• Pruned / int8 / W4A8 / GGUF: smaller versions of the same model. The full files are tens of gigabytes; the pruned W4A8 FL2VA file is about 12.5 GB and the GGUF Q4_K_M is about 11.4 GB.

The default graph runs 20 steps, which is slow on consumer cards. Almost everything people build on top of H3 is about making it faster, fitting it in 16 GB, or getting longer and sharper results. That is what the rest of this guide is organised around.

Which MiniMax H3 Workflow Do You Need?

Start from what you want to make, then open the matching guide. Each one has the exact files, folders, settings and a free workflow JSON.

Core Files and Where They Go

These are the files most H3 workflows on this site use. You do not need all of them: pick one diffusion model, one text encoder route, both VAEs, and the speed LoRA for your chosen method.

The clip_projections folder is created by ComfyUI-ClipProj the first time ComfyUI starts with it installed. If a shared workflow asks for a path like minimax\qwen3vl-32b-minimax-h3-q2_k.gguf, the author kept files in a subfolder: either re-pick the file in the loader or create ComfyUI/models/text_encoders/minimax/. If you are unsure where any file goes, my ComfyUI model folders cheat sheet covers every folder.

The W4A8 model needs a ComfyUI build newer than 7 August 2026, when asym_w4a8_int8 support landed in mainline. Update ComfyUI before you load it.

Custom Node Packs

Install these through ComfyUI-Manager (search the pack name), or with git clone into ComfyUI/custom_nodes/ followed by pip install -r requirements.txt inside the pack folder using ComfyUI’s own Python:

• ComfyUI-MiniMax-H3-Turbo (Larryvrh): MiniMaxH3TurboLoRA, MiniMaxH3TurboSampler.
• ComfyUI-ClipProj (nicolab28): ClipProjLoader for the 4B encoder + projection route.
• ComfyUI-GGUF (city96): UnetLoaderGGUF, CLIPLoaderGGUF for GGUF files. See my ComfyUI-GGUF guide.
• ComfyUI-KJNodes: DiffusionModelLoaderKJ, PathchSageAttentionKJ.
• ComfyUI-VideoHelperSuite: VHS_VideoCombine to save MP4 with the audio muxed in.
• ComfyUI-Easy-Use: VRAM clean-up nodes used between stages.
• ComfyUI-Frame-Interpolation for RIFE, and SeedVR2 for 1080p upscaling (finishing guides below).

Most of these are in my list of must-have ComfyUI custom nodes.

Speed Options: Turbo, Acc, Pika H3 Flex, FastH3 Omni and GGUF

The stock 20 steps are too slow for daily use. These are the routes that cut steps, and when to pick each.

Two mistakes I see a lot: loading the Turbo LoRA through a plain LoraLoaderModelOnly (it must go through MiniMaxH3TurboLoRA), and using the beta scheduler with the Turbo sampler (use simple). Also note that GGUF is not automatically faster: file size does not predict H3 speed, so measure on your own card before you switch.

New to quant formats? GGUF quantization explained covers what Q4_K_M and Q2_K mean.

Running MiniMax H3 on 16 GB VRAM

H3 runs on 16 GB cards such as the RTX 5080, 4080 and 4070 Ti Super, but only with the right choices:

1. Use a pruned W4A8 or int8 model, not the full bf16 weights.
2. Use the 4B encoder + ClipProj (2.8 GB) instead of the 32B encoder (15.7 GB). This is the single biggest saving.
3. Keep the default 0.4 MP (864×480 at 16:9) for the first pass, and upscale afterwards.
4. Have plenty of system RAM. H3 streams weights from RAM when they do not fit in VRAM. 64 GB is comfortable; 32 GB works for short clips.
5. Clear VRAM between stages (sampling, VAE decode, upscale) with the Easy-Use clean-up nodes.
6. Install SageAttention for faster attention, or disable the Sage node if you do not have it.

If you still get torch.OutOfMemoryError or Allocation on device, work through my MiniMax H3 out of memory fixes, which go stage by stage: model load, text encoder, sampler and VAE decode. For a wider view of what fits in each VRAM tier, see how much VRAM you need for ComfyUI. If ComfyUI drops its connection mid-render, that is usually RAM running out; see ComfyUI stuck on Reconnecting.

Prompting MiniMax H3

H3 follows a structured prompt better than a loose sentence. Describe, in order: the subject, the action, the setting, the camera move, the lighting and style, and then the sound (ambient sound, effects and any dialogue). Put spoken lines in quotes and say who speaks them. Keep one main action per clip; long lists of events tend to blur together in a 6-second shot.

My MiniMax H3 prompt guide has copy-paste templates for T2V, I2V, FL2V and reference-to-video, a list of camera moves that work, dialogue tags and fixes for common prompt problems.

LoRAs for Style, Camera and Characters

Beyond the speed LoRAs, H3 has a growing set of LoRAs for realism, camera motion (including 360 orbits), retro looks such as VHS and 1930s animation, VR180 and character swap. All of them go in ComfyUI/models/loras/. Stack the speed LoRA first, then one style or camera LoRA at a time, and lower style strength if motion breaks. The full list with links, trigger words and strengths is in the best MiniMax H3 LoRAs.

Smoother and Sharper: FPS and Upscaling

H3 renders at 24 fps and around 0.4 MP on 16 GB. Three finishing routes, all of which keep the audio in sync:

• 72 fps: RIFE VFI x3, with an optional 1080p upscale. MiniMax H3 RIFE workflow.
• Exact 1080p 60 fps: RIFE x5 then keep every 2nd frame, plus SeedVR2 for the upscale. 1080p 60 fps workflow.
• 2K and 4K: the native latent upscaler, the fast Sol-H3-VAE and NVIDIA’s Open-H3-Regenerate-2K refiner in a two-pass graph. 2K/4K upscale guide.

Pixel upscalers (SeedVR2) sharpen what is already there; the latent route re-generates detail at the higher size, which looks better but costs more time and VRAM.

Longer Videos and Editing

A single H3 clip is short. Two tools get you further:

• Extender: turns one or more reference images into several connected clips with the same character, continuous motion and one audio track. MiniMax H3 Extender workflow.
• Section editor (Frame Infill): re-generates only part of an existing video, keeps the original audio and can use a mask. MiniMax H3 section editor.

For still images of a consistent character (thumbnails, storyboards), the reference-to-image workflow runs Ref2VA with a length of 5 frames and picks one with ImageFromBatch.

Common Errors and Fixes

For general ComfyUI errors (CUDA, torch, missing modules), see ComfyUI errors and fixes.

Frequently Asked Questions

Is MiniMax H3 free to run locally?

Yes. The weights are open and run in ComfyUI on your own GPU. You only pay for hardware and electricity.

How much VRAM does MiniMax H3 need?

16 GB is the practical minimum for comfortable use with a pruned W4A8 or int8 model and the 4B encoder + ClipProj. Plenty of system RAM (64 GB) helps, because H3 streams weights from RAM.

Which file do I use for text-to-video?

The FL2VA weights, for example minimax_h3_fl2va_pruned-w4a8_convrot_pruned.safetensors, with no first or last frame connected.

Does MiniMax H3 make audio?

Yes. It generates synchronized stereo audio in the same pass. Decode it with minimax_h3_audio_vae_fp32.safetensors and save it with the video.

Turbo LoRA, Acc LoRA or Pika H3 Flex: which is best?

Turbo (6 steps) is the everyday choice for T2V and I2V. The official Acc LoRA (8 steps) is a safe pick for Ref2VA and FL2VA. Pika H3 Flex is a single 6-step LoRA that some people prefer for motion. Try the one that matches your workflow first.

Is GGUF faster than safetensors for H3?

Not necessarily. GGUF saves disk and sometimes memory, but an int8 safetensors model can sample faster. Measure on your own card.

Where do MiniMax H3 LoRAs go?

In ComfyUI/models/loras/. The Turbo LoRA loads through MiniMaxH3TurboLoRA; style and camera LoRAs load with a normal LoRA loader.

Final Thoughts

If you are new to H3, do it in this order: set up the text-to-video workflow with the Turbo LoRA, then try image-to-video and reference-to-video, and only then add finishing (RIFE, SeedVR2 or the latent upscaler) and the Extender for longer films. Keep the 4B encoder + ClipProj on 16 GB cards, keep the scheduler on simple, and bookmark this page: I update it whenever a new H3 tool lands.

Comparing H3 with other open video models? See the best local AI video models, the Wan 2.2 16 GB setup and the LTX-2.5 16 GB setup.

Sources

• Comfy-Org/MiniMax-H3 (official ComfyUI repack: FL2VA, VAEs, 32B encoder)
• Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI (pruned W4A8)
• leejet/MiniMax-H3-GGUF and realrebelai/MiniMax-H3_GGUFs
• larryvrh/MiniMax-H3-Turbo-Lora
• NicoLab28/ClipProj-MiniMax-H3 and Merserk/qwen3vl-4b-int4-convrot