Skip to guide

Pika H3 Flex in ComfyUI: Reference Images to Video

Animate reference images with the MiniMax H3 reference checkpoint and Pika adapter. See the saved settings, model download locations, workflow screenshots and licence notes.

In this guide

Short answer: Animate reference images with the MiniMax H3 reference checkpoint and Pika adapter. The listed files total about 42.04 GB on disk. The repositories list MiniMax H3 Community License Agreement.

Pika H3 Flex in ComfyUI: Reference Images to Video workflow graph in ComfyUI
Workflow graph. Open the image to inspect the nodes.

Output example

This workflow was run on a PC with an RTX 3090 using a photoreal physical robot reference and natural recorded English narration. The same robot fills both picture inputs. A Load Audio node was added in the working copy to the supported standalone audio-reference socket; the saved video retains the model-generated audio.

Photoreal physical robot retains its off-white ceramic facial panels, brushed steel crown and chest, fasteners, blue-grey eyes, articulated dark mouth, exposed neck pistons and shoulder joints. It faces the camera in a stable workshop portrait, with small mouth/jaw movements and a closed-mouth ending. No obvious cuts, added characters or gross deformation in 13 sampled frames and three full-size frames. Audio is present and decodes.

The saved video is 864 × 1184, 158 frames at 24 fps, approximately 6.58 seconds. The selector was 3:4 portrait at 0.98 MP, with six seconds requested. The run kept the LCM sampler, Pika reference LoRA strength 1.0 and saved manual sigma schedule, with seed 405095103756821.

The recording and generated soundtrack passed technical decoding checks.

Input references

Photoreal physical robot with ceramic face panels, brushed steel, visible fasteners, dark articulated mouth and exposed neck mechanisms
Picture 1 and Picture 2 use this same robot as an identity and material reference. This is not a test of blending two different identities.

The robot image was generated for this example using OpenAI imagegen. It was inspected before generation; the article copies are WebP versions of the original 1086 × 1448 PNG.

Audio 1: natural narration reference, 5.25 seconds. Input and output are provided separately for listening comparison.

Recording: The Tale of Peter Rabbit by Beatrix Potter, read by Kevin Devine for LibriVox, marked with an explicit public-domain dedication on its source page. The source interval 24.8–30.05 seconds was converted to mono 24 kHz PCM WAV without voice synthesis. The requested sentence is “They lived with their mother in a sandbank, underneath the root of a very big fir-tree.” Its transcript was aligned automatically; that is not listening review.

The working copy connects LoadAudio 151 → 136.ref_audios.ref_audio_0, with the audio VAE connected. The output soundtrack follows SamplerCustomAdvanced → VAEDecodeAudio → CreateVideo; the source recording was not dubbed onto the result.

Prompt used for this example
subject_definitions:
<Subject 1> is the physical humanoid robot in <Picture 1> and <Picture 2>. Both pictures repeat the same robot identity reference. Preserve its off-white ceramic face plates with fine wear and fasteners, brushed steel crown and chest, blue-grey optical eyes, dark articulated lips and jaw, exposed neck pistons, cables and shoulder joints. This is a manufactured robot with solid mechanical materials.
<Audio 1> is the natural recorded male English narration reference. Use its calm conversational timbre, clear articulation and natural unhurried rhythm for the robot's speaking voice.

summary:
[reference generation] A photorealistic physical robot faces the camera in its softly lit workshop and speaks one short sentence with small coordinated mechanical mouth movements.

retention_analysis:
<Picture 1>: fully_preserved robot identity, materials, face geometry and workshop lighting; framed as a chest-up portrait.
<Picture 2>: fully_preserved duplicate of the same robot; not a second robot or a second location.
<Audio 1>: attribute_transfer natural male speaking timbre and cadence; use as a voice reference for the exact dialogue below.

