Training a character LoRA that actually holds

Identity fails in the dataset, not in the sampler. Nodezy fixes the set before it trains

Identity fails in the dataset, not in the sampler. Small frames, odd ratios, baked-in text wreck training — so the popup flags them before you spend.

Fix the set

Too small: outpaint or upscale. Odd ratio: outpaint or crop. Text in the frame: keep or inpaint. Apply all fixes. Then generate.

Name, type (Style / Object / Character), drop up to 5 refs, generate the model. Character keeps a face and a person consistent across scenes.

Pick the model on the Image or Video node. You stop prompting the face and start directing the scene. Same face, every frame.

> Community <

Sep 2, 20264 mins read

Badges, levels, payouts sit in settings. The studio still earns when someone else runs your node.

> Canvas <

Aug 3, 20265 mins read

Workflows, models, artist LoRAs — every file you already paid to learn, on the same canvas.

> Canvas <

Jul 15, 20266 mins read

Frames and groups are secondary. The feature is mutating the still in place: larger, smiling, recropped.