Joint text-to-video post-training stops sharing one reward: prompt and generator get separate credit, lifting a matched LTX-2.5 control from 61.47 to 64.46

Text-to-video systems now expose two trainable adaptation channels: a language-side prompt policy and the video generator itself. Prompt-side methods previously improved conditioning with a fixed generator, generator post-training changed the diffusion or flow model alone, and recent systems optimize both under one terminal reward. Once both modules are trainable, the same scalar reward must serve both attribution and coordination, whereas with a single updated module it served as supervision for that module alone. Under one LTX-2.5, Qwen3.5-9B prompt-optimizer and Flow-GRPO harness, the strictly matched joint-training control Matched Global-LOO reaches 61.47 on T2V-CompBench and 55.58 on StoryEval, while the full FlowCred reaches 64.46 and 58.08.

Hierarchical credit assignment uses matched prompt-seed rollout grids to separate reward variation caused by prompt choice from the residual caused by generator randomness once the prompt is fixed. A global leave-one-out baseline removes only the current trajectory and still mixes prompt-to-prompt variation into the generator advantage; hierarchical credit makes the generator advantage a prompt-conditioned residual instead. At a fixed decoded-video budget, generator-advantage variance drops by about 31% and matched seeds reduce prompt-credit variance by about 27%, with the matched-seed ablation reaching 62.63 on T2V-CompBench versus 62.21 for independently sampled seeds.

Functional coordination probes alternative current-policy prompts at states actually encountered by the generator, and reuses the same-state response geometry for generator-aware prompt shaping and reliability-gated generator routing. Even after statistical credit is separated, a prompt change can still duplicate a local improvement that the generator reward step would make in the same direction; coordination targets that overlap. Across three independent training runs, prompt shaping alone adds +0.29/+0.65/+0.77 on VBench/StoryEval/T2V-CompBench, generator routing alone adds +0.38/+0.47/+1.04, and the combination adds +0.65/+1.37/+1.83.

The demonstrated scope is online joint reward post-training on LTX-2.5 with a Qwen3.5-9B prompt optimizer and Flow-GRPO, an 8,000-request training pool, and evaluation on VBench, VBench-2.0, T2V-CompBench and StoryEval at unchanged decoded-video and reward-model-call budgets. Functional coordination requires two backend-local objects: a response vector from changing text conditioning at a fixed state, and a local improvement covector supplied by the generator updater; Flow-GRPO, RVM and differentiable reward flow provide both directly. Preference objectives with only a trajectory-level update and no local improvement covector, and closed API-only generators, are outside the current formulation.

Functional coordination is local and first-order; its prompt term is auxiliary shaping and its generator rule is a minimum-overlap regularizer rather than a conserved transfer of gradient mass. The realized-update audit is described as a descriptive consistency check of trained behavior, not an exact transfer of causal responsibility. Intervals from three outer replicates are used descriptively, without stronger significance claims. The increment is positive in every tested backbone and updater setting, though its magnitude varies.