Pika H3 Flex LoRA Explained: FL2VA and Ref2VA Adapters, ComfyUI Setup and Community LoRAs

Pika H3 Flex is a LoRA. Pika did not release a new checkpoint: the open release is two LoRA adapters that sit on top of the 33B-parameter MiniMax H3 base and turn its 50-step sampler into a six-step one. That is why so many “Pika H3 Flex LoRA” questions are really about which file goes where, which settings it needs and whether it can be stacked with other LoRAs. This guide answers those from the official sources.

The two official adapters

Together they are about 5.5 GB. The base model they need is about 211 GB. Both live in Pika-Labs/Pika-H3-Flex on Hugging Face; download and install steps are in our GitHub and Hugging Face guide.

What is inside the LoRA

Pika’s technical write-up describes how the adapters were made:

1. Teacher: MiniMax H3, fine-tuned on Pika’s in-house audio-video data, still sampling in 50 steps.
2. Distillation: a six-step student trained with Projected Distribution Matching Distillation (PDMD), a more stable variant of DMD.
3. Reward learning: two more LoRAs on top of the six-step model — ReFL for aesthetics and clarity, DPO for instruction following and realism — then merged into one adapter at no extra inference cost.

So one Flex adapter carries three things at once: the step reduction, Pika’s fine-tune and the reward tuning. More detail in how the 6-step sampler works.

Loading the LoRA in ComfyUI

ComfyUI has official templates (video_pika_h3_flex_t2v, video_pika_h3_flex_fl2va, video_pika_h3_flex_ref2va). If you build the graph yourself, the ComfyUI tutorial sets it up like this:

• Put the BF16 adapters in models/loras/; the base, text encoder and VAEs are shared with ComfyUI’s MiniMax H3 templates.
• Load the adapter with LoraLoaderModelOnly at strength 1.0 after the diffusion model loader.
• Sampler lcm, sigmas from ManualSigmas (six steps, so at least seven sigma values), CFG 1.0 through BasicGuider.

The quantized base files in ComfyUI (int8_convrot transformer and VAE, an NVFP4 text encoder) are what make lower-memory setups possible compared with the full-precision files the official CLI uses.

Loading the LoRA in Diffusers

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "MiniMaxAI/MiniMax-H3", dtype=torch.bfloat16, device_map="cuda"
)
pipe.load_lora_weights("Pika-Labs/Pika-H3-Flex")

Use the six-step settings from the repo’s examples/README.md; Pika notes that generic six-step Euler sampling is not equivalent to its schedule.

Can you stack other LoRAs on Pika H3 Flex?

Not speed LoRAs. Flex is already distilled to six steps. ComfyUI’s guidance is not to stack turbo or “Lightning” LoRAs on top — two acceleration adapters fight each other and the output degrades. Strength other than 1.0 also moves you off the schedule the adapter was trained for.

Style or character LoRAs trained for MiniMax H3 are a different question: they target the same base, so they may load, but nothing official says they keep working after distillation. Test at low strength and compare against Flex alone.

Community LoRAs and what people are reporting

The open release has started a wave of community experiments, mostly shared on X, in ComfyUI guides and on Reddit. Treat all of the following as unverified community reports, not Pika claims:

• A community “Pika-H3-Flex-Turbo-LoRA” has been shared as a single standard LoRA file that claims 2-, 4- and 8-step settings (draft, fast and best quality) without custom nodes.
• Users report running MiniMax H3 with Flex on a 24 GB RTX 3090 using quantized ComfyUI files, with similar video quality to other speed LoRAs but mixed audio quality.
• A Japanese test write-up on note.com measured roughly 3.6× faster generation in ComfyUI after adding the two adapters.

If you try a community LoRA, check its license (it inherits MiniMax H3 Community License terms if it is derived from the model), and compare it with the official adapter on the same seed and prompt.

Can you train your own Pika H3 Flex LoRA?

There are no official training instructions. The GitHub repo and model card cover inference only, and Pika’s training data is proprietary. Pika says it offers customized training as a service (enterprise@pika.art). Community trainers that support MiniMax H3 may work, but a LoRA trained on the 50-step base will not automatically behave the same way under Flex’s six-step schedule.

Too much hassle?

Running LoRAs locally makes sense for research and private pipelines. If you just want clips — with 1080p to 4K upscaling and audio included — use Pika H3 Flex text to video or image to video in the browser.

FAQ

Is Pika H3 Flex a LoRA or a full model? A LoRA. Two adapters (FL2VA and Ref2VA) on the MiniMax H3 base.

Which LoRA file do I need? FL2VA for text to video and start/end frames; Ref2VA for reference-to-video. ComfyUI uses the BF16 files in the comfy/ folder.

What strength should the LoRA use? 1.0, with the six-step ManualSigmas schedule, lcm sampler and CFG 1.0.

Can I use a Lightning or turbo LoRA with it? No — Flex is already a speed adapter, and stacking another breaks the output.