NVIDIA CMD Long Video

Most local video models give you six seconds and call it a day. NVIDIACMDLongVideo is the node that stretches NVIDIA CMD's causal rollout to about 501 pixel frames - roughly 31 seconds at 16 fps - without overflowing the card. It's the pack's answer to the single biggest reason CMD long videos don't run locally: the official implementation caches every latent frame of the rollout, and 126 latents of KV will eat a 32GB card alive.

This node is deliberately identical to NVIDIACMDImageToVideo in shape, with one hard requirement baked in: the cmd_model you feed it must be a long checkpoint. If you wire in a chunk1_short model, it raises a ValueError telling you to load chunk1_long or chunk4_long instead. No silent garbage output - it just refuses.

How it works

The trick is the circular KV cache. Instead of keeping all 126 latents in memory, the adapter caps the attention cache at local_attn_size (21 for chunk1_long), so the model only ever holds a rolling window of context. That's what turns a would-be overflow into a manageable peak: measured on the author's RTX 5090 32GB, the long workflow peaks around 23.8 GiB and takes about 267 seconds. Still not cheap - but it runs, where the official pipeline would die on the same card.

The generation itself mirrors the short node: Cosmos-Reason1 grounds your prompt, your image gets scaled to 832×480, the student DiT rolls out the latents, and the VAE decodes to frames. Few-step, no CFG, no negative prompt - same philosophy, just a longer horizon.

The inputs

Exactly the same surface as NVIDIACMDImageToVideo:

• cmd_model - must come from a loader set to chunk1_long (or chunk4_long). The node checks this and throws if it's wrong.
• image - the starting frame, scaled to the model's fixed 832×480.
• prompt - multiline text. Required, and worth writing like a real sentence - Reason1 is a vision-language model and reads natural language far better than a comma list.
• seed - INT, default 0. Locks the rollout; same seed + image + prompt reproduces the clip.
• camera (optional) - a CMD_CAMERA value, though the tested workflows keep long video camera-free.

Output is frames, an IMAGE batch of ~501 frames you can pipe straight into NVIDIACMDSaveVideo (which will name it with a 501f suffix) or any standard encoder.

Gotchas

• This is a 5090-class workload. The author tested on an RTX 5090 32GB. At ~23.8 GiB peak you're still above what a 16GB card can hold, so check your free VRAM before queueing.
• The node runs a memory preflight. If the estimated footprint exceeds its safety budget, it refuses to start rather than OOM mid-generate. That's a feature - it means "free up VRAM / pick a lighter preset," not "broken."
• Long + camera is not the tested combo. If you want camera movement, use the camera checkpoint with the short camera workflow instead; the README's verified paths keep them separate.
• It's still 480p and still noncommercial. Same Cosmos-Predict2.5 lineage, same NVIDIA OneWay Noncommercial license on the weights - research/education only.

One long rollout at ~4.5 minutes per clip is not iteration-friendly; treat this node as the "final take" step, and hammer out your prompt on the short node first.