Open-source generative AI workflow tooling

Generate images and video
with your AI agent.

Your agent calls one API to train a character LoRA, generate images, and cut them into video — self-hosted on your own GPU, no cloud bill, no ComfyUI graph to wire by hand.

Get the repoRead the docs
POST /api/image/generate
Astronaut on the moon — an SDXL output from the API call above
One call, no ComfyUI graph. See more →
Agent-driven API
Your agent describes a shot through the API. No ComfyUI graph to wire by hand.
Self-hosted, $0 per shot
Runs on your own GPU — no cloud bill, no data leaves it.
Runs on modest hardware
Tuned for a single consumer GPU — no data center, no exotic rig.

What you can make

Every one of these came out of a self-hosted FlixML install on a single consumer GPU. The stills are SDXL, the clip is Wan 2.2. FlixML also drives lip-sync (InfiniteTalk), instruction edits (Qwen) and FLUX.2 — full breakdown, VRAM included, in the model directory.

Rain-soaked neon city street at night, reflections on wet asphalt

SDXL

Cinematic scene from a text prompt.

Mountain valley at golden hour with mist between the pines

SDXL

Landscape, no reference image.

Sunlit architectural atrium in concrete and timber

SDXL

Architectural interior, hard daylight.

Extreme macro of a dew-covered spider web backlit at dawn

SDXL

Macro detail, backlit.

Armoured dragon on a rooftop in a neon-lit city at night

SDXL

Creature work, neon-lit night scene.

Wan 2.2

That still, animated — image in, video out.

FlixML Studio

Studio is the app: a gallery, a workflow catalog, and one HTTP API your agent talks to. You ask for a shot in plain language, it runs the workflow on your own GPUs, and the result lands in your gallery. Open source under AGPL-3.0 and free to run — we run it here every day, and every generated image on this page came out of it.

View the repoInstall it
FlixML Studio: a chat where a user asks for a lighthouse shot, beside the Studio gallery filling with the generated stills and a workflow catalog listing what's installed.

Workflows

14 workflows ship with Studio, across 5 kinds of job: text to image, image to image, text to video, image to video, and video to video. Each has a page with its parameters, the VRAM it needs and what a run costs.

All workflows
A glazed ceramic teapot on a wooden table beside a window in morning light.

Text → Image

SDXL Base Image

sdxl_base
A blue and white porcelain teapot on the same table and in the same light as the brown one it was generated from.

Image → Image

SDXL Image-to-Image

sdxl_img2img
The blue and white porcelain teapot seen from a front-right quarter angle, spout left and handle right, on the same table.

Image → Image

Qwen Multi-Angle (re-angle a still)

qwen_multiangle

Image → Video

Wan 2.2 Image-to-Video

wan22_i2v

Guides

How to run the workflows behind image and video generation, model by model — plus what's moving in the wider model landscape.

All articles
A studio microphone on a dark desk beside a glowing green audio waveform.
Local lip-sync with InfiniteTalk
Run it locally with GGUF quantization for talking characters at $0 a shot.
Read guide
A dark green wave caught mid-break, spray frozen along its crest.
Animating stills with Wan 2.2 I2V
Turn a single image into a coherent clip, and stop the face from drifting between frames.
Read guide
Green and black network cables tangled on a wet dark surface, two connectors lit in the foreground.
Why FlixML drives ComfyUI with an API, not a node graph
How an AI agent describes a shot instead of wiring the graph by hand.
Read guide
A consumer graphics card on a workbench under hard light, memory chips in sharp detail.
MiniMax-H3: what the model everyone's running actually requires
Its ComfyUI build pulled 19.2M downloads in a month, and it needs 15GB+ of VRAM. Here's the real math, and why that matters for what you can reproduce.
Read guide