ComfyUI Video Generation Fails with CUDA Out of Memory

This problem looks like a single error, but several independent layers can produce the same visible result. In ComfyUI, the specific symptom is this: A video workflow fails during sampling even though a single image generation fits. The practical objective is to restore the affected feature, determine the actual failing layer, and confirm the result without sacrificing data or account protection.

Symptoms and first checks

A video workflow fails during sampling even though a single image generation fits. The key distinction is whether the failure happens before the service, game, application or device becomes available, or only when the user performs the specific action. Duplicate the workflow and halve frame count or resolution without replacing checkpoints.

• Exact observation: A video workflow fails during sampling even though a single image generation fits.
• First useful comparison: Duplicate the workflow and halve frame count or resolution without replacing checkpoints.
• Expected healthy state: The reduced workflow completes, then settings can be increased one parameter at a time while VRAM stays below the device limit.

Keep a short timeline: when did the behavior start, what software or network condition changed immediately beforehand, and is the failure consistent on every attempt? A successful unrelated feature should not be treated as proof that the affected layer is healthy.

Why the failure happens

Video frame count, temporal attention, latent size and model residency can multiply peak VRAM demand. This is a hypothesis that should be confirmed through the comparison above. A model that downloads successfully is not necessarily compatible with a tokenizer, runtime, GPU or API mode. Keep data and model artifacts separate so experiments do not overwrite production indexes or checkpoints.

Small reproducible inputs are valuable: fix the random seed where supported, preserve the exact model hash and run a single request before enabling batching. For retrieval failures, separate the search layer from the final generated answer If the test result contradicts the suspected cause, do not push forward with a broad reset. Return to the recorded baseline and investigate the next boundary instead.

Before you change anything

Preserve the known working model, environment lockfile, workflow and any custom prompt templates. Capture VRAM capacity, model architecture, device selection and the exact failing phase before upgrading dependencies.

1. Duplicate the workflow and halve frame count or resolution without replacing checkpoints.
2. Watch nvidia-smi memory over nodes and identify whether the allocation spike happens during VAE decoding or temporal sampling.
3. Note the original settings and capture any relevant error identifier without copying passwords or session tokens.

Do not begin with formatting a disk, deleting a player profile, disabling secure boot, bypassing TLS checks or running registry cleaners. The controlled tests below are designed to find the narrow failure before considering any operation that could affect unrelated data.

Interpreting the diagnostic results

Use the following evidence map after completing the Beginner checks. Each row provides an interpretation, not an instruction to apply every possible fix.

Watch nvidia-smi memory over nodes and identify whether the allocation spike happens during VAE decoding or temporal sampling. If the evidence points away from this layer, return to the previous row. The process is intentionally sequential: a configuration change only has diagnostic value when the prior state is recorded.

Optional read-only diagnostic — Shell (read-only): Run this only on a system you administer. Replace example values with your own legitimate paths as needed.

nvidia-smi

Mistakes that make this harder

• Avoid changing unrelated layers. Test a minimal prompt or tiny input first. Memory pressure, tokenizer faults, and API serialization issues usually become easier to distinguish when you remove extra models, agents and extensions
• Do not confuse a temporary workaround with a fix. Video frame count, temporal attention, latent size and model residency can multiply peak VRAM demand.
• Do not erase the evidence. Keep the original settings and note whether a video workflow fails during sampling even though a single image generation fits.

Decision rule: If the proposed change does not produce the expected result in a repeatable test, restore the previous configuration. A problem that remains identical needs better evidence, not a larger reset.

Confirm the repair

The reduced workflow completes, then settings can be increased one parameter at a time while VRAM stays below the device limit. Use the same device, title, workload or service path on which the failure was observed. Compare the outcome directly with the original screenshot or error log, and repeat the test twice if the failure was intermittent.

For data-processing systems, preserve a set of representative inputs and expected outputs. Confirm both functional correctness and resource use before restoring a production workload If one test passes but the underlying account, certificate, performance metric or data synchronization is still wrong, mark the result as partial and continue with the Advanced checks rather than claiming a complete repair.

If it still fails

Watch nvidia-smi memory over nodes and identify whether the allocation spike happens during VAE decoding or temporal sampling. If ComfyUI still fails the original operation, revert temporary test settings that did not help and capture the precise evidence for the vendor or administrator. Include what worked, what failed and whether the comparison was conducted on the same build and connection.

Do not keep repeating account resets, firmware flashes or driver removals when there is no measurable improvement. For unsupported hardware, managed work devices or possible physical failure, ask the relevant vendor or qualified technician to confirm the supported route forward.

Avoiding a recurrence

Use smaller approved settings, enable the workflow’s supported offload options and unload unnecessary models before generating. Keep the exact working driver, firmware version, app build or policy setting documented. Future changes should be tested against the same baseline, especially after operating-system, network or application updates.