detailed_description:
[Shot 1] Photographic live-action product cinematography of <Subject 1> standing in the same bright workshop as its references. A frontal chest-up portrait keeps the whole head, exposed neck mechanism and upper shoulder joints visible against softly blurred shelves and windows. Soft window light creates real highlights on brushed steel and subtle shadows along the ceramic panel seams. The camera stays fixed and level for one continuous six-second shot. The robot looks into the lens and speaks in a calm, natural male English voice (S1), using <Audio 1> as the voice reference: <d>[en] They lived with their mother in a sandbank, underneath the root of a very big fir-tree.</d> Its dark mechanical lips and hinged jaw articulate gently with the spoken syllables; the visible neck actuators make only tiny plausible adjustments. Keep the face plate rigid, individual fasteners and panel boundaries stable, and shoulders nearly still. After the sentence, its mouth settles softly closed while it maintains a friendly attentive gaze. Preserve the physical ceramic and metal robot throughout, with no human skin, no human elderly man, no cartoon or toy styling, no extra characters, no scene cuts, no captions and no added text.

overall_soundscape:
One clean natural male speaking voice, close and intelligible, with the timbre and cadence of <Audio 1>. Very quiet workshop room tone underneath. No electronic voice effect, no music and no competing machinery.

non_diegetic_music:
N/A

Observed limits: This single example does not establish voice fidelity or precise lip synchronization. One prompt/seed fixture; duplicated picture references do not test blending distinct identities. Visual review used sampled frames rather than full perceptual playback; fine temporal artifacts may be missed. Selector requested 3:4 at 0.98 MP and six seconds; alignment produced 864 × 1184 and 158 frames (6.583 seconds).

The featured image documents the workflow graph. The output example above comes from a separate completed run, with its recorded prompt and limitations.

Download and load the workflow

Download workflow (JSON)

Drag the downloaded JSON file onto the ComfyUI canvas to load it.

Install the model files and any custom nodes listed below. If the workflow uses input images, video or audio, select your own files in the input nodes before running it.

Which files go where

Place each file under your ComfyUI directory in the folder shown. A text encoder turns the prompt into model conditioning. A VAE decodes the sampled representation into image or audio data. A LoRA is an adapter loaded alongside the base model.

PartFileSizeFolder
vaeminimax_h3_video_vae_int8_convrot.safetensors2.81 GBmodels/vae
vaeminimax_h3_audio_vae_fp32.safetensors0.61 GBmodels/vae
diffusion_modelsminimax_h3_ref2va_pruned_int8_convrot.safetensors20.97 GBmodels/diffusion_models
text_encodersqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors15.69 GBmodels/text_encoders
loraspika_h3_flex_ref2va_v1_bf16.safetensors1.96 GBmodels/loras

The listed files total about 42.04 GB on disk. Disk size is not a measured VRAM requirement. Reuse matching files already installed rather than downloading a second copy.

Settings in this workflow

ControlSaved value
Samplerlcm
LoRA strength1.0
Duration5 seconds
Frame rate24 fps
Reference node dimensions1344 × 768
Frame count124
Reference sizingmatch

Saved sigma schedule: 1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0.

How to use it

  1. Load your two reference images into the Picture 1 and Picture 2 inputs.
  2. In the prompt, refer to each picture by its connection order, using <Picture 1> and <Picture 2>. Describe the action, camera and sound.
  3. Keep the reference-to-video checkpoint and matching Pika ref2va LoRA selected. Set the duration and output size.
  4. Generate when you are ready, then inspect the saved video and its audio.

What to check before changing the graph

The reference checkpoint and LoRA are a pair. The fl2va adapter from the text-to-video workflow is a different file. Keep the two image connections in the same order as the picture tags in your prompt. Use references you have permission to use, and obtain consent for a person’s likeness or voice.

Licence

The repositories list MiniMax H3 Community License Agreement. This is a community licence with conditions. Read the current terms before commercial use; a public download does not mean unrestricted use. See our model licence tracker for the wider collection.

Troubleshooting

ProblemLikely causeFix
Missing nodesYour installation differs from the captured environmentUpdate ComfyUI and reload the template before reconnecting nodes.
Model missing from a dropdownFile is in the wrong folder or the list is staleCheck the exact filename and folder above, then refresh the model list.
Unexpected output size or lengthA connected input overrides an internal widgetTrace the size and duration connections from the outer graph.

Where to go next

FAQ

Are these output examples?

The featured image shows the workflow graph. The Output example section shows the actual saved generation from a PC with an RTX 3090.

Can I change the saved settings?

Yes. Keep the documented configuration as a reference and change one control at a time. Samplers, adapters and noise schedules work together, so treat a replacement recipe as a separate configuration.

Sources

Keep building

Get the cheat sheets and planners or browse all ComfyUI workflows.

Need help building a production workflow? Talk to Automate Digital.