VAE and latent channel mismatch: verify the exact variant, not just its family name
An unexpected channel count or latent dimension calls for checking the model, VAE and latent-construction nodes together. Image and video latents do not all share the same structure.
Symptoms and scope
An unexpected channel count or latent dimension calls for checking the model, VAE and latent-construction nodes together. Image and video latents do not all share the same structure.
Error fragments for search; IDs, values and filenames may vary:
expected input
channels
size of tensor
Source-supported context
Official model troubleshooting includes channel-count conflicts. Wan 2.2 documentation uses different VAE configurations for the 5B TI2V and 14B routes, so the family label alone does not justify substitution. Source 1 Source 2
Cases to distinguish
1. The VAE belongs to another model or variant.
2. An image, video or cross-family encoded latent is connected to an incompatible sampler/model.
3. An older template's dimensions, frames or specialized latent node has the same interface type but a different internal layout.
Diagnostic sequence
The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.
Step 1. Record expected and actual dimensions and identify whether encoding, sampling or decoding failed.
Step 2. Check the exact variant's official template for diffusion, VAE and latent-node choices.
Step 3. Restore its original encode–sample–decode chain, removing intermediate latents copied from other workflows. If the template and model permit a smaller image or shorter clip, test one supported reduction at a time.
Step 4. If template file instructions disagree, retain both URLs and versions and mark the relationship as unresolved instead of choosing the newest-looking filename.
Completion check
The matched chain passes sampling and decoding; add back only supported dimensions and optional controls after successful saving.
Limitations and cautions
Zero-padding, cropping channels or adjusting width/height does not generally fix architecture differences. No Wan variant or GPU capacity has been runtime-tested here.
Original sources
- ComfyUI model troubleshooting — original record checked 2026-09-21.
- Wan2.2 native workflows — original record checked 2026-09-21.
Localization prepared: 2026-09-22. The source-check dates above were inherited from the existing article; this translation does not claim they were all rechecked today. No GPU run, minimum-memory measurement or runtime guarantee is supplied.
Related troubleshooting and guides
Check the next possible cause
The same symptom can come from a different cause. Work through these related entries in order.
- VAE out of memory after sampling finishes: diagnose decoding separately
VAEDecodeCompleted sampling still leaves latent data that must be decoded into pixels. A VAE failure after sampling therefore differs from memory exhaustion at the start of sampling. - mat1 and mat2 shapes cannot be multiplied: inspect the model component chain
mat1 and mat2 shapes cannot be multipliedMatrix dimensions help locate a failure during sampling, encoding or adapter use, but one numeric pair does not identify a unique missing model.
Was this page helpful?
Anonymous. Only a yes/no counter is stored; no account, IP address or device details.
Sources & references
Source-reviewed against current official ComfyUI model troubleshooting and Wan2.2 workflow pages on 2026-09-25. The diagnostic sequence is editorial guidance, not a site-run model/GPU test or repair guarantee; older source-check dates in the body remain historical.
01ComfyUI model troubleshootingSource checked: 2026-09-2102Wan2.2 native workflowsSource checked: 2026-09-21Report an issue · 72980a7e-ba30-5c96-9fb9-e93a3f57cf5f
Related reading
Pages linked to this one by the editors.
Distinguish 5B TI2V from the 14B text/image-to-video routes. A shared Wan 2.2 label does not make weights, VAE, encoder and latents freely interchangeable.
- Wan 2.2