Skip to content

Difforum 1.0: camera direction layer with timeline Director, video-model bridges and AE/Blender export - #1

Merged
chillithebillis merged 4 commits into
mainfrom
v1
Sep 27, 2026
Merged

chillithebillis merged 4 commits into
mainfrom
v1

Conversation

@chillithebillis

@chillithebillis chillithebillis commented Sep 27, 2026 •

Copy link
Copy Markdown
Owner

Difforum 1.0: a camera direction layer

Difforum becomes a direction layer. You direct a shot on a visual timeline and render it with any renderer.

What changes

  • Director (timeline): a multi-track editor with Scenes (prompt + mood), Camera (a visual picker with 25 moves, speed, amount, lens, easing and audio reaction) and Energy (a drawable denoise curve). The animated preview is computed by the same Python engine that renders. A single direction wire feeds the renderers.
  • Setup with target grids: LTX 8k+1, MiniMax H3 17k+5 @ 24 fps, Wan 4k+1.
  • Bridges: Guide Frames, Keyframes, Camera → Prompt, LTX Guides (core LTXVAddGuide) and H3 Shot (first/last frame, segments longer than 20 s).
  • Camera Export / Import for After Effects (.jsx), Blender (.py) and JSON, plus exporters in tools/.
  • One feedback engine behind the Feedback Sampler, Live Sampler and Storyboard. Depth follows the image in 3D, and pseudo-3D replaces the old silent freeze when no depth map is connected. Cadence crossfades between keys, revealed areas are repainted, colour anchors per scene, and the z-buffer is deterministic on CUDA/MPS/CPU.
  • Fewer nodes: 40 → 25, plus a Render Options node so the sampler stays small. The 0.x nodes stay registered, hidden and deprecated, under Difforum/legacy, so old workflows still open. See docs/MIGRATION.md.
  • Fixes: the top-level core import that collided with other packs; three templates with camera values in the wrong fields; and templates that rendered a static camera.
  • Project: 9 templates in ComfyUI's template browser, generated from the node definitions. docs/NODES.md is generated too. Also CI, a registry publish workflow, and about 40 MB of media replaced with small WebP files.

Tests

  • pytest: 51 tests, including the 0.x engine checks, the new engine/direction/nodes/export suites, and round trips through JSON, Blender and AE.
  • Loaded in a current ComfyUI with the current frontend: all 25 nodes register, the Director widget renders and plays, all 9 templates load with valid widget values, and 01_storyboard_no_model runs end to end.
  • LTX Guides was run against the real core LTXVAddGuide. The Feedback Sampler ran through ComfyUI's real common_ksampler / VAE on a tiny randomly initialised SD1.5 model (the plumbing was verified; image quality was not tested).

Not tested yet

Renders with real checkpoints (SDXL/Flux), LTX-2.5 and H3 on real models, and the AE/Blender scripts inside those apps. Rotation order in AE should be checked on a real comp.


Generated by Claude Code

@chillithebillis
chillithebillis merged commit d997003 into main Sep 27, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant