OmniChar

You start by building a character. In OmniChar Cloud that takes 1-3 photos and nothing to install; locally you name the character, add your reference photos and the app standardises them into a .char file. Models that read references directly, such as FLUX.2 and MiniMax H3 Reference to Video, need no training at all. For models without a reference channel, such as Krea 2, the same photos are turned into a LoRA on the same canvas, and the file carries both.

Next you generate on a node canvas. Drop a node for Z-Image Turbo, FLUX.2, Krea 2, MiniMax H3 or LTX-2.5, wire in a prompt and the character, and run it on your own GPU through Inline Core, the generation engine built into the app, with no account or external server. Every render becomes a take under a shot, so you can compare takes, promote one and roll back, and shots wire into one another to form a sequence you can read in order.

Finally you assemble and reuse. A Video Director node builds a timeline from frame outputs with layered audio and exports the combined video, and every image generated locally carries the graph that made it, so dropping it back on the canvas rebuilds the pipeline. Locally the install is one set of commands with uv and a webui script, or you can launch it on RunPod.

• •Portable .char Characters - One file holds a character's face, body, clothes and voice, built once and used with FLUX.2, MiniMax H3 and Krea 2 locally and hosted models such as Nano Banana Pro. Identity scoring shows how closely each take matches the character.
• •ComfyUI Nodes and SDK - Free ComfyUI nodes add a Load Character node to workflows you already use, characters built in ComfyUI open in OmniChar's Characters panel, and the omnichar-sdk lets developers integrate .char files. Shared workflows open on the canvas with their nodes wired and required models listed.
• •Local Image and Video Generation - Inline Core runs Z-Image Turbo, FLUX.2, Krea 2, MiniMax H3 and LTX-2.5 on NVIDIA, AMD or CPU. LTX-2.5 and MiniMax H3 produce video with a synchronised soundtrack in the same pass, with no per-second billing when you run them yourself.
• •LoRA Training on the Canvas - Wire Load Dataset into Caption into Train LoRA, watch the loss curve live, and stop and resume from checkpoints. It trains LoRAs for LTX-2.5, Krea 2, Z-Image, FLUX.2 and MiniMax H3, including Control LoRAs learned from pairs of video clips.
• •Takes, Sequences and a Timeline - Regenerating never destroys earlier attempts, shots link into a sequence, and the Video Director and Trim nodes cut the finished video with audio layers without leaving the app.
• •Cloud and API Nodes - OmniChar Cloud runs the models on hosted GPUs in the browser at up to 4K, while API Nodes bring hosted models such as GPT Image 2, Nano Banana and Seedance onto the same canvas using your own provider key, with a rough price estimate on each node.

OmniChar is built for AI filmmakers and creators making short films, ads or social series who need the same character to look the same from shot to shot, without training a new LoRA for every model they try. ComfyUI users can add .char files to the workflows they already run instead of rebuilding them, and developers can work with characters through the SDK and API.

Hobbyists with a capable GPU get a free, local setup for image and video generation and LoRA training without writing scripts; the three image models train at 512px on a 16GB card, while LTX-2.5 training wants 48GB. Creators without a GPU can start in the browser on OmniChar Cloud with free credits and pay only for what they generate. Agencies and brands building digital twins of consenting talent can look at the early-access likeness licensing, which records the owner, scope and revocation status inside the file.