VaceEncodingBlock fails on extremely narrow input images (kernel size > input size)

[livepeer-tessa]

Summary

The VaceEncodingBlock encounters a convolution error when processing input images that have been resized to extremely narrow dimensions (width as small as 2 pixels), which is smaller than the 3x3 convolution kernel.

Error Details

Multiple ERROR-level logs from fal.ai backend (28 Feb 2026):

Pipelines affected:

• krea-realtime-video (majority of occurrences)
• streamdiffusionv2

Error pattern:

Error in block: (vace_encoding, VaceEncodingBlock)
Error details: Calculated padded input size per channel: (257 x 2). Kernel size: (3 x 3). Kernel size can't be greater than actual input size

And for streamdiffusionv2:

Calculated padded input size per channel: (513 x 2). Kernel size: (3 x 3). Kernel size can't be greater than actual input size

Analysis

• The padded input has dimensions like (257 x 2) or (513 x 2) — the width dimension is only 2 pixels
• The 3x3 convolution kernel cannot process inputs smaller than 3 pixels in any dimension
• The 257/513 heights suggest standard resolutions (256+1 for padding or 512+1)
• The 2-pixel width is abnormally narrow, likely caused by:
  1. Extreme aspect ratio input images/videos
  2. Incorrect resize/crop operations upstream
  3. A bug in resolution calculation when switching aspect ratios

Suggested Fix

1. Add input validation in VaceEncodingBlock to reject or pad inputs below minimum dimensions
2. Add a minimum dimension check earlier in the pipeline (e.g., in input preprocessing)
3. Consider logging a more descriptive warning when inputs are near the minimum size

Sample Timestamps (28 Feb 2026 UTC)

• 09:14:34 - 09:14:39 (krea-realtime-video)
• 09:17:14 (streamdiffusionv2)
• 09:19:43 - 09:22:48 (krea-realtime-video)

Environment

• Backend: fal.ai
• App ID: 3564d791-2447-4a41-9ab8-53909db8c657
• Blocks affected: VaceEncodingBlock

[livepeer-tessa]

New Occurrence — 2026-03-15 (transposed dimensions)

Fresh prod hit today with the height dimension this time (previously width=2, now height=2):

Error in block: (vace_encoding, VaceEncodingBlock)
Error details: Calculated padded input size per channel: (2 x 513). Kernel size: (3 x 3).
               Kernel size can't be greater than actual input size

• Pipeline: krea-realtime-video
• App: github_f1lhgmk5v76a0ev1w0u378by-scope-app--prod
• Job: 5193400c-da0f-4eef-8bdd-dd0fdd26c1db
• Window: 10:48:28 – 10:59:33 UTC (11 minutes, 665 seconds)
• Error count: ~2,372 errors at ~4/second (saturating the Loki 500-entry fetch limit repeatedly)

Original issue had (513 x 2) — width=2. Today's is (2 x 513) — height=2. Both axes can underflow, same root cause.

PR #696 adds a spatial padding guard (the spatial mirror of the temporal guard in PR #674). Fix pings @mjh1 @emranemran for review.

[livepeer-tessa]

New Occurrence: 2026-03-14 & 2026-03-15 (Grafana/Loki)

Auto-detected via Grafana error monitoring.

New variant: (2 x 513) — extremely narrow height (previously seen as narrow width)

Prior occurrences showed (257 x 2) and (513 x 2) — width of 2 pixels.

New observations show the transposed case: height of 2 pixels, width of 513:

Error in block: (vace_encoding, VaceEncodingBlock)
Error details: Calculated padded input size per channel: (2 x 513). Kernel size: (3 x 3).
               Kernel size can't be greater than actual input size

Stack trace (2026-03-15 10:58:30 UTC):

