fix(vace): guard against sub-minimum spatial dims in WAN VAE encoder

[livepeer-tessa]

Summary

Fixes #713 — krea-realtime-video / WAN2.1 VAE: Conv2d crashes with (2 × 513) input < (3 × 3) spatial kernel in VaceEncodingBlock.

---

Root Cause

The WAN VAE encoder has 3 spatial downsample stages, each implemented as:

nn.Sequential(nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2)))

Tracing the minimum viable height through all three stages:

Stage	Input H	After ZeroPad	After Conv2d(3, stride=2)
DS1	2	3	1
DS2	1	2	CRASH (2 < 3)

An input with H=2 survives DS1 but crashes at DS2 with PyTorch's opaque:

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

The minimum safe spatial input is 8px in both dimensions (derived by working backwards through all 3 stages to ensure each padded intermediate is ≥ 3).

The crash was triggered at line 123 of vace/utils/encoding.py when encoding reference images with vae.encode_to_latent(refs_stacked, use_cache=False).

---

Changes

src/scope/core/pipelines/wan2_1/vace/utils/encoding.py

• Added MIN_VAE_SPATIAL_SIZE = 8 constant with derivation comment
• Added _check_spatial_size() helper that raises a descriptive ValueError early (before the cryptic PyTorch kernel-size error) for main frame encoding paths
• Added a warning + skip guard for reference image encoding: since refs are supplementary conditioning, we log a warning and continue without ref conditioning for that chunk — degrading gracefully rather than crashing the session 500+ times
• Moved the mask-channel padding and latent concatenation into the else branch so ref_latent_batch is never accessed on the skip path (was previously a latent NameError risk)
• Added import logging and module-level logger

---

Behaviour After Fix

• Ref images too small (< 8px in H or W): Warning logged, chunk continues without reference conditioning. Session does not crash.
• Main frames too small: Descriptive ValueError surfaced immediately with resolution guidance (vs. cryptic PyTorch kernel message after several encoder stages).

---

Testing

This fix is defensive (no new code paths in the happy path). Verified with:

• python3 -c "import ast; ast.parse(open('src/scope/core/pipelines/wan2_1/vace/utils/encoding.py').read()); print('ok')" ✅

Reviewers: @mjh1 @emranemran

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: f2eb7808-7580-4d60-be2b-a8debd7e4c29

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/713-vace-encoding-min-spatial-size

📝 Coding Plan

• Generate coding plan for human review comments

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

Tip

You can enable review details to help with troubleshooting, context usage and more.

Enable the reviews.review_details setting to include review details such as the model used, the time taken for each step and more in the review comments.

[github-actions]

🚀 fal.ai Preview Deployment

App ID	daydream/scope-pr-714--preview
WebSocket	wss://fal.run/daydream/scope-pr-714--preview/ws
Commit	6cc5d17

Testing

Connect to this preview deployment by running this on your branch:

uv run build && SCOPE_CLOUD_APP_ID="daydream/scope-pr-714--preview/ws" uv run daydream-scope

🧪 E2E tests will run automatically against this deployment.

[github-actions]

✅ E2E Tests passed

Status	passed
fal App	daydream/scope-pr-714--preview
Run	View logs

Test Artifacts

Check the workflow run for screenshots.