DaSiWa AutoLong SVI Pro Stream Writer

SVI 2.0 Pro is the reason you can finally make a 20–25 second continuous take on Wan 2.2, but it doesn't generate it as one blob - the workflow still runs in segments, and someone has to make those segments into a single file without a visible seam. That's this node's whole job. DaSiWaAutoLongStreamSVIPro is the current writer in the ComfyUI-DaSiWa-AutoLong pack, and it's the one you'll actually reach for on a modern DaSiWa SVI workflow.

The mechanism is the interesting part, because it sidesteps the two things that usually kill long-form video: RAM blow-up and MP4 stitching. Each segment's already-interpolated frames are piped straight into one persistent libx264 process as raw frames, so there's never a full video sitting in memory and never a set of clips to concatenate afterwards. When a segment ends, the node re-queues the workflow automatically; on the last segment it closes the encoder and you get one finished MP4 with a player and download button right inside the node.

The inputs that matter

Most of the plumbing - index, total_segments, session_id - comes wired from the Start node, and you leave it alone. The ones you actually set:

• overlap_frames - how many raw continuation frames get carried into the next segment. The README is blunt about the values: 1 for plain single-tail-frame continuation, 5 for SVI 2.0, 4 for SVI 2.0 Pro (one SVI 2.0 Pro temporal latent corresponds to 4 frames). This is the setting people get wrong, and the pack insists SVI 2.0 stays at 5 and 2.0 Pro stays at 4.
• interpolation_multiplier - must match your RIFE stage exactly: 1 if you don't interpolate, 2 for a 16→32 FPS upscale. The node uses it to skip the duplicated overlap region precisely, so a mismatch means duplicated frames or dropped content.
• frame_rate - default 32; the rate your output should play at.
• crf / preset - H.264 quality/speed, defaults 17 and "medium".
• filename_prefix - where the file lands, default video/DaSiWa_AUTO_LONG.

There's one optional input, samples (a LATENT). Wire the SVI motion latent here and the node stashes it in the session; the Start node then passes it back as previous_latent on the next iteration. That's what lets SVI carry temporal context across segments rather than just reusing one pixel frame.

Outputs and the wrap-up

filenames (VHS_FILENAMES) and video_path (STRING), same as the legacy writer. On the final segment the node also renders the finished file inline so you can preview and download without hunting through your output folder.

Install and gotchas

Install through ComfyUI Manager (search "ComfyUI-DaSiWa-AutoLong") or:

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

Restart, done. No Python deps, no model downloads - just ffmpeg on the system, which VideoHelperSuite environments already have.

Where people get burned: don't queue the same workflow while an auto-run is going (the re-queue logic demands exactly one running prompt), and if you cancel midway, reset iteration to 0 on the Start node or the session state is gone. Every segment must be the same resolution and frame rate - the encoder hard-stops if the frame size changes, and it also requires even width and height because of the yuv420p output. And resist the urge to fiddle with overlap_frames between runs while a workflow is mid-flight; warm-up state and skip math assume it's constant.