Fix per-frame timestep allocation for video2world mode in CosmosPredict2

In video2world (image2world) mode, conditioning frames were receiving the
same noisy timestep as all other frames. This caused the transformer to
treat the clean conditioning frame as a fully-noised input, breaking
temporal coherence and causing severe quality degradation.

Fix: after replacing conditioning frames in model_input, expand t to
shape (B, T) and zero out timesteps for frames where condition_mask=1,
signaling to the model that those frames are already clean (t=0).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: csy2077

fastgen/networks/cosmos_predict2/network.py

@@ -1352,6 +1352,13 @@ def forward(
                 conditioning_latents_full = conditioning_latents
             # Replace conditioning frames
             model_input = conditioning_latents_full * condition_mask_C + x_t * (1 - condition_mask_C)
+            # Per-frame timesteps: assign timestep 0 to conditioning (clean) frames so the
+            # model knows they are noise-free and should not be denoised. Without this, the
+            # model treats the clean first frame as a fully-noised frame, causing incoherent
+            # video2world (image2world) generation.
+            t_expanded = t.unsqueeze(1).expand(B, T)
+            mask_B_T = condition_mask[:, 0, :, 0, 0]  # (B, T)
+            t = t_expanded * (1 - mask_B_T)
 
         model_outputs = self.transformer(
             x_B_C_T_H_W=model_input,