Text to video
Describe the scene, action and camera movement. Generate an AI video from your words.
The camera follows her
across the rooftop.”
Create with Seedance 2.5, MiniMax H3 official API or open MiniMax H3 BF16. Generate video from text, images or reference media, with model hosting handled for you.
Sign in to create · Generation is billed by model and plan

SHOT 01
SHOT 03Generate video from a prompt, animate a still image, or guide a shot with reference media. koma brings these generation modes into one workspace.
Describe the scene, action and camera movement. Generate an AI video from your words.
Bring a still image into motion. Use a first frame, or first and last frames, to guide the shot.
Guide generation with images, video or audio. Supported media types depend on the model.
Input modes vary by model. Compare supported models below. Images are creative illustrations.
Seedance 2.5, MiniMax H3 official API, and open MiniMax H3 BF16. Each supports Reference to Video. Compare official APIs by generation cost, or let koma handle the machines and setup for the open model.

From US$0.16/sec480p · Rates vary by output settings
Text, frames and multimedia references.

From US$0.12/sec768P · A lower starting cost
Multimedia references. A lower starting cost.

Setup by koma.Official open weights · BF16
Machines and workflows, ready to use.

About US$0.97 per 15 s480p · generated at fixed 6 steps
Try different takes in 6 steps.
Starting prices are before tax, at the resolutions shown. The MiniMax H3 official API includes 5 image references; additional images cost US$0.06 each. Illustrations show creative styles. See all options and rates in Plans.
Seedance 2.5 and MiniMax H3 official APIs, directly in koma. Start with a prompt, define frames or bring reference media into your next shot.
IMAGE REFERENCE
SCENE & STYLE
A scene, an action, a camera move. Start from a written prompt.
Define the opening and ending images of your shot.
Give characters, settings and visual style a clear reference.
Bring motion and sound into the generation direction.
Both official models offer these modes. Input combinations follow the selected mode; clip lengths and resolutions are listed above.
From sourcing machines and installing ComfyUI to loading weights, connecting nodes and testing workflows, koma takes care of the setup. The open model becomes a tool you can create with.
Source and configure GPUs, storage and the runtime environment.
Install ComfyUI, its dependencies and the model weights.
Configure the nodes and connect the MiniMax H3 generation workflow.
Initialize the environment and verify it with a test generation.
Shared generation or dedicated compute. Create directly in koma.
Your focus: the film.
The setup is on us.
Official BF16 weights for the generation model and text encoder.
BF16 is a 16-bit floating-point format. The generation model and text encoder use official BF16 weights, without lower-bit quantization or distilled few-step checkpoints. koma prepares the environment they need.
Choose 1–100 steps, with 20 as the default. Adjust the settings for your shot.
A first experiment, a recurring project or a focused production session. Find the option that fits the way you create.

Start small. Try a few clips and different models.
Open MiniMax H3 BF16 bills on compute time: US$0.30 per compute minute, about US$3.00 for 15 s at 480P estimated at the default 20 steps, prepaid on submit and settled after completion. Or choose an official API for the inputs you need.

Social posts, campaigns and stories on a regular schedule.
Choose around your recurring usage and included benefits. Seedance 2.5 and MiniMax H3 official APIs cover text and multimodal inputs.

Short films and multi-shot projects with focused iteration.
US$5.90 for 3 hours. Exclusive RTX 5090 compute, with ComfyUI and the MiniMax H3 workflow prepared for you. No additional per-video charge during the session.
Prices are in US dollars before tax. Dedicated time runs continuously after successful deployment and a test video; charging starts at that point. Official APIs are billed separately. See Plans for eligibility and full terms.
A few practical details
about models, compute and creating.
No. For open MiniMax H3 generation, koma prepares the compute environment, deploys ComfyUI, loads the model and sets up its workflow. You prepare your references and prompts, submit generations and collect your videos in koma. This is a prepared MiniMax H3 BF16 workflow; uploading custom ComfyUI workflows is not supported.
The MiniMax H3 generation model and text encoder use the official BF16 weights. BF16 is a 16-bit floating-point format. koma does not replace these with lower-bit quantized weights or a distilled few-step checkpoint. Available output settings follow the model configuration.
This MiniMax H3 BF16 workflow currently uses reference images or reference videos plus a prompt: up to 9 images and 3 videos (2–15 seconds each, 15 seconds in total). For text-only, first/last-frame or standalone audio references, choose the matching inputs in Seedance 2.5 or the MiniMax H3 official API.
With koma CLI, an AI assistant that can run commands can work with your scripts and references, prepare shots, submit jobs and download results. You can continue choosing takes in the browser. Video generation still runs on the selected cloud service.
Per-generation use runs through shared services: official APIs bill by video second; open-source models bill on actual compute time, prepaid when you submit and settled on actual usage afterwards. Dedicated compute gives you an exclusive MiniMax H3 BF16 machine for a continuous time block, with no additional per-video charges during that session.
AI script text collaboration is free, and each new creation includes 4 bonus images with its first generation. Later image generation and video generation are billed according to the model, settings and plan; open models also require compute. Sign in to create, and see Plans for current benefits and rates.