[fal.ai] longlive/VACE: vace_input_masks must have 1 channel, got 3 — depth map fed as RGB instead of grayscale

[livepeer-tessa]

Summary

The longlive pipeline crashes with a ValueError in VaceEncodingBlock._encode_with_conditioning when vace_input_masks has 3 channels instead of the expected 1. This happens when the video-depth-anything pipeline output is used as vace_input_frames in a graph — the depth map appears to arrive as a 3-channel (RGB) tensor rather than the 1-channel (grayscale) format that VACE requires.

The error fires continuously (~160+ times in a single session), causing the entire longlive pipeline to fail on every chunk for the duration of the session.

---

Error Logs (Grafana/Loki — 2026-04-10 16:36–16:38 UTC)

fal_app: github_f1lhgmk5v76a0ev1w0u378by-scope-app--prod
fal_job_id: 42337cb5-17d4-440c-82d5-1e2aa30a485d
session: 3c6a7a72

Error in block: (vace_encoding, VaceEncodingBlock)
Error details: VaceEncodingBlock._encode_with_conditioning: vace_input_masks must have 1 channel, got 3
  File "/app/src/scope/core/pipelines/wan2_1/vace/blocks/vace_encoding.py", line 207, in __call__
  File "/app/src/scope/core/pipelines/wan2_1/vace/blocks/vace_encoding.py", line 767, in _encode_with_conditioning
ValueError: VaceEncodingBlock._encode_with_conditioning: vace_input_masks must have 1 channel, got 3

Repeated every ~400ms for the full session (160+ occurrences).

---

Triggering Graph Config

The user had configured the following pipeline graph:

input → video-depth-anything → longlive (vace_input_frames) → rife → output

With params:

• vace_enabled: True
• vace_use_input_video: True
• vace_context_scale: 0.6

The video-depth-anything pipeline output is fed directly to longlive on the vace_input_frames port. The depth pipeline likely produces a 3-channel output (depth map replicated across RGB channels) but VaceEncodingBlock expects a single-channel grayscale mask.

---

Affected File

src/scope/core/pipelines/wan2_1/vace/blocks/vace_encoding.py

• Line 207: __call__
• Line 767: _encode_with_conditioning

---

Note: Distinct from Related Issues

• VACE crash #721: vace_input_masks shape mismatch — temporal frames count (B,1,13 vs B,1,12). Different dimension.
• krea-realtime-video / WAN2.1 VAE: Conv2d crashes with (2 × 513) input < (3 × 3) spatial kernel in VaceEncodingBlock #713: Conv2d spatial kernel underflow — different error entirely.
• This issue: Channel count mismatch (C=3 instead of C=1).

---

Suggested Fix

In _encode_with_conditioning (line ~767), add a channel check and auto-convert before the validation raises:

if vace_input_masks.shape[1] != 1:
    # Convert RGB depth map to single-channel by averaging
    vace_input_masks = vace_input_masks.mean(dim=1, keepdim=True)

Or alternatively, validate and raise a more helpful error directing the user to ensure their mask source produces single-channel output.

The input should ideally be converted/normalized at the graph edge level (when the depth map is wired to the vace_input_frames port) rather than inside the block itself.

---

Impact

• Severity: High — the pipeline silently keeps running but produces no useful output; every chunk fails
• Frequency: ~160+ errors in a single ~2-hour session (2026-04-10)
• Affects: Any user wiring video-depth-anything → longlive with VACE enabled