[Lingbot World] Bound image condition storage and extend temporal RoPE (#6838)

Signed-off-by: wtz2333 <2955110911@qq.com>
This commit is contained in:
wtz2333 authored and GitHub committed 2026-09-15 18:49:55 +08:00
1 parent 4a8ac29775
commit a3dee6fdf5
12 files changed
+592 -195

No files matched your search

+5 -3
View File
@@ -48,9 +48,11 @@ python examples/offline_inference/diffusion/lingbot_world_v2.py \
```
The current causal-fast checkpoint uses four DMD sampling steps. Raw frame
counts must be `9 + 12k`, up to 117 frames. Camera files may contain a
longer bounded source trajectory (the official example contains 269 frames);
the runtime consumes only the prefix needed by the request.
counts must be `9 + 12k` for a non-negative integer `k`; for example, 129 frames
produce eleven AR blocks. Both camera files must contain the same number of
source frames, at least as many as requested and at most 4096. The runtime
consumes only the prefix needed by the request. Longer rollouts also require
sufficient device memory.
## Realtime in-process generation (deprecated)