Pika H3 Flex in ComfyUI: 6-Step MiniMax H3 LoRA Setup, Files and Settings

Pika has just released something rare: its own tuned version of an open video model, free to download. Pika H3 Flex is a post-trained MiniMax H3 that renders a full clip with sound in only 6 sampling steps, compared with 20 to 50 for the stock model. It comes as two small LoRA adapters that sit on top of the H3 files you may already have, and ComfyUI added official templates for it on launch day. This guide covers what Pika H3 Flex is, every file you need with its exact ComfyUI folder, the settings that matter, and how it compares with the other H3 speed LoRAs.

What Is Pika H3 Flex?

Pika (Pika Labs) is best known for its hosted video app. With H3 Flex it took MiniMax H3, the 33B open omni model that makes video and audio together, and post-trained it in three stages:

1. Supervised fine-tuning on Pika’s in-house audio-video dataset, at the teacher’s normal 50 steps.
2. Projected Distribution Matching Distillation (PDMD) to bring the step count down from 50 to 6. Pika says standard DMD was unstable in its runs, so it used this newer method.
3. Reward learning with ReFL (for aesthetics and clarity) and DPO (for instruction following, realism and visual quality). These were trained as two LoRAs and merged into one adapter.

The important point for local users is that there is no new architecture. Flex ships as LoRA adapters in the native MiniMax H3 module layout. If you already run H3 in ComfyUI, you only need to download two new files. Pika’s original adapters are FP32 at 2.77 GB each. Comfy converted them to BF16 at 1.96 GB each, and those are the files you want.

The output is the same format as normal H3: 480p or 768p, 11 aspect ratios (21:9 to 9:21), 24 fps, 5 to 15 seconds, with 32 kHz stereo audio mixed into the MP4. Durations round up to the frame grid, so a 5-second request gives 124 frames and 15 seconds gives 362.

The Two Adapters: FL2VA and Ref2VA

H3 has two transformer branches, and Pika trained an adapter for each. You load the one that matches your task.

Ref2VA accepts up to 9 images, 3 videos and 3 audio clips, with 12 files in total, and at least one image or video. Each video or audio clip must be 2 to 15 seconds, and each type is capped at 15 seconds combined. These are the same limits as the stock Ref2VA, which I covered in my H3 reference to video workflow.

Files You Need and Where They Go

All files are on the official Comfy-Org/MiniMax-H3 repo. The base files are shared with the normal H3 templates, so if you already ran H3 you only need the two LoRAs at the bottom of the table.

Your folder tree should end up like this:

ComfyUI/
  models/
    diffusion_models/
      minimax_h3_fl2va_pruned_int8_convrot.safetensors
      minimax_h3_ref2va_pruned_int8_convrot.safetensors
    text_encoders/
      qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    vae/
      minimax_h3_video_vae_int8_convrot.safetensors
      minimax_h3_audio_vae_fp32.safetensors
    loras/
      pika_h3_flex_fl2va_v1_bf16.safetensors
      pika_h3_flex_ref2va_v1_bf16.safetensors

Custom nodes: none are needed. The templates use only core nodes such as LoraLoaderModelOnly, ManualSigmas and BasicGuider, so there is nothing to put in ComfyUI/custom_nodes/. If you want to download from the command line, the Hugging Face CLI works well:

pip install -U "huggingface_hub[cli]"
cd ComfyUI/models
hf download Comfy-Org/MiniMax-H3 loras/pika_h3_flex_fl2va_v1_bf16.safetensors --local-dir ./_h3
hf download Comfy-Org/MiniMax-H3 loras/pika_h3_flex_ref2va_v1_bf16.safetensors --local-dir ./_h3
mv ./_h3/loras/*.safetensors ./loras/

The CLI keeps the repo’s subfolder, which is why the files are moved into ComfyUI/models/loras/ at the end. Do the same for any base file you are missing, moving each into the folder shown in the table.

