Stale VACE conditioning frames persist across chunks in LongLive/MemFlow/RewardForcing

[ryanontheinside]

Description

When vace_input_frames / vace_input_masks are not provided in a chunk's kwargs, LongLive, MemFlow, and RewardForcing pipelines fail to clear them from PipelineState. The previous chunk's conditioning frames silently persist and get reused by VaceEncodingBlock, which sees non-None values and re-enters the conditioning path.

This is inconsistent with how other transient inputs are handled — first_frame_image, last_frame_image, vace_ref_images, etc. are all explicitly cleared when absent from kwargs.

KreaRealtimeVideoPipeline is the only pipeline that currently handles this correctly.

Root cause

Each pipeline's _generate() bulk-writes kwargs to state, then manually clears a hardcoded list of transient keys. vace_input_frames and vace_input_masks were added to Krea but not to the other three pipelines.

More broadly, this pattern — duplicating the same transient-clearing logic across every pipeline subclass — violates the Open-Closed Principle. Adding a new transient parameter requires modifying every pipeline, which is how this bug happened in the first place.

Affected pipelines

• longlive/pipeline.py
• memflow/pipeline.py
• reward_forcing/pipeline.py

Potential fix

Centralize the bulk-write + transient-clearing into a shared utility (e.g. in defaults.py):

TRANSIENT_STATE_KEYS = [
    "transition", "video", "vace_ref_images",
    "first_frame_image", "last_frame_image",
    "vace_input_frames", "vace_input_masks",
]

def apply_kwargs_to_state(state, kwargs):
    for k, v in kwargs.items():
        state.set(k, v)
    for key in TRANSIENT_STATE_KEYS:
        if key not in kwargs:
            state.set(key, None)

Each pipeline's _generate() replaces the manual loop + clearing block with a single apply_kwargs_to_state(self.state, kwargs) call. One place to declare transient keys, no per-pipeline duplication.

Note

In practice this may be partially masked during normal operation because PipelineProcessor.process_chunk() continuously feeds vace_input_frames every chunk when vace_enabled and vace_use_input_video are both true. The bug surfaces when the video input stops or those flags change mid-stream.

[yondonfu]

Related to #134