DaSiWa AutoLong Start

If you've run one of DaSiWa's Wan 2.2 workflows, you know the ritual: generate a segment, save the tail frame, wire it back in as the new first frame, queue again, repeat until you have a clip, then stitch the pieces in another program. It works, and it is miserable. This node is the fix. DaSiWaAutoLongStart is the control brain of the ComfyUI-DaSiWa-AutoLong pack - you queue the workflow exactly once and it keeps re-queueing itself until every segment of your long video is written to a single file.

The context matters here. Wan 2.2 is the last open Wan, and its native ceiling is still 81 frames. SVI 2.0 Pro is what pushed past that into 20–25 second continuous takes, but it does it in segments, and someone still has to chain those segments. That someone is this pack, and this node is where the chain begins.

What it actually does

You feed it the very first frame and a segment count, and it acts as the session controller. On segment one it hands your I2V stage the initial image. On every segment after that it hands it the tail of the previous segment - the frame that became the last frame of the last run - so motion carries forward instead of restarting from a random pose. Behind the scenes it keeps a small session dictionary keyed by the node's unique ID, resets it when iteration is 0, and pulls the saved continuation frame on later iterations.

The clever part is that it doesn't loop anything itself. When a segment finishes, the writer node mutates the workflow's own prompt, bumps iteration by one, and puts the whole thing back on the queue. That's the "auto" in AutoLong, and it's why nothing accumulates in RAM: each segment generates, writes its frames into one persistent FFmpeg stream, and frees its tensors before the next run starts. No Easy-Use loop nodes, no growing list of images.

The inputs that matter

• initial_image - the first frame of the whole video. Only used on iteration 0.
• total_segments - how many segments make up the video. Default 3, up to 100.
• iteration - leave this at 0. The node updates it internally when it re-queues itself. The README hammers this because it's the one setting that breaks everything if you touch it mid-run.

What comes out

The outputs that matter: start_image wires into your I2V first-frame conditioning, and index / total / session_id feed the stream writer so it knows which segment it's on and which session to append to. If you're doing SVI, previous_latent and motion_latent_count are the motion-conditioning handoff - the writer stores the last segment's latent and this node passes it back on the next iteration, which is how SVI keeps temporal continuity across segments instead of just re-using one pixel frame.

Install

ComfyUI Manager is the easy route - search "ComfyUI-DaSiWa-AutoLong" and hit install, or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/kokotu/ComfyUI-DaSiWa-AutoLong.git

Then restart ComfyUI. There are no Python requirements and no model files to download; the only system dependency is ffmpeg, which any environment with VideoHelperSuite already has. Output lands in ComfyUI/output/video under the DaSiWa_AUTO_LONG prefix.

Where people get burned

Don't queue the same workflow again while the auto-run is still going - the re-queue logic expects exactly one running prompt, and it will error out. And if you manually cancel partway, set iteration back to 0 before your next run, or the node will complain that continuation state is missing. Keep resolution and frame rate identical across segments too; the encoder refuses to continue a stream that changes size. Treat it right and you get a minute of video from a single click.