Step-by-Step Setup in ComfyUI

1. Update ComfyUI. The text-to-video template needs ComfyUI 0.43.2 or newer. The FL2VA and Ref2VA templates work from 0.30.0. Portable users run update/update_comfyui.bat. Git installs run git pull inside the ComfyUI/ folder, then pip install -r requirements.txt. Desktop app users just update the app.
2. Download the files above into their folders, then restart ComfyUI or press R to refresh the model lists.
3. Open the template. Go to Workflow > Browse Templates > Video and pick one of the three: Pika H3 Flex: Text to Video (video_pika_h3_flex_t2v), Pika H3 Flex: Reference to Video (video_pika_h3_flex_ref2va) or Pika H3 Flex: First and Last Frame (video_pika_h3_flex_fl2va).
4. Check the loaders. The diffusion model loader should show the matching pruned int8 base, and the LoraLoaderModelOnly node should show the matching Pika adapter at strength 1.0. FL2VA LoRA goes with the FL2VA base, and Ref2VA LoRA with the Ref2VA base.
5. Write your prompt, add any frames or references, set the length, and click Run.

Recommended Settings

The templates ship with Pika’s settings already in place, so the main job is to not change them by accident.

If you edit the sigma list, keep at least seven values, because seven sigmas give six steps. Community tester WAS reported popping on fine detail, such as distant birds, when dropping to four steps, so six is the safe minimum.

For higher resolution, a community test by NebSH ran a 6-step 960×544 pass, a latent upscale, and a 4-step 1920×1088 pass, then upscaled to 2560×1440, all in 3 minutes 17 seconds. ComfyUI core added H3 latent upscaling on October 28, which makes this two-pass approach easy to build. For a full 1080p 60 fps finish, my RIFE and SeedVR2 workflow works on Flex output too.

Using Flex With Your Old H3 Workflows

You do not have to use the new templates. To add Flex to any older MiniMax H3 workflow:

1. Add a LoraLoaderModelOnly node after the diffusion model loader and pick the matching Pika LoRA at 1.0.
2. Remove any other speed LoRA (turbo, Lightning, PTD, 8-step Acc).
3. Change the sampler to lcm.
4. Replace BasicScheduler with ManualSigmas and paste Pika’s schedule from the Flex template.
5. Set CFG to 1.0. If your workflow uses CFGGuider, switch to BasicGuider.

This matters for workflows like my H3 Extender multi-clip workflow, where every clip is a separate render. Going from 20 steps to 6 for each clip cuts the total time a lot.

Quality: Pika’s Own Benchmarks

Pika compared Flex with stock MiniMax H3 and Fal’s H3 Max at 768p. These are Pika’s numbers, not independent tests, so treat them as a guide.

So the claim is that a 6-step Flex render looks at least as good as a full-step stock render, with alignment roughly level. Pika is also honest about where it loses: it says competitors are still preferred for physics, multi-scene stories, lighting and materials, and action-heavy reference tasks.

On speed, Pika measured its server setup (eight B300 GPUs) going from 3.409 seconds to 2.129 seconds for a 5-second 768p clip with audio, a 1.60x gain from serving tricks alone. On a home PC the bigger win is simply 6 steps instead of 20.

Pika H3 Flex vs Other H3 Speed LoRAs

H3 already had several speed-up LoRAs before Flex. Here is how they compare.

Kijai ran the same text-to-video prompt through the DMAD speed LoRA and Flex at six steps each, and reported DMAD failed that case while Flex did not. djbfilmz compared 6-step Flex with a 20-step run and planned to keep 20 steps for final renders, which is a sensible workflow: use Flex to find the right prompt and seed fast, then do one slow final render. For the full list of H3 add-ons, see my best MiniMax H3 LoRAs list.

Remember the main rule: do not stack Flex with any of the others. Two distillation LoRAs on top of each other break the output. Style and character LoRAs are a different matter, but test them, because they were trained against the undistilled model.

Running It on a 16 GB GPU

Pika does not publish a VRAM figure. Its own CLI was tested on an H100 80 GB, and it says large requests may need --cpu-offload and lots of system RAM. In ComfyUI the files are the same pruned int8 base and nvfp4 text encoder used by the normal H3 templates, so if stock H3 already runs on your card, Flex will too. The LoRA adds about 2 GB of weights.

On a 16 GB card, these steps help:

• Start at 480p and 5 seconds, then move up to 768p.
• Launch ComfyUI with --reserve-vram 1.5 so Windows and the browser keep some headroom.
• Have plenty of system RAM (64 GB is comfortable), because model parts get swapped out of the GPU.
• Close other GPU apps during the render.

My MiniMax H3 out of memory fixes cover this in detail.

Prompting Tips

Comfy’s tutorial includes some useful tips from testing Flex:

• Camera moves: describe the move inside the shot sentence and give it an end point, for example “the camera slowly pushes in and stops on her face”. Exact degrees or cut timestamps can be off by 0.2 to 0.3 seconds.
• Framing: describe framing through content. “A small figure alone in a huge empty hall” works better than just “wide shot”.
• Contact: fine contact, such as a hand stopping a few centimetres from something, is unreliable.
• Text: short labels on products and props stay readable through camera moves, which is useful for ads.
• Ref2VA prompts: use the full reference format with subject_definitions and summary, and label references as <Subject 1>, <Picture 1>, <Video 1> or <Audio 1>.

Prompts can be 1 to 8,000 characters. My MiniMax H3 prompt guide has ready-made templates that work with Flex.

Limitations

• ComfyUI renders were not verified by Pika. Pika’s adapter readme says the file and tensor conversion were checked, but no end-to-end ComfyUI render was verified before upload. Expect small differences from Pika’s own package.
• Local output stops at 768p. 1080p, 2K and 4K upsampling, five separate audio stems (dialogue, score, SFX, foley, ambience) and prompt enhancement are only in Pika’s hosted API.
• Same weaknesses as H3: physics, multi-scene stories and fast action are still not its strength.
• Licence: the adapters are MiniMax H3 model derivatives under the MiniMax H3 Community License Agreement, which has territorial and commercial-use conditions. Read it before paid work.

FAQ

Is Pika H3 Flex free?

Yes. The weights are free on Hugging Face under the MiniMax H3 Community License. Commercial use has conditions, so read the licence first. Pika’s hosted API with 4K output is a separate paid service.

Do I need a custom node for Pika H3 Flex?

No. It uses only core ComfyUI nodes. Update ComfyUI to 0.43.2 or newer and open the Pika H3 Flex templates from the Video section of the template browser.

Where do I put the Pika H3 Flex LoRA files?

In ComfyUI/models/loras/. The base H3 diffusion models go in ComfyUI/models/diffusion_models/, the text encoder in ComfyUI/models/text_encoders/, and both VAEs in ComfyUI/models/vae/.

Can I use Flex with the 8-step Acc LoRA?

No. Both are speed LoRAs, and stacking two distillation LoRAs breaks the output. Use one or the other.

Can I use fewer than 6 steps?

You can try, but testers saw popping on fine detail at 4 steps. Six steps is what the model was trained for.

Which adapter should I use for image to video?

The FL2VA adapter. It handles text to video, first frame, last frame, and first plus last frame. Use Ref2VA only for multi-reference work such as character swap or motion transfer.

Final Thoughts

Pika H3 Flex is the most useful H3 release since the model itself. It cuts render time to roughly a third of a 20-step run, Pika’s own scores say quality holds up or improves, and setup is just two LoRA files and a template. If you already run MiniMax H3, download the two adapters today and use them for every draft render. Keep a slow stock render for the final shot if you notice physics or lighting problems.

More H3 guides:

• Best List of LoRAs for MiniMax H3
• MiniMax H3 Out of Memory Fixes for 16 GB GPUs
• Best Local AI Video Models in 2026

Sources

• Hugging Face: Pika-Labs/Pika-H3-Flex and GitHub inference package
• ComfyUI docs: Pika H3 Flex tutorial
• Pika: H3 Flex technical write-up and launch post
• ComfyUI Wiki: Pika H3 Flex news (October 29, 2026)
• Hugging Face: Comfy-Org/MiniMax-H3