File "/app/src/scope/core/pipelines/wan2_1/vace/blocks/vace_encoding.py", line 199, in __call__
    self._encode_extension_mode(
File "/app/src/scope/core/pipelines/wan2_1/vace/blocks/vace_encoding.py", line 402, in _encode_extension_mode
    z0 = vace_encode_frames(
File "/app/src/scope/core/pipelines/wan2_1/vace/utils/encoding.py", line 123, in vace_encode_frames
    ref_latent_out = vae.encode_to_latent(refs_stacked, use_cache=False)
File "/app/src/scope/core/pipelines/wan2_1/vae/wan.py", line 139, in encode_to_latent
    latent = self._encode_with_cache(pixel, scale, self._create_encoder_cache())
File "/app/src/scope/core/pipelines/wan2_1/vae/wan.py", line 159, in _encode_with_cache
    out = self.model.encoder(
File "/app/src/scope/core/pipelines/wan2_1/vae/modules/vae.py", line 380, in forward
    x = layer(x, feat_cache, feat_idx)
File "/app/src/scope/core/pipelines/wan2_1/vae/modules/vae.py", line 163, in forward
    x = self.resample(x)  # <-- Conv2D: 3x3 kernel on 2xN spatial input
RuntimeError: Calculated padded input size per channel: (2 x 513). Kernel size: (3 x 3).
             Kernel size can't be greater than actual input size

Details

• Pipeline: krea-realtime-video (extension mode, not reference-only mode)
• Session: 2c7ce7e9 (fal.ai)
• Occurrences: ~250+ in a ~1.5 minute window (same session, every chunk — log storm)
• Prior occurrence: Also seen 2026-03-14 14:05 UTC (same pattern)
• Code path: _encode_extension_mode → vace_encode_frames → vae.encode_to_latent → Conv2D resample layer

The fix suggested in the issue (minimum dimension validation) would cover both orientations.

[livepeer-tessa]

Update: PR #696 covers the new transposed variant

The new occurrences from 2026-03-14/15 (height=2, width=513) confirm the error affects both orientations:

Date	Shape	Axis
2026-02-28	(257×2) / (513×2)	width = 2
2026-03-14/15	(2×513)	height = 2

PR #696 already guards against both — it pads whichever of height or width is < 3. The fix is orientation-agnostic.

Tagging reviewers for awareness: @emranemran @mjh1 — the volume on 2026-03-15 was ~250+ errors in 1.5 minutes from a single session, so merging sooner rather than later would help.

[livepeer-tessa]

New occurrence — 2026-03-15

Same error pattern observed again, this time with transposed dimensions (height=2 instead of width=2):

Error in block: (vace_encoding, VaceEncodingBlock)
Error details: Calculated padded input size per channel: (2 x 513). Kernel size: (3 x 3). Kernel size can't be greater than actual input size

Pipeline: krea-realtime-video
Session: [2c7ce7e9]
Time: 2026-03-15 10:55:00 – 10:59:33 UTC
Occurrences: 500+ in ~5 minutes (every ~234ms), same flood pattern as before

fal.ai job details:

• job_id: 5193400c-da0f-4eef-8bdd-dd0fdd26c1db
• node_id: ce7099c4-0da2-e85f-2b20-51355980f222
• worker: fal-jobs/GPU-H100/5193400c-da0f-4eef-8bdd-dd0fdd26c1db
• app_id: b34bc0cd-90e4-4898-8142-3c905ca0f13c (prod)

Stack trace (same call path as before):

scope/core/pipelines/krea_realtime_video/pipeline.py:271 _generate
  → diffusers modular_pipeline.py:932 __call__
  → vace_encoding.py:199 __call__
  → vace_encoding.py:402 _encode_extension_mode
  → vace/utils/encoding.py:123 vace_encode_frames
  → wan2_1/vae/wan.py:139 encode_to_latent
  → wan2_1/vae/wan.py:159 _encode_with_cache
  → wan2_1/vae/modules/vae.py:380 forward
  → vae.py:163 self.resample(x)   ← fails here
  → torch.nn.modules.conv.Conv2d (3x3 kernel, height=2 input)

Note this occurrence has dimensions (2 x 513) — the height dimension is 2 pixels (vs previous report's (257/513 x 2) where width was 2). Same root cause, different aspect ratio: extremely tall/narrow input after VAE downsampling.

These 500+ floods occurred just before logging from this worker stopped entirely at 10:59:33 UTC. No fal.ai logs have been received from prod since that time (~3 days ago as of this report).