Sampler AR Video

This is the SAMPLER you reach for when your checkpoint is an autoregressive video model - the Causal Forcing / Self Forcing line of checkpoints that made "real-time, streaming video generation on a single 4090" a headline in 2026. Where every video model before it denoised a whole clip at once, these generate the clip block by block, each block conditioned on the frames that came before. That's what lets them run long and run fast, and this node is the piece that drives the loop.

SamplerARVideo landed in ComfyUI core in May 2026 (the "auto-regressive video generation" commit) alongside its two partners: EmptyARVideoLatent, which makes the 5D [B, C, T, H, W] latent you'll denoise, and ARVideoI2V, which encodes a start image and seeds the KV cache so you get image-to-video without a separate I2V architecture. If you see a workflow for these models, it's these three nodes plus the usual sampler scaffolding.

How it works

Under the hood it registers the sampler ar_video with one piece of configuration: how many frames each autoregressive block handles. The node's own tooltip says it plainly - 1 = framewise, 3 = chunkwise, and it must match the checkpoint's training mode. "Must" isn't editorializing; the model was trained to predict either one frame at a time or a chunk at a time, and feeding it the wrong block size is like speaking to it in a dialect it never learned. Results come out blurry, drifting, or just wrong.

The inputs and outputs that matter

This is a deliberately minimal node - one input, one output:

• num_frame_per_block (INT, 1–64, default 1). Framewise is the safe default for most AR checkpoints; go to 3 only if the model card says chunkwise. 64 is technically allowed but nothing trains that way.
• SAMPLER output - wire it into the sampler port of SamplerCustomAdvanced.

Everything else (noise, guider, sigmas, latent) still comes in from the surrounding graph. The AR model is typically run through a CFG-guider at low strength and a short sigma schedule - these are distilled for speed, so they want few steps, not twenty.

Getting the pieces

The node itself ships with ComfyUI - nothing to install, no Manager needed. The part you do need to fetch is the checkpoint, and those are plain diffusion checkpoints you drop into your models directory. The Causal-Forcing releases are on Hugging Face (look for the zhuhz22 / Causal-Forcing repos); the same folder layout as any video model applies.

Where people get burned

• Wrong block size. If output quality collapses or motion looks broken, check num_frame_per_block against the checkpoint's stated training mode before you blame the prompt.
• Forgetting it's distilled. Treating an AR video model like a 20-step diffusion model wastes the point - these want few, aggressive steps, and piling on steps tends to degrade rather than refine.
• VRAM. Real-time video is the reward, but the KV cache seeding and block loop still want a serious card; people routinely offload or drop resolution when the first run OOMs.

It's new, it's niche, and there's still a visible quality drop versus the big offline models in side-by-side tests. But if "generate thousands of frames without the clip falling apart" is your goal, this is the sampler that gets you there.