← Back to the film

Creation, in context.

Visual transcript of the 96-second WeGen2 film. The soundtrack is original instrumental music, with no narration. Model-generated images are presented through editorial motion design; the animation is not a model-generated video.

  1. 0–6 seconds. The WeGen2 identity opens the film: native pixel-space multimodal creation. An idea. A world of possibilities.
  2. 6–10 seconds. Reference-conditioned editing: change the green mountain city into a winter landscape.
  3. 10–15 seconds. A handcrafted paper world is edited into an illuminated night scene.
  4. 15–19 seconds. Photographic detail: a portrait in the rain, with close-ups of lighting and materials.
  5. 19–23 seconds. An illustrated fantasy film poster with model-generated lettering.
  6. 23–29 seconds. A WeGen2 creation-studio poster combines Chinese writing with miniature landscapes, architecture, and fashion scenes. Generated from one text prompt; lettering is part of the model output.
  7. 29–38 seconds. A source person and eight garment and accessory references. The references form a contact sheet, then move beside the generated outfit.
  8. 38–49 seconds. The generated outfit appears. Highlighted regions link to the blazer, shirt, jeans, sneakers, bag, belt and necklace references.
  9. 49–54 seconds. Successive image editing: add a reply to a fictional chat, then revise the messages.
  10. 54–60 seconds. A cake-making comic: eight sequentially generated panels presented together on one canvas.
  11. 60–63 seconds. Behind every image, one connected model.
  12. 63–71 seconds. Native pixels: learned patch modules connect RGB images and the backbone. No separate pretrained vision encoder or latent autoencoder is used at inference.
  13. 71–80 seconds. A-MoT: separate understanding and generation branches communicate through joint attention. Generation has image and timestep residual pathways inside the generation branch.
  14. 80–89 seconds. A-MoE: three differently colored tokens each pass through the shared expert. The router selects different expert subsets, and colored flows trace each token through the shared and routed paths. Expert count is schematic.
  15. 89–96 seconds. WeGen2: 61B total parameters, approximately 7B activated per generation token. Explore the project page.

Outfit highlights indicate visible correspondence, not exact product replication. The eyewear shape differs and is not highlighted. Detail crops come from the displayed originals.