🔶 KSampler txt2video img2video - Advanced v1

If you're after "make an animation without a proper video model or a beefy card," this is the flagship node of the pack that promises it. Chaosaiart-Nodes' whole pitch, in the author's own words from the original release thread, is generating "frame by frame" through "Batch Count" rather than batch size, so the images stay "self-similar" - meaning each frame is a fresh img2img pass seeded from the previous frame's output, never holding a batch of frames in VRAM at once. This node is the sampler that actually does that per-frame generation, and it's built to run inside the pack's frame-loop system rather than standalone.

Under the hood it's not a single denoise pass - it's two, chained together. splitt_by_steps picks the step at which the pass splits, and denoise_part_1/denoise_part_2 set how strong each half is (defaults 0.6 and 1.0). This mirrors what the README calls the "Fixed 0.5" method, which it says was specifically tuned for SD 1.5-era checkpoints doing frame-to-frame animation - it's not an arbitrary design choice, it's the pack's answer to the flicker and drift that a single denoise pass tends to produce across a chain of frames.

The fields you'll actually touch: Image_Mode (aspect ratio: square, 16:9 widescreen, or 9:16 portrait) and Image_Size (360p/480p/HD/Full HD) set your output canvas. Img2img_input_Size decides what happens if your incoming image doesn't match that canvas - resize, crop, or override. seed_mode is worth reading twice: it independently controls whether the seed used for the very first frame (txt2video_seed_First_IMG) and the seed used for every frame after (seed_activ) stay fixed or increment, because frame one has no prior image to work from and often wants different seed behavior than the rest of the chain. steps, cfg, sampler_name, and scheduler behave like a normal KSampler - since this pack targets SD 1.5-class checkpoints, DPM++ 2M Karras at 20-30 steps is the safe starting point rather than anything tuned for a flow-matching model. positive, negative, model, and vae are the standard wiring. The optional start_Image seeds frame one with a real picture for img2video instead of pure noise; zoom_frame hooks into the pack's zoom/pan animation feature via its own custom type, which isn't covered in this article.

Required but easy to miss: activ_frame - this node won't do anything sensible without a chaosaiart_restarter_advanced node feeding it the current frame number.

Outputs: Info (a text status readout), IMAGE (the decoded frame, ready to display or cache), and SAMPLER (the raw latent, in case you want to keep working in latent space).

Install via ComfyUI Manager - search "Chaosaiart-Nodes" - or Manager's "Install via Git URL" with https://github.com/chaosaiart/Chaosaiart-Nodes. By hand: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes, restart. Linux also needs pip install opencv-python and pip install tqdm inside your venv.

The most common confusion: queuing this once and expecting motion. It's a per-frame node in a loop-driven system - without the restarter actually re-running the workflow multiple times, you'll get one static image, not an animation.