Follow-up to #14737: drop output_type="latent" from modular decode blocks #15071

#14737 standardizes the latents the core denoise group leaves in the state: always unpacked and still normalized, mirroring what the VAE encoder blocks produce, pinned by test_latents_output_in_canonical_form. In modular pipelines that already gives users latents for free. Skip the decode block and latents stays in the state, or call pipe(..., output="latents").

Several decode blocks still accept output_type="latent", copied from the standard pipelines. Keeping that path alongside the standard form is redundant and a source of inconsistency: it is a second, separate piece of code that decides what "latent output" means, and it does not have to agree with what the denoise group produces. Today it does not:

Same argument, different tensors, written into an output slot typed as images or videos. And every time the denoise group changes, as it does for flux in #14737, this branch has to be revisited too. That is maintenance burden and complexity for no benefit.

Proposal: remove the output_type == "latent" branches and keep output_type to pil / np / pt, as the template already documents. The latents in the state are the one latent output. The ltx2 branches can go with #14612.

Related: #14730, #14737.

This issue was written by Claude Code (Claude Fable 5.1) and posted on behalf of @yiyixuxu, who reviewed and approved the text.