nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[feat] v2: Wan2.2-I2V-A14B (MoE i2v) — combine boundary-routed experts + i2v conditioning

Browse Source 
Reuses everything: 2 WanTransformer3DModel experts + BoundaryTimestepRouting (from the A14B MoE pattern),
the CLIP image encoder + first-frame [mask|cond] conditioning + the i2v program (from the Fun-InP i2v
port), and the shared WanDenoiseLoop (i2v hooks + the boundary expert). No new adapter. CPU-verified: the
toy MoE i2v runs end-to-end (2 experts + boundary + conditioning -> finite video); resolves with i2v caps.
Structural (GPU-pending: 2x14B, like the A14B T2V). Wan family now largely covered (T2V 1.3B/14B/TI2V-5B/
A14B, causal SF, i2v 1.3B/14B/A14B). CPU mini 211/2.
This commit is contained in:
Will Lin
2026-07-04 17:15:34 +00:00
committed by Peiyuan Zhang
parent 7464ef8308
commit fc3332550c
3 changed files with 63 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/inference/basic/V2_PORTING_STATUS.md
+2 -1
View File
@@ -55,7 +55,8 @@ Wan2.1 model — covered by wan21; those runtime modes are out of scope.
| `basic_ltx2_3_distilled_i2v*.py` | LTX-2.3-Distilled I2V | Image conditioning on the ltx2 single-stage program (encode the input image + condition the denoise). |
## Backlog — goal: port ALL fastvideo models (63 total: 8 ported · 21 reuse-arch · 34 new-arch)
Ported (11): the 7 above + **Wan2.1-T2V-14B** + the **Wan2.1 i2v cluster** — Wan2.1-Fun-1.3B-InP
Ported (12): the 7 above + **Wan2.1-T2V-14B** + **Wan2.2-I2V-A14B** (MoE i2v, structural) + the
**Wan2.1 i2v cluster** — Wan2.1-Fun-1.3B-InP
(GPU-verified: CLIP encoder + first-frame [mask|cond] -> 36ch DiT) and Wan2.1-I2V-14B-480P/720P (reuse the
same i2v card; 14B weights GPU-pending). i2v required a real sub-system: the image-encoder adapter, the
loop's i2v conditioning hooks, and the Wan adapter's 36ch concat (all landed + T2V-unregressed).