[Kandinsky 5] Fix I2V conditioning: don't inject the image latent into visual_cond channels (#14282)

Fix Kandinsky 5 I2V: don't duplicate image latent into visual_cond channels

Kandinsky5I2VPipeline.prepare_latents injected the conditioning image both as the
first latent frame and into the visual_cond channels; the reference implementation
(kandinskylab/kandinsky-5) only does the former. The duplicate over-conditions the
first frame and produces mesh/visual artifacts. Remove the redundant injection to
match the reference.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Author: 3 people

src/diffusers/pipelines/kandinsky5/pipeline_kandinsky_i2v.py

@@ -725,7 +725,6 @@ def prepare_latents(
             )
 
             visual_cond_mask[:, 0:1] = 1
-            visual_cond[:, 0:1] = image_latents
 
             latents = torch.cat([latents, visual_cond, visual_cond_mask], dim=